Upload 126 files

.flake8

@@ -0,0 +1,3 @@
+[flake8]
+select = E3, E4, F
+per-file-ignores = roop/core.py:E402

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+docs/screenshot.png filter=lfs diff=lfs merge=lfs -text
+roop-unleashed-main/docs/screenshot.png filter=lfs diff=lfs merge=lfs -text

.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,37 @@
+---
+name: Bug report
+about: Create a report to help us improve
+title: ''
+labels: ''
+assignees: ''
+
+---
+
+**Describe the bug**
+A clear and concise description of what the bug is.
+
+**To Reproduce**
+Steps to reproduce the behavior:
+1. Go to '...'
+2. Click on '....'
+3. Scroll down to '....'
+4. See error
+
+**Details**
+What OS are you using?
+- [ ] Linux
+- [ ] Linux in WSL
+- [ ] Windows 
+- [ ] Mac
+
+Are you using a GPU?
+- [ ] No. CPU FTW
+- [ ] NVIDIA
+- [ ] AMD
+- [ ] Intel
+- [ ] Mac
+
+**Which version of roop unleashed are you using?**
+
+**Screenshots**
+If applicable, add screenshots to help explain your problem.

.github/workflows/stale.yml

@@ -0,0 +1,29 @@
+# This workflow warns and then closes issues and PRs that have had no activity for a specified amount of time.
+#
+# You can adjust the behavior by modifying this file.
+# For more information, see:
+# https://github.com/actions/stale
+name: Mark stale issues and pull requests
+
+on:
+  schedule:
+  - cron: '32 0 * * *'
+
+jobs:
+  stale:
+
+    runs-on: ubuntu-latest
+    permissions:
+      issues: write
+      pull-requests: write
+
+    steps:
+    - uses: actions/stale@v5
+      with:
+        repo-token: ${{ secrets.GITHUB_TOKEN }}
+        stale-issue-message: 'This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.'
+        stale-pr-message: 'This PR is stale because it has been open 45 days with no activity. Remove stale label or comment or this will be closed in 10 days.'
+        close-issue-message: 'This issue was closed because it has been stalled for 5 days with no activity.'
+        days-before-stale: 30
+        days-before-close: 5
+        days-before-pr-close: -1

.gitignore

@@ -0,0 +1,15 @@
+.vs
+.idea
+models
+temp
+__pycache__
+*.pth
+/start.bat
+/env
+.vscode
+output
+temp
+config.yaml
+run.bat
+venv
+start.sh

Dockerfile

@@ -0,0 +1,18 @@
+FROM python:3.11
+
+# making app folder
+WORKDIR /app
+
+# copying files
+COPY . .
+
+# installing requirements
+RUN apt-get update 
+RUN apt-get install ffmpeg -y
+RUN pip install --upgrade pip 
+RUN pip install -r ./requirements.txt 
+
+# launching gradio app 
+ENV GRADIO_SERVER_NAME="0.0.0.0"
+EXPOSE 7860
+ENTRYPOINT python ./run.py

LICENSE

@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU Affero General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero General Public License from time to time.  Such new versions
+will be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU Affero General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU Affero General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU Affero General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License as published
+    by the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+<https://www.gnu.org/licenses/>.

README.md

@@ -1,14 +1,253 @@
----
-title: Faceswap Tool
-emoji: 👀
-colorFrom: green
-colorTo: blue
-sdk: gradio
-sdk_version: 5.9.1
-app_file: app.py
-pinned: false
-license: mit
-short_description: 'Roop_Unleashed '
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# roop-unleashed
+
+[Changelog](#changelog) • [Usage](#usage) • [Wiki](https://github.com/C0untFloyd/roop-unleashed/wiki)
+
+
+Uncensored Deepfakes for images and videos without training and an easy-to-use GUI.
+
+
+![Screen](https://github.com/C0untFloyd/roop-unleashed/assets/131583554/6ee6860d-efbe-4337-8c62-a67598863637)
+
+### Features
+
+- Platform-independant Browser GUI
+- Selection of multiple input/output faces in one go
+- Many different swapping modes, first detected, face selections, by gender
+- Batch processing of images/videos
+- Masking of face occluders using text prompts or automatically
+- Optional Face Upscaler/Restoration using different enhancers
+- Preview swapping from different video frames
+- Live Fake Cam using your webcam
+- Extras Tab for cutting videos etc.
+- Settings - storing configuration for next session
+- Theme Support
+
+and lots more...
+
+
+## Disclaimer
+
+This project is for technical and academic use only.
+Users of this software are expected to use this software responsibly while abiding the local law. If a face of a real person is being used, users are suggested to get consent from the concerned person and clearly mention that it is a deepfake when posting content online. Developers of this software will not be responsible for actions of end-users.
+**Please do not apply it to illegal and unethical scenarios.**
+
+In the event of violation of the legal and ethical requirements of the user's country or region, this code repository is exempt from liability
+
+### Installation
+
+Please refer to the [wiki](https://github.com/C0untFloyd/roop-unleashed/wiki).
+
+#### macOS Installation
+Simply run the following command. It will check and install all dependencies if necessary.
+
+`/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/PJF16/roop-unleashed/master/installer/macOSinstaller.sh)"`
+
+
+
+### Usage
+
+- Windows: run the `windows_run.bat` from the Installer.
+- Linux: `python run.py`
+- macOS: `sh runMacOS.sh`
+- Dockerfile:
+  ```shell
+  docker build -t roop-unleashed . && docker run -t \
+    -p 7860:7860 \
+    -v ./config.yaml:/app/config.yaml \
+    -v ./models:/app/models \
+    -v ./temp:/app/temp \
+    -v ./output:/app/output \
+    roop-unleashed
+  ```
+
+<a target="_blank" href="https://colab.research.google.com/github/C0untFloyd/roop-unleashed/blob/main/roop-unleashed.ipynb">
+  <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
+</a>
+  
+
+Additional commandline arguments are currently unsupported and settings should be done via the UI.
+
+> Note: When you run this program for the first time, it will download some models roughly ~2Gb in size.
+
+
+
+
+### Changelog
+
+**31.12.2024** v4.4.0 Hotfix
+
+Bugfix: Updated Colab to use present Cuda Drivers
+Bugfix: Live-Cam not working because of new face swapper
+Set default swapping model back to Insightface
+
+Happy New Year!
+
+
+**30.12.2024** v4.4.0
+
+- Added random face selection mode
+- Added alternative face swapping model with 128px & 256 px output ([ReSwapper](https://github.com/somanchiu/ReSwapper/tree/main))
+- Video repair added to Extras Tab
+- Updated most packages to newer versions. CUDA >= 12.4 now required!
+- Several minor bugfixes and QoL Changes
+
+
+**28.9.2024** v4.3.1
+
+- Bugfix: Several possible memory leaks
+- Added different output modes, e.g. to virtual cam stream
+- New swapping mode "All input faces"
+- Average total fps displayed and setting for autorun
+
+
+**16.9.2024** v4.2.8
+
+- Bugfix: Starting roop-unleashed without NVIDIA gpu but cuda option enabled
+- Bugfix: Target Faces couldn't be moved left/right
+- Bugfix: Enhancement and upscaling working again in virtual cam
+- Corrupt videos caught when adding to target files, displaying warning msg
+- Source Files Component cleared after face detection to release temp files
+- Added masking and mouth restore options to virtual cam
+
+
+**9.9.2024** v4.2.3
+
+- Hotfix for gradio pydantic issue with fastapi
+- Upgraded to Gradio 4.43 hoping it will fix remaining issues
+- Added new action when no face detected -> use last swapped
+- Specified image format for image controls - opening new tabs on preview images possible again!
+- Hardcoded image output format for livecam to jpeg - might be faster than previous webp
+- Chain events to be only executed if previous was a success
+
+
+**5.9.2024** v4.2.0
+
+- Added ability to move input & target faces order
+- New CLI Arguments override settings
+- Small UI changes to faceswapping tab
+- Added mask option and code for restoration of original mouth area
+- Updated gradio to v4.42.0
+- Added CLI Arguments --server_share and --cuda_device_id
+- Added webp image support
+
+
+**15.07.2024** v4.1.1
+
+- Bugfix: Post-processing after swapping
+
+
+**14.07.2024** v4.1.0
+
+- Added subsample upscaling to increase swap resolution
+- Upgraded gradio
+
+
+**12.05.2024** v4.0.0
+
+- Bugfix: Unnecessary init every frame in live-cam
+- Bugfix: Installer downloading insightface package each run
+- Added xseg masking to live-cam
+- Added realesrganx2 to frame processors
+- Upgraded some requirements
+- Added subtypes and different model support to frame processors
+- Allow frame processors to change resolutions of videos
+- Different OpenCV Cap for MacOS Virtual Cam
+- Added complete frame processing to extras tab
+- Colorize, upscale and misc filters added
+
+
+**22.04.2024** v3.9.0
+
+- Bugfix: Face detection bounding box corrupt values at weird angles
+- Rewrote mask previewing to work with every model
+- Switching mask engines toggles text interactivity
+- Clearing target files, resets face selection dropdown
+- Massive rewrite of swapping architecture, needed for xseg implementation
+- Added DFL Xseg Support for partial face occlusion
+- Face masking only runs when there is a face detected
+- Removed unnecessary toggle checkbox for text masking
+
+
+**22.03.2024** v3.6.5
+
+- Bugfix: Installer pulling latest update on first installation
+- Bugfix: Regression issue, blurring/erosion missing from face swap
+- Exposed erosion and blur amounts to UI
+- Using same values for manual masking too
+
+
+**20.03.2024** v3.6.3
+
+- Bugfix: Workaround for Gradio Slider Change Bug
+- Bugfix: CSS Styling to fix Gradio Image Height Bug
+- Made face swapping mask offsets resolution independant
+- Show offset mask as overlay
+- Changed layout for masking
+
+
+**18.03.2024** v3.6.0
+
+- Updated to Gradio 4.21.0 - requiring many changes under the hood
+- New manual masking (draw the mask yourself)
+- Extras Tab, streamlined cutting/joining videos
+- Re-added face selection by gender (on-demand loading, default turned off)
+- Removed unnecessary activate live-cam option
+- Added time info to preview frame and changed frame slider event to allow faster changes
+
+
+**10.03.2024** v3.5.5
+
+- Bugfix: Installer Path Env
+- Bugfix: file attributes
+- Video processing checks for presence of ffmpeg and displays warning if not found
+- Removed gender + age detection to speed up processing. Option removed from UI
+- Replaced restoreformer with restoreformer++
+- Live Cam recoded to run separate from virtual cam and without blocking controls
+- Swapping with only 1 target face allows selecting from several input faces
+
+
+
+**08.01.2024** v3.5.0
+
+- Bugfix: wrong access options when creating folders
+- New auto rotation of horizontal faces, fixing bad landmark positions (expanded on ![PR 364](https://github.com/C0untFloyd/roop-unleashed/pull/364))
+- Simple VR Option for stereo Images/Movies, best used in selected face mode
+- Added RestoreFormer Enhancer - https://github.com/wzhouxiff/RestoreFormer
+- Bumped up package versions for onnx/Torch etc.   
+
+
+**16.10.2023** v3.3.4
+
+**11.8.2023** v2.7.0
+
+Initial Gradio Version - old TkInter Version now deprecated
+
+- Re-added unified padding to face enhancers
+- Fixed DMDNet for all resolutions
+- Selecting target face now automatically switches swapping mode to selected
+- GPU providers are correctly set using the GUI (needs restart currently)
+- Local output folder can be opened from page
+- Unfinished extras functions disabled for now
+- Installer checks out specific commit, allowing to go back to first install
+- Updated readme for new gradio version
+- Updated Colab
+
+
+# Acknowledgements
+
+Lots of ideas, code or pre-trained models borrowed from the following projects:
+
+https://github.com/deepinsight/insightface<br />
+https://github.com/s0md3v/roop<br />
+https://github.com/AUTOMATIC1111/stable-diffusion-webui<br /> 
+https://github.com/Hillobar/Rope<br />
+https://github.com/TencentARC/GFPGAN<br />   
+https://github.com/kadirnar/codeformer-pip<br />
+https://github.com/csxmli2016/DMDNet<br />
+https://github.com/glucauze/sd-webui-faceswaplab<br />
+https://github.com/ykk648/face_power<br />
+
+<br />
+<br />
+Thanks to all developers!
+

clip/__init__.py

@@ -0,0 +1 @@
+from .clip import *

clip/bpe_simple_vocab_16e6.txt.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a
+size 1356917

clip/clip.py

@@ -0,0 +1,241 @@
+import hashlib
+import os
+import urllib
+import warnings
+from typing import Any, Union, List
+
+import torch
+from PIL import Image
+from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize
+from tqdm import tqdm
+
+from .model import build_model
+from .simple_tokenizer import SimpleTokenizer as _Tokenizer
+
+try:
+    from torchvision.transforms import InterpolationMode
+    BICUBIC = InterpolationMode.BICUBIC
+except ImportError:
+    BICUBIC = Image.BICUBIC
+
+
+
+__all__ = ["available_models", "load", "tokenize"]
+_tokenizer = _Tokenizer()
+
+_MODELS = {
+    "RN50": "https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
+    "RN101": "https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
+    "RN50x4": "https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt",
+    "RN50x16": "https://openaipublic.azureedge.net/clip/models/52378b407f34354e150460fe41077663dd5b39c54cd0bfd2b27167a4a06ec9aa/RN50x16.pt",
+    "RN50x64": "https://openaipublic.azureedge.net/clip/models/be1cfb55d75a9666199fb2206c106743da0f6468c9d327f3e0d0a543a9919d9c/RN50x64.pt",
+    "ViT-B/32": "https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
+    "ViT-B/16": "https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt",
+    "ViT-L/14": "https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt",
+    "ViT-L/14@336px": "https://openaipublic.azureedge.net/clip/models/3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02/ViT-L-14-336px.pt",
+}
+
+
+def _download(url: str, root: str):
+    os.makedirs(root, exist_ok=True)
+    filename = os.path.basename(url)
+
+    expected_sha256 = url.split("/")[-2]
+    download_target = os.path.join(root, filename)
+
+    if os.path.exists(download_target) and not os.path.isfile(download_target):
+        raise RuntimeError(f"{download_target} exists and is not a regular file")
+
+    if os.path.isfile(download_target):
+        if hashlib.sha256(open(download_target, "rb").read()).hexdigest() == expected_sha256:
+            return download_target
+        else:
+            warnings.warn(f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file")
+
+    with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
+        with tqdm(total=int(source.info().get("Content-Length")), ncols=80, unit='iB', unit_scale=True, unit_divisor=1024) as loop:
+            while True:
+                buffer = source.read(8192)
+                if not buffer:
+                    break
+
+                output.write(buffer)
+                loop.update(len(buffer))
+
+    if hashlib.sha256(open(download_target, "rb").read()).hexdigest() != expected_sha256:
+        raise RuntimeError("Model has been downloaded but the SHA256 checksum does not not match")
+
+    return download_target
+
+
+def _convert_image_to_rgb(image):
+    return image.convert("RGB")
+
+
+def _transform(n_px):
+    return Compose([
+        Resize(n_px, interpolation=BICUBIC),
+        CenterCrop(n_px),
+        _convert_image_to_rgb,
+        ToTensor(),
+        Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
+    ])
+
+
+def available_models() -> List[str]:
+    """Returns the names of available CLIP models"""
+    return list(_MODELS.keys())
+
+
+def load(name: str, device: Union[str, torch.device] = "cuda" if torch.cuda.is_available() else "cpu", jit: bool = False, download_root: str = None):
+    """Load a CLIP model
+
+    Parameters
+    ----------
+    name : str
+        A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict
+
+    device : Union[str, torch.device]
+        The device to put the loaded model
+
+    jit : bool
+        Whether to load the optimized JIT model or more hackable non-JIT model (default).
+
+    download_root: str
+        path to download the model files; by default, it uses "~/.cache/clip"
+
+    Returns
+    -------
+    model : torch.nn.Module
+        The CLIP model
+
+    preprocess : Callable[[PIL.Image], torch.Tensor]
+        A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input
+    """
+    if name in _MODELS:
+        model_path = _download(_MODELS[name], download_root or os.path.expanduser("~/.cache/clip"))
+    elif os.path.isfile(name):
+        model_path = name
+    else:
+        raise RuntimeError(f"Model {name} not found; available models = {available_models()}")
+
+    with open(model_path, 'rb') as opened_file:
+        try:
+            # loading JIT archive
+            model = torch.jit.load(opened_file, map_location=device if jit else "cpu").eval()
+            state_dict = None
+        except RuntimeError:
+            # loading saved state dict
+            if jit:
+                warnings.warn(f"File {model_path} is not a JIT archive. Loading as a state dict instead")
+                jit = False
+            state_dict = torch.load(opened_file, map_location="cpu")
+
+    if not jit:
+        model = build_model(state_dict or model.state_dict()).to(device)
+        if str(device) == "cpu":
+            model.float()
+        return model, _transform(model.visual.input_resolution)
+
+    # patch the device names
+    device_holder = torch.jit.trace(lambda: torch.ones([]).to(torch.device(device)), example_inputs=[])
+    device_node = [n for n in device_holder.graph.findAllNodes("prim::Constant") if "Device" in repr(n)][-1]
+
+    def _node_get(node: torch._C.Node, key: str):
+        """Gets attributes of a node which is polymorphic over return type.
+        
+        From https://github.com/pytorch/pytorch/pull/82628
+        """
+        sel = node.kindOf(key)
+        return getattr(node, sel)(key)
+
+    def patch_device(module):
+        try:
+            graphs = [module.graph] if hasattr(module, "graph") else []
+        except RuntimeError:
+            graphs = []
+
+        if hasattr(module, "forward1"):
+            graphs.append(module.forward1.graph)
+
+        for graph in graphs:
+            for node in graph.findAllNodes("prim::Constant"):
+                if "value" in node.attributeNames() and str(_node_get(node, "value")).startswith("cuda"):
+                    node.copyAttributes(device_node)
+
+    model.apply(patch_device)
+    patch_device(model.encode_image)
+    patch_device(model.encode_text)
+
+    # patch dtype to float32 on CPU
+    if str(device) == "cpu":
+        float_holder = torch.jit.trace(lambda: torch.ones([]).float(), example_inputs=[])
+        float_input = list(float_holder.graph.findNode("aten::to").inputs())[1]
+        float_node = float_input.node()
+
+        def patch_float(module):
+            try:
+                graphs = [module.graph] if hasattr(module, "graph") else []
+            except RuntimeError:
+                graphs = []
+
+            if hasattr(module, "forward1"):
+                graphs.append(module.forward1.graph)
+
+            for graph in graphs:
+                for node in graph.findAllNodes("aten::to"):
+                    inputs = list(node.inputs())
+                    for i in [1, 2]:  # dtype can be the second or third argument to aten::to()
+                        if _node_get(inputs[i].node(), "value") == 5:
+                            inputs[i].node().copyAttributes(float_node)
+
+        model.apply(patch_float)
+        patch_float(model.encode_image)
+        patch_float(model.encode_text)
+
+        model.float()
+
+    return model, _transform(model.input_resolution.item())
+
+
+def tokenize(texts: Union[str, List[str]], context_length: int = 77, truncate: bool = False) -> Union[torch.IntTensor, torch.LongTensor]:
+    """
+    Returns the tokenized representation of given input string(s)
+
+    Parameters
+    ----------
+    texts : Union[str, List[str]]
+        An input string or a list of input strings to tokenize
+
+    context_length : int
+        The context length to use; all CLIP models use 77 as the context length
+
+    truncate: bool
+        Whether to truncate the text in case its encoding is longer than the context length
+
+    Returns
+    -------
+    A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length].
+    We return LongTensor when torch version is <1.8.0, since older index_select requires indices to be long.
+    """
+    if isinstance(texts, str):
+        texts = [texts]
+
+    sot_token = _tokenizer.encoder["<|startoftext|>"]
+    eot_token = _tokenizer.encoder["<|endoftext|>"]
+    all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]
+    #if packaging.version.parse(torch.__version__) < packaging.version.parse("1.8.0"):
+    #    result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+    #else:
+    result = torch.zeros(len(all_tokens), context_length, dtype=torch.int)
+
+    for i, tokens in enumerate(all_tokens):
+        if len(tokens) > context_length:
+            if truncate:
+                tokens = tokens[:context_length]
+                tokens[-1] = eot_token
+            else:
+                raise RuntimeError(f"Input {texts[i]} is too long for context length {context_length}")
+        result[i, :len(tokens)] = torch.tensor(tokens)
+
+    return result

clip/clipseg.py

@@ -0,0 +1,538 @@
+import math
+from os.path import basename, dirname, join, isfile
+import torch
+from torch import nn
+from torch.nn import functional as nnf
+from torch.nn.modules.activation import ReLU
+
+
+def get_prompt_list(prompt):
+    if prompt == 'plain':
+        return ['{}']    
+    elif prompt == 'fixed':
+        return ['a photo of a {}.']
+    elif prompt == 'shuffle':
+        return ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.']
+    elif prompt == 'shuffle+':
+        return ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.',
+                            'a cropped photo of a {}.', 'a good photo of a {}.', 'a photo of one {}.',
+                            'a bad photo of a {}.', 'a photo of the {}.']
+    else:
+        raise ValueError('Invalid value for prompt')        
+
+
+def forward_multihead_attention(x, b, with_aff=False, attn_mask=None):
+    """ 
+    Simplified version of multihead attention (taken from torch source code but without tons of if clauses). 
+    The mlp and layer norm come from CLIP.
+    x: input.
+    b: multihead attention module. 
+    """
+
+    x_ = b.ln_1(x)
+    q, k, v = nnf.linear(x_, b.attn.in_proj_weight, b.attn.in_proj_bias).chunk(3, dim=-1)
+    tgt_len, bsz, embed_dim = q.size()
+
+    head_dim = embed_dim // b.attn.num_heads
+    scaling = float(head_dim) ** -0.5
+
+    q = q.contiguous().view(tgt_len, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
+    k = k.contiguous().view(-1, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
+    v = v.contiguous().view(-1, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
+
+    q = q * scaling
+
+    attn_output_weights = torch.bmm(q, k.transpose(1, 2)) #  n_heads * batch_size, tokens^2, tokens^2
+    if attn_mask is not None:
+
+
+        attn_mask_type, attn_mask = attn_mask
+        n_heads = attn_output_weights.size(0) // attn_mask.size(0)
+        attn_mask = attn_mask.repeat(n_heads, 1)
+        
+        if attn_mask_type == 'cls_token':
+            # the mask only affects similarities compared to the readout-token.
+            attn_output_weights[:, 0, 1:] = attn_output_weights[:, 0, 1:] * attn_mask[None,...]
+            # attn_output_weights[:, 0, 0] = 0*attn_output_weights[:, 0, 0]
+
+        if attn_mask_type == 'all':
+            # print(attn_output_weights.shape, attn_mask[:, None].shape)
+            attn_output_weights[:, 1:, 1:] = attn_output_weights[:, 1:, 1:] * attn_mask[:, None]
+        
+    
+    attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
+
+    attn_output = torch.bmm(attn_output_weights, v)
+    attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+    attn_output = b.attn.out_proj(attn_output)
+
+    x = x + attn_output
+    x = x + b.mlp(b.ln_2(x))
+
+    if with_aff:
+        return x, attn_output_weights
+    else:
+        return x
+
+
+class CLIPDenseBase(nn.Module):
+
+    def __init__(self, version, reduce_cond, reduce_dim, prompt, n_tokens):
+        super().__init__()
+
+        import clip
+
+        # prec = torch.FloatTensor
+        self.clip_model, _ = clip.load(version, device='cpu', jit=False)
+        self.model = self.clip_model.visual
+
+        # if not None, scale conv weights such that we obtain n_tokens.
+        self.n_tokens = n_tokens
+
+        for p in self.clip_model.parameters():
+            p.requires_grad_(False)
+
+        # conditional
+        if reduce_cond is not None:
+            self.reduce_cond = nn.Linear(512, reduce_cond)
+            for p in self.reduce_cond.parameters():
+                p.requires_grad_(False)
+        else:
+            self.reduce_cond = None        
+
+        self.film_mul = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        self.film_add = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        
+        self.reduce = nn.Linear(768, reduce_dim)
+
+        self.prompt_list = get_prompt_list(prompt)     
+
+        # precomputed prompts
+        import pickle
+        if isfile('precomputed_prompt_vectors.pickle'):
+            precomp = pickle.load(open('precomputed_prompt_vectors.pickle', 'rb'))
+            self.precomputed_prompts = {k: torch.from_numpy(v) for k, v in precomp.items()}        
+        else:
+            self.precomputed_prompts = dict()
+    
+    def rescaled_pos_emb(self, new_size):
+        assert len(new_size) == 2
+
+        a = self.model.positional_embedding[1:].T.view(1, 768, *self.token_shape)
+        b = nnf.interpolate(a, new_size, mode='bicubic', align_corners=False).squeeze(0).view(768, new_size[0]*new_size[1]).T
+        return torch.cat([self.model.positional_embedding[:1], b])
+
+    def visual_forward(self, x_inp, extract_layers=(), skip=False, mask=None):
+        
+
+        with torch.no_grad():
+
+            inp_size = x_inp.shape[2:]
+
+            if self.n_tokens is not None:
+                stride2 = x_inp.shape[2] // self.n_tokens
+                conv_weight2 = nnf.interpolate(self.model.conv1.weight, (stride2, stride2), mode='bilinear', align_corners=True)
+                x = nnf.conv2d(x_inp, conv_weight2, bias=self.model.conv1.bias, stride=stride2, dilation=self.model.conv1.dilation)
+            else:
+                x = self.model.conv1(x_inp)  # shape = [*, width, grid, grid]
+
+            x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+            x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+
+            x = torch.cat([self.model.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+
+            standard_n_tokens = 50 if self.model.conv1.kernel_size[0] == 32 else 197
+
+            if x.shape[1] != standard_n_tokens:
+                new_shape = int(math.sqrt(x.shape[1]-1))
+                x = x + self.rescaled_pos_emb((new_shape, new_shape)).to(x.dtype)[None,:,:]
+            else:
+                x = x + self.model.positional_embedding.to(x.dtype)
+
+            x = self.model.ln_pre(x)
+
+            x = x.permute(1, 0, 2)  # NLD -> LND
+
+            activations, affinities = [], []
+            for i, res_block in enumerate(self.model.transformer.resblocks):
+                
+                if mask is not None:
+                    mask_layer, mask_type, mask_tensor = mask
+                    if mask_layer == i or mask_layer == 'all':
+                        # import ipdb; ipdb.set_trace()
+                        size = int(math.sqrt(x.shape[0] - 1))
+                        
+                        attn_mask = (mask_type, nnf.interpolate(mask_tensor.unsqueeze(1).float(), (size, size)).view(mask_tensor.shape[0], size * size))
+                        
+                    else:
+                        attn_mask = None
+                else:
+                    attn_mask = None
+
+                x, aff_per_head = forward_multihead_attention(x, res_block, with_aff=True, attn_mask=attn_mask)
+
+                if i in extract_layers:
+                    affinities += [aff_per_head]
+
+                    #if self.n_tokens is not None:
+                    #    activations += [nnf.interpolate(x, inp_size, mode='bilinear', align_corners=True)]
+                    #else:
+                    activations += [x]
+
+                if len(extract_layers) > 0 and i == max(extract_layers) and skip:
+                    print('early skip')
+                    break
+                
+            x = x.permute(1, 0, 2)  # LND -> NLD
+            x = self.model.ln_post(x[:, 0, :])
+
+            if self.model.proj is not None:
+                x = x @ self.model.proj
+
+            return x, activations, affinities
+
+    def sample_prompts(self, words, prompt_list=None):
+
+        prompt_list = prompt_list if prompt_list is not None else self.prompt_list
+
+        prompt_indices = torch.multinomial(torch.ones(len(prompt_list)), len(words), replacement=True)
+        prompts = [prompt_list[i] for i in prompt_indices]
+        return [promt.format(w) for promt, w in zip(prompts, words)]
+
+    def get_cond_vec(self, conditional, batch_size):
+        # compute conditional from a single string
+        if conditional is not None and type(conditional) == str:
+            cond = self.compute_conditional(conditional)
+            cond = cond.repeat(batch_size, 1)
+
+        # compute conditional from string list/tuple
+        elif conditional is not None and type(conditional) in {list, tuple} and type(conditional[0]) == str:
+            assert len(conditional) == batch_size
+            cond = self.compute_conditional(conditional)
+
+        # use conditional directly
+        elif conditional is not None and type(conditional) == torch.Tensor and conditional.ndim == 2:
+            cond = conditional
+
+        # compute conditional from image
+        elif conditional is not None and type(conditional) == torch.Tensor:
+            with torch.no_grad():
+                cond, _, _ = self.visual_forward(conditional)
+        else:
+            raise ValueError('invalid conditional')
+        return cond   
+
+    def compute_conditional(self, conditional):
+        import clip
+
+        dev = next(self.parameters()).device
+
+        if type(conditional) in {list, tuple}:
+            text_tokens = clip.tokenize(conditional).to(dev)
+            cond = self.clip_model.encode_text(text_tokens)
+        else:
+            if conditional in self.precomputed_prompts:
+                cond = self.precomputed_prompts[conditional].float().to(dev)
+            else:
+                text_tokens = clip.tokenize([conditional]).to(dev)
+                cond = self.clip_model.encode_text(text_tokens)[0]
+        
+        if self.shift_vector is not None:
+            return cond + self.shift_vector
+        else:
+            return cond
+
+
+def clip_load_untrained(version):
+    assert version == 'ViT-B/16'
+    from clip.model import CLIP
+    from clip.clip import _MODELS, _download
+    model = torch.jit.load(_download(_MODELS['ViT-B/16'])).eval()
+    state_dict = model.state_dict()
+
+    vision_width = state_dict["visual.conv1.weight"].shape[0]
+    vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
+    vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
+    grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
+    image_resolution = vision_patch_size * grid_size
+    embed_dim = state_dict["text_projection"].shape[1]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith(f"transformer.resblocks")))
+
+    return CLIP(embed_dim, image_resolution, vision_layers, vision_width, vision_patch_size, 
+        context_length, vocab_size, transformer_width, transformer_heads, transformer_layers)    
+
+
+class CLIPDensePredT(CLIPDenseBase):
+
+    def __init__(self, version='ViT-B/32', extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4, prompt='fixed', 
+                 extra_blocks=0, reduce_cond=None, fix_shift=False,
+                 learn_trans_conv_only=False,  limit_to_clip_only=False, upsample=False, 
+                 add_calibration=False, rev_activations=False, trans_conv=None, n_tokens=None, complex_trans_conv=False):
+        
+        super().__init__(version, reduce_cond, reduce_dim, prompt, n_tokens)
+        # device = 'cpu'
+
+        self.extract_layers = extract_layers
+        self.cond_layer = cond_layer
+        self.limit_to_clip_only = limit_to_clip_only
+        self.process_cond = None
+        self.rev_activations = rev_activations
+        
+        depth = len(extract_layers)
+
+        if add_calibration:
+            self.calibration_conds = 1
+
+        self.upsample_proj = nn.Conv2d(reduce_dim, 1, kernel_size=1) if upsample else None
+
+        self.add_activation1 = True
+
+        self.version = version
+        
+        self.token_shape = {'ViT-B/32': (7, 7), 'ViT-B/16': (14, 14)}[version]
+
+        if fix_shift:
+            # self.shift_vector = nn.Parameter(torch.load(join(dirname(basename(__file__)), 'clip_text_shift_vector.pth')), requires_grad=False)
+            self.shift_vector = nn.Parameter(torch.load(join(dirname(basename(__file__)), 'shift_text_to_vis.pth')), requires_grad=False)
+            # self.shift_vector = nn.Parameter(-1*torch.load(join(dirname(basename(__file__)), 'shift2.pth')), requires_grad=False)
+        else:
+            self.shift_vector = None
+
+        if trans_conv is None:
+            trans_conv_ks = {'ViT-B/32': (32, 32), 'ViT-B/16': (16, 16)}[version]
+        else:
+            # explicitly define transposed conv kernel size
+            trans_conv_ks = (trans_conv, trans_conv)
+
+        if not complex_trans_conv:
+            self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+        else:
+            assert trans_conv_ks[0] == trans_conv_ks[1]
+
+            tp_kernels = (trans_conv_ks[0] // 4, trans_conv_ks[0] // 4)
+
+            self.trans_conv = nn.Sequential(
+                nn.Conv2d(reduce_dim, reduce_dim, kernel_size=3, padding=1),
+                nn.ReLU(),
+                nn.ConvTranspose2d(reduce_dim, reduce_dim // 2, kernel_size=tp_kernels[0], stride=tp_kernels[0]),
+                nn.ReLU(),
+                nn.ConvTranspose2d(reduce_dim // 2, 1, kernel_size=tp_kernels[1], stride=tp_kernels[1]),               
+            )
+
+#        self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+        
+        assert len(self.extract_layers) == depth
+
+        self.reduces = nn.ModuleList([nn.Linear(768, reduce_dim) for _ in range(depth)])
+        self.blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(len(self.extract_layers))])
+        self.extra_blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(extra_blocks)])
+        
+        # refinement and trans conv
+
+        if learn_trans_conv_only:
+            for p in self.parameters():
+                p.requires_grad_(False)
+            
+            for p in self.trans_conv.parameters():
+                p.requires_grad_(True)
+
+        self.prompt_list = get_prompt_list(prompt)
+
+
+    def forward(self, inp_image, conditional=None, return_features=False, mask=None):
+
+        assert type(return_features) == bool
+
+        inp_image = inp_image.to(self.model.positional_embedding.device)
+
+        if mask is not None:
+            raise ValueError('mask not supported')
+
+        # x_inp = normalize(inp_image)
+        x_inp = inp_image
+
+        bs, dev = inp_image.shape[0], x_inp.device
+
+        cond = self.get_cond_vec(conditional, bs)
+
+        visual_q, activations, _ = self.visual_forward(x_inp, extract_layers=[0] + list(self.extract_layers))
+
+        activation1 = activations[0]
+        activations = activations[1:]
+
+        _activations = activations[::-1] if not self.rev_activations else activations
+
+        a = None
+        for i, (activation, block, reduce) in enumerate(zip(_activations, self.blocks, self.reduces)):
+            
+            if a is not None:
+                a = reduce(activation) + a
+            else:
+                a = reduce(activation)
+
+            if i == self.cond_layer:
+                if self.reduce_cond is not None:
+                    cond = self.reduce_cond(cond)
+                
+                a = self.film_mul(cond) * a + self.film_add(cond)
+
+            a = block(a)
+
+        for block in self.extra_blocks:
+            a = a + block(a)
+
+        a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
+
+        size = int(math.sqrt(a.shape[2]))
+
+        a = a.view(bs, a.shape[1], size, size)
+
+        a = self.trans_conv(a)
+
+        if self.n_tokens is not None:
+            a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear', align_corners=True) 
+
+        if self.upsample_proj is not None:
+            a = self.upsample_proj(a)
+            a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear')
+
+        if return_features:
+            return a, visual_q, cond, [activation1] + activations
+        else:
+            return a,
+
+
+
+class CLIPDensePredTMasked(CLIPDensePredT):
+
+    def __init__(self, version='ViT-B/32', extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4, 
+                 prompt='fixed', extra_blocks=0, reduce_cond=None, fix_shift=False, learn_trans_conv_only=False, 
+                 refine=None, limit_to_clip_only=False, upsample=False, add_calibration=False, n_tokens=None):
+
+        super().__init__(version=version, extract_layers=extract_layers, cond_layer=cond_layer, reduce_dim=reduce_dim, 
+                         n_heads=n_heads, prompt=prompt, extra_blocks=extra_blocks, reduce_cond=reduce_cond, 
+                         fix_shift=fix_shift, learn_trans_conv_only=learn_trans_conv_only,
+                         limit_to_clip_only=limit_to_clip_only, upsample=upsample, add_calibration=add_calibration,
+                         n_tokens=n_tokens)
+
+    def visual_forward_masked(self, img_s, seg_s):
+        return super().visual_forward(img_s, mask=('all', 'cls_token', seg_s))
+
+    def forward(self, img_q, cond_or_img_s, seg_s=None, return_features=False):
+
+        if seg_s is None:
+            cond = cond_or_img_s
+        else:
+            img_s = cond_or_img_s
+
+            with torch.no_grad():
+                cond, _, _ = self.visual_forward_masked(img_s, seg_s)
+
+        return super().forward(img_q, cond, return_features=return_features)
+
+
+
+class CLIPDenseBaseline(CLIPDenseBase):
+
+    def __init__(self, version='ViT-B/32', cond_layer=0, 
+                extract_layer=9, reduce_dim=128, reduce2_dim=None, prompt='fixed', 
+                 reduce_cond=None, limit_to_clip_only=False, n_tokens=None):
+        
+        super().__init__(version, reduce_cond, reduce_dim, prompt, n_tokens)
+        device = 'cpu'
+
+        # self.cond_layer = cond_layer
+        self.extract_layer = extract_layer
+        self.limit_to_clip_only = limit_to_clip_only
+        self.shift_vector = None
+
+        self.token_shape = {'ViT-B/32': (7, 7), 'ViT-B/16': (14, 14)}[version]
+        
+        assert reduce2_dim is not None
+
+        self.reduce2 = nn.Sequential(
+            nn.Linear(reduce_dim, reduce2_dim),
+            nn.ReLU(),
+            nn.Linear(reduce2_dim, reduce_dim)
+        )
+        
+        trans_conv_ks = {'ViT-B/32': (32, 32), 'ViT-B/16': (16, 16)}[version]
+        self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+
+
+    def forward(self, inp_image, conditional=None, return_features=False):
+
+        inp_image = inp_image.to(self.model.positional_embedding.device)
+
+        # x_inp = normalize(inp_image)
+        x_inp = inp_image
+
+        bs, dev = inp_image.shape[0], x_inp.device
+
+        cond = self.get_cond_vec(conditional, bs)
+
+        visual_q, activations, affinities = self.visual_forward(x_inp, extract_layers=[self.extract_layer])
+
+        a = activations[0]
+        a = self.reduce(a)
+        a = self.film_mul(cond) * a + self.film_add(cond)
+
+        if self.reduce2 is not None:
+            a = self.reduce2(a)
+
+        # the original model would execute a transformer block here
+
+        a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
+
+        size = int(math.sqrt(a.shape[2]))
+
+        a = a.view(bs, a.shape[1], size, size)
+        a = self.trans_conv(a)
+
+        if return_features:
+            return a, visual_q, cond, activations
+        else:
+            return a,
+
+
+class CLIPSegMultiLabel(nn.Module):
+
+    def __init__(self, model) -> None:
+        super().__init__()
+
+        from third_party.JoEm.data_loader import get_seen_idx, get_unseen_idx, VOC
+
+        self.pascal_classes = VOC
+
+        from clip.clipseg import CLIPDensePredT
+        from general_utils import load_model
+        # self.clipseg = load_model('rd64-vit16-neg0.2-phrasecut', strict=False)
+        self.clipseg = load_model(model, strict=False)
+        
+        self.clipseg.eval()
+
+    def forward(self, x):
+
+        bs = x.shape[0]
+        out = torch.ones(21, bs, 352, 352).to(x.device) * -10
+
+        for class_id, class_name in enumerate(self.pascal_classes):
+        
+            fac = 3 if class_name == 'background' else 1
+
+            with torch.no_grad():
+                pred = torch.sigmoid(self.clipseg(x, class_name)[0][:,0]) * fac
+
+            out[class_id] += pred
+
+
+        out = out.permute(1, 0, 2, 3)
+
+        return out
+
+        # construct output tensor
+                    

clip/model.py

@@ -0,0 +1,436 @@
+from collections import OrderedDict
+from typing import Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1):
+        super().__init__()
+
+        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
+        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.relu1 = nn.ReLU(inplace=True)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.relu2 = nn.ReLU(inplace=True)
+
+        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
+
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+        self.relu3 = nn.ReLU(inplace=True)
+
+        self.downsample = None
+        self.stride = stride
+
+        if stride > 1 or inplanes != planes * Bottleneck.expansion:
+            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
+            self.downsample = nn.Sequential(OrderedDict([
+                ("-1", nn.AvgPool2d(stride)),
+                ("0", nn.Conv2d(inplanes, planes * self.expansion, 1, stride=1, bias=False)),
+                ("1", nn.BatchNorm2d(planes * self.expansion))
+            ]))
+
+    def forward(self, x: torch.Tensor):
+        identity = x
+
+        out = self.relu1(self.bn1(self.conv1(x)))
+        out = self.relu2(self.bn2(self.conv2(out)))
+        out = self.avgpool(out)
+        out = self.bn3(self.conv3(out))
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu3(out)
+        return out
+
+
+class AttentionPool2d(nn.Module):
+    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2 + 1, embed_dim) / embed_dim ** 0.5)
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.flatten(start_dim=2).permute(2, 0, 1)  # NCHW -> (HW)NC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x[:1], key=x, value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False
+        )
+        return x.squeeze(0)
+
+
+class ModifiedResNet(nn.Module):
+    """
+    A ResNet class that is similar to torchvision's but contains the following changes:
+    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
+    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
+    - The final pooling layer is a QKV attention instead of an average pool
+    """
+
+    def __init__(self, layers, output_dim, heads, input_resolution=224, width=64):
+        super().__init__()
+        self.output_dim = output_dim
+        self.input_resolution = input_resolution
+
+        # the 3-layer stem
+        self.conv1 = nn.Conv2d(3, width // 2, kernel_size=3, stride=2, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(width // 2)
+        self.relu1 = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.relu2 = nn.ReLU(inplace=True)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.relu3 = nn.ReLU(inplace=True)
+        self.avgpool = nn.AvgPool2d(2)
+
+        # residual layers
+        self._inplanes = width  # this is a *mutable* variable used during construction
+        self.layer1 = self._make_layer(width, layers[0])
+        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
+        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
+        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
+
+        embed_dim = width * 32  # the ResNet feature dimension
+        self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim, heads, output_dim)
+
+    def _make_layer(self, planes, blocks, stride=1):
+        layers = [Bottleneck(self._inplanes, planes, stride)]
+
+        self._inplanes = planes * Bottleneck.expansion
+        for _ in range(1, blocks):
+            layers.append(Bottleneck(self._inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        def stem(x):
+            x = self.relu1(self.bn1(self.conv1(x)))
+            x = self.relu2(self.bn2(self.conv2(x)))
+            x = self.relu3(self.bn3(self.conv3(x)))
+            x = self.avgpool(x)
+            return x
+
+        x = x.type(self.conv1.weight.dtype)
+        x = stem(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.attnpool(x)
+
+        return x
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        ret = super().forward(x.type(torch.float32))
+        return ret.type(orig_type)
+
+
+class QuickGELU(nn.Module):
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = LayerNorm(d_model)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, d_model * 4)),
+            ("gelu", QuickGELU()),
+            ("c_proj", nn.Linear(d_model * 4, d_model))
+        ]))
+        self.ln_2 = LayerNorm(d_model)
+        self.attn_mask = attn_mask
+
+    def attention(self, x: torch.Tensor):
+        self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
+        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attention(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])
+
+    def forward(self, x: torch.Tensor):
+        return self.resblocks(x)
+
+
+class VisionTransformer(nn.Module):
+    def __init__(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.output_dim = output_dim
+        self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)
+
+        scale = width ** -0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width))
+        self.ln_pre = LayerNorm(width)
+
+        self.transformer = Transformer(width, layers, heads)
+
+        self.ln_post = LayerNorm(width)
+        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+    def forward(self, x: torch.Tensor):
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+        x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        x = self.ln_post(x[:, 0, :])
+
+        if self.proj is not None:
+            x = x @ self.proj
+
+        return x
+
+
+class CLIP(nn.Module):
+    def __init__(self,
+                 embed_dim: int,
+                 # vision
+                 image_resolution: int,
+                 vision_layers: Union[Tuple[int, int, int, int], int],
+                 vision_width: int,
+                 vision_patch_size: int,
+                 # text
+                 context_length: int,
+                 vocab_size: int,
+                 transformer_width: int,
+                 transformer_heads: int,
+                 transformer_layers: int
+                 ):
+        super().__init__()
+
+        self.context_length = context_length
+
+        if isinstance(vision_layers, (tuple, list)):
+            vision_heads = vision_width * 32 // 64
+            self.visual = ModifiedResNet(
+                layers=vision_layers,
+                output_dim=embed_dim,
+                heads=vision_heads,
+                input_resolution=image_resolution,
+                width=vision_width
+            )
+        else:
+            vision_heads = vision_width // 64
+            self.visual = VisionTransformer(
+                input_resolution=image_resolution,
+                patch_size=vision_patch_size,
+                width=vision_width,
+                layers=vision_layers,
+                heads=vision_heads,
+                output_dim=embed_dim
+            )
+
+        self.transformer = Transformer(
+            width=transformer_width,
+            layers=transformer_layers,
+            heads=transformer_heads,
+            attn_mask=self.build_attention_mask()
+        )
+
+        self.vocab_size = vocab_size
+        self.token_embedding = nn.Embedding(vocab_size, transformer_width)
+        self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width))
+        self.ln_final = LayerNorm(transformer_width)
+
+        self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim))
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+        self.initialize_parameters()
+
+    def initialize_parameters(self):
+        nn.init.normal_(self.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.positional_embedding, std=0.01)
+
+        if isinstance(self.visual, ModifiedResNet):
+            if self.visual.attnpool is not None:
+                std = self.visual.attnpool.c_proj.in_features ** -0.5
+                nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std)
+
+            for resnet_block in [self.visual.layer1, self.visual.layer2, self.visual.layer3, self.visual.layer4]:
+                for name, param in resnet_block.named_parameters():
+                    if name.endswith("bn3.weight"):
+                        nn.init.zeros_(param)
+
+        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
+        attn_std = self.transformer.width ** -0.5
+        fc_std = (2 * self.transformer.width) ** -0.5
+        for block in self.transformer.resblocks:
+            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+
+        if self.text_projection is not None:
+            nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.context_length, self.context_length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    @property
+    def dtype(self):
+        return self.visual.conv1.weight.dtype
+
+    def encode_image(self, image):
+        return self.visual(image.type(self.dtype))
+
+    def encode_text(self, text):
+        x = self.token_embedding(text).type(self.dtype)  # [batch_size, n_ctx, d_model]
+
+        x = x + self.positional_embedding.type(self.dtype)
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.ln_final(x).type(self.dtype)
+
+        # x.shape = [batch_size, n_ctx, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
+
+        return x
+
+    def forward(self, image, text):
+        image_features = self.encode_image(image)
+        text_features = self.encode_text(text)
+
+        # normalized features
+        image_features = image_features / image_features.norm(dim=1, keepdim=True)
+        text_features = text_features / text_features.norm(dim=1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale = self.logit_scale.exp()
+        logits_per_image = logit_scale * image_features @ text_features.t()
+        logits_per_text = logits_per_image.t()
+
+        # shape = [global_batch_size, global_batch_size]
+        return logits_per_image, logits_per_text
+
+
+def convert_weights(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+        if isinstance(l, nn.MultiheadAttention):
+            for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]:
+                tensor = getattr(l, attr)
+                if tensor is not None:
+                    tensor.data = tensor.data.half()
+
+        for name in ["text_projection", "proj"]:
+            if hasattr(l, name):
+                attr = getattr(l, name)
+                if attr is not None:
+                    attr.data = attr.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+
+def build_model(state_dict: dict):
+    vit = "visual.proj" in state_dict
+
+    if vit:
+        vision_width = state_dict["visual.conv1.weight"].shape[0]
+        vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
+        vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
+        grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
+        image_resolution = vision_patch_size * grid_size
+    else:
+        counts: list = [len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]]
+        vision_layers = tuple(counts)
+        vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
+        output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5)
+        vision_patch_size = None
+        assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0]
+        image_resolution = output_width * 32
+
+    embed_dim = state_dict["text_projection"].shape[1]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith("transformer.resblocks")))
+
+    model = CLIP(
+        embed_dim,
+        image_resolution, vision_layers, vision_width, vision_patch_size,
+        context_length, vocab_size, transformer_width, transformer_heads, transformer_layers
+    )
+
+    for key in ["input_resolution", "context_length", "vocab_size"]:
+        if key in state_dict:
+            del state_dict[key]
+
+    convert_weights(model)
+    model.load_state_dict(state_dict)
+    return model.eval()

clip/simple_tokenizer.py

@@ -0,0 +1,132 @@
+import gzip
+import html
+import os
+from functools import lru_cache
+
+import ftfy
+import regex as re
+
+
+@lru_cache()
+def default_bpe():
+    return os.path.join(os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz")
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r'\s+', ' ', text)
+    text = text.strip()
+    return text
+
+
+class SimpleTokenizer(object):
+    def __init__(self, bpe_path: str = default_bpe()):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        merges = gzip.open(bpe_path).read().decode("utf-8").split('\n')
+        merges = merges[1:49152-256-2+1]
+        merges = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v+'</w>' for v in vocab]
+        for merge in merges:
+            vocab.append(''.join(merge))
+        vocab.extend(['<|startoftext|>', '<|endoftext|>'])
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
+        self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE)
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + ( token[-1] + '</w>',)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token+'</w>'
+
+        while True:
+            bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+
+                if word[i] == first and i < len(word)-1 and word[i+1] == second:
+                    new_word.append(first+second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = ' '.join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = whitespace_clean(basic_clean(text)).lower()
+        for token in re.findall(self.pat, text):
+            token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
+            bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = ''.join([self.decoder[token] for token in tokens])
+        text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ')
+        return text

clip/vitseg.py

@@ -0,0 +1,286 @@
+import math
+from posixpath import basename, dirname, join
+# import clip
+from clip.model import convert_weights
+import torch
+import json
+from torch import nn
+from torch.nn import functional as nnf
+from torch.nn.modules import activation
+from torch.nn.modules.activation import ReLU
+from torchvision import transforms
+
+normalize = transforms.Normalize(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711))
+
+from torchvision.models import ResNet
+
+
+def process_prompts(conditional, prompt_list, conditional_map):
+    # DEPRECATED
+            
+    # randomly sample a synonym
+    words = [conditional_map[int(i)] for i in conditional]
+    words = [syns[torch.multinomial(torch.ones(len(syns)), 1, replacement=True).item()] for syns in words]
+    words = [w.replace('_', ' ') for w in words]
+
+    if prompt_list is not None:
+        prompt_indices = torch.multinomial(torch.ones(len(prompt_list)), len(words), replacement=True)
+        prompts = [prompt_list[i] for i in prompt_indices]
+    else:
+        prompts = ['a photo of {}'] * (len(words))
+
+    return [promt.format(w) for promt, w in zip(prompts, words)]
+
+
+class VITDenseBase(nn.Module):
+    
+    def rescaled_pos_emb(self, new_size):
+        assert len(new_size) == 2
+
+        a = self.model.positional_embedding[1:].T.view(1, 768, *self.token_shape)
+        b = nnf.interpolate(a, new_size, mode='bicubic', align_corners=False).squeeze(0).view(768, new_size[0]*new_size[1]).T
+        return torch.cat([self.model.positional_embedding[:1], b])
+
+    def visual_forward(self, x_inp, extract_layers=(), skip=False, mask=None):
+        
+        with torch.no_grad():
+
+            x_inp = nnf.interpolate(x_inp, (384, 384))
+
+            x = self.model.patch_embed(x_inp)
+            cls_token = self.model.cls_token.expand(x.shape[0], -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+            if self.model.dist_token is None:
+                x = torch.cat((cls_token, x), dim=1)
+            else:
+                x = torch.cat((cls_token, self.model.dist_token.expand(x.shape[0], -1, -1), x), dim=1)
+            x = self.model.pos_drop(x + self.model.pos_embed)
+
+            activations = []
+            for i, block in enumerate(self.model.blocks):
+                x = block(x)
+
+                if i in extract_layers:
+                    # permute to be compatible with CLIP
+                    activations += [x.permute(1,0,2)]                
+
+            x = self.model.norm(x)
+            x = self.model.head(self.model.pre_logits(x[:, 0]))
+
+            # again for CLIP compatibility
+            # x = x.permute(1, 0, 2)
+
+        return x, activations, None
+
+    def sample_prompts(self, words, prompt_list=None):
+
+        prompt_list = prompt_list if prompt_list is not None else self.prompt_list
+
+        prompt_indices = torch.multinomial(torch.ones(len(prompt_list)), len(words), replacement=True)
+        prompts = [prompt_list[i] for i in prompt_indices]
+        return [promt.format(w) for promt, w in zip(prompts, words)]
+
+    def get_cond_vec(self, conditional, batch_size):
+        # compute conditional from a single string
+        if conditional is not None and type(conditional) == str:
+            cond = self.compute_conditional(conditional)
+            cond = cond.repeat(batch_size, 1)
+
+        # compute conditional from string list/tuple
+        elif conditional is not None and type(conditional) in {list, tuple} and type(conditional[0]) == str:
+            assert len(conditional) == batch_size
+            cond = self.compute_conditional(conditional)
+
+        # use conditional directly
+        elif conditional is not None and type(conditional) == torch.Tensor and conditional.ndim == 2:
+            cond = conditional
+
+        # compute conditional from image
+        elif conditional is not None and type(conditional) == torch.Tensor:
+            with torch.no_grad():
+                cond, _, _ = self.visual_forward(conditional)
+        else:
+            raise ValueError('invalid conditional')
+        return cond   
+
+    def compute_conditional(self, conditional):
+        import clip
+
+        dev = next(self.parameters()).device
+
+        if type(conditional) in {list, tuple}:
+            text_tokens = clip.tokenize(conditional).to(dev)
+            cond = self.clip_model.encode_text(text_tokens)
+        else:
+            if conditional in self.precomputed_prompts:
+                cond = self.precomputed_prompts[conditional].float().to(dev)
+            else:
+                text_tokens = clip.tokenize([conditional]).to(dev)
+                cond = self.clip_model.encode_text(text_tokens)[0]
+        
+        return cond
+
+
+class VITDensePredT(VITDenseBase):
+
+    def __init__(self, extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4, prompt='fixed', 
+                 depth=3, extra_blocks=0, reduce_cond=None, fix_shift=False,
+                 learn_trans_conv_only=False, refine=None, limit_to_clip_only=False, upsample=False, 
+                 add_calibration=False, process_cond=None, not_pretrained=False):
+        super().__init__()
+        # device = 'cpu'
+
+        self.extract_layers = extract_layers
+        self.cond_layer = cond_layer
+        self.limit_to_clip_only = limit_to_clip_only
+        self.process_cond = None
+        
+        if add_calibration:
+            self.calibration_conds = 1
+
+        self.upsample_proj = nn.Conv2d(reduce_dim, 1, kernel_size=1) if upsample else None
+
+        self.add_activation1 = True
+
+        import timm 
+        self.model = timm.create_model('vit_base_patch16_384', pretrained=True)
+        self.model.head = nn.Linear(768, 512 if reduce_cond is None else reduce_cond)
+
+        for p in self.model.parameters():
+            p.requires_grad_(False)
+
+        import clip
+        self.clip_model, _ = clip.load('ViT-B/16', device='cpu', jit=False)
+        # del self.clip_model.visual
+        
+        
+        self.token_shape = (14, 14)
+
+        # conditional
+        if reduce_cond is not None:
+            self.reduce_cond = nn.Linear(512, reduce_cond)
+            for p in self.reduce_cond.parameters():
+                p.requires_grad_(False)
+        else:
+            self.reduce_cond = None
+
+        # self.film = AVAILABLE_BLOCKS['film'](512, 128)
+        self.film_mul = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        self.film_add = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        
+        # DEPRECATED
+        # self.conditional_map = {c['id']: c['synonyms'] for c in json.load(open(cond_map))}
+        
+        assert len(self.extract_layers) == depth
+
+        self.reduces = nn.ModuleList([nn.Linear(768, reduce_dim) for _ in range(depth)])
+        self.blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(len(self.extract_layers))])
+        self.extra_blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(extra_blocks)])
+
+        trans_conv_ks = (16, 16)
+        self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+
+        # refinement and trans conv
+
+        if learn_trans_conv_only:
+            for p in self.parameters():
+                p.requires_grad_(False)
+            
+            for p in self.trans_conv.parameters():
+                p.requires_grad_(True)
+
+        if prompt == 'fixed':
+            self.prompt_list = ['a photo of a {}.']
+        elif prompt == 'shuffle':
+            self.prompt_list = ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.']
+        elif prompt == 'shuffle+':
+            self.prompt_list = ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.',
+                                'a cropped photo of a {}.', 'a good photo of a {}.', 'a photo of one {}.',
+                                'a bad photo of a {}.', 'a photo of the {}.']
+        elif prompt == 'shuffle_clip':
+            from models.clip_prompts import imagenet_templates
+            self.prompt_list = imagenet_templates
+
+        if process_cond is not None:
+            if process_cond == 'clamp' or process_cond[0] == 'clamp':
+
+                val = process_cond[1] if type(process_cond) in {list, tuple} else 0.2
+
+                def clamp_vec(x):
+                    return torch.clamp(x, -val, val)
+
+                self.process_cond = clamp_vec
+
+            elif process_cond.endswith('.pth'):
+                
+                shift = torch.load(process_cond)
+                def add_shift(x):
+                    return x + shift.to(x.device)
+
+                self.process_cond = add_shift
+
+        import pickle
+        precomp = pickle.load(open('precomputed_prompt_vectors.pickle', 'rb'))
+        self.precomputed_prompts = {k: torch.from_numpy(v) for k, v in precomp.items()}
+
+
+    def forward(self, inp_image, conditional=None, return_features=False, mask=None):
+
+        assert type(return_features) == bool
+
+        # inp_image = inp_image.to(self.model.positional_embedding.device)
+
+        if mask is not None:
+            raise ValueError('mask not supported')
+
+        # x_inp = normalize(inp_image)
+        x_inp = inp_image
+
+        bs, dev = inp_image.shape[0], x_inp.device
+
+        inp_image_size = inp_image.shape[2:]
+
+        cond = self.get_cond_vec(conditional, bs)
+
+        visual_q, activations, _ = self.visual_forward(x_inp, extract_layers=[0] + list(self.extract_layers))
+
+        activation1 = activations[0]
+        activations = activations[1:]
+
+        a = None
+        for i, (activation, block, reduce) in enumerate(zip(activations[::-1], self.blocks, self.reduces)):
+            
+            if a is not None:
+                a = reduce(activation) + a
+            else:
+                a = reduce(activation)
+
+            if i == self.cond_layer:
+                if self.reduce_cond is not None:
+                    cond = self.reduce_cond(cond)
+                
+                a = self.film_mul(cond) * a + self.film_add(cond)
+
+            a = block(a)
+
+        for block in self.extra_blocks:
+            a = a + block(a)
+
+        a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
+
+        size = int(math.sqrt(a.shape[2]))
+
+        a = a.view(bs, a.shape[1], size, size)
+
+        if self.trans_conv is not None:
+            a = self.trans_conv(a)
+
+        if self.upsample_proj is not None:
+            a = self.upsample_proj(a)
+            a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear')
+
+        a = nnf.interpolate(a, inp_image_size)
+
+        if return_features:
+            return a, visual_q, cond, [activation1] + activations
+        else:
+            return a,

config_colab.yaml

@@ -0,0 +1,14 @@
+clear_output: true
+force_cpu: false
+max_threads: 3
+memory_limit: 0
+output_image_format: png
+output_template: '{file}_{time}'
+output_video_codec: libx264
+output_video_format: mp4
+provider: cuda
+selected_theme: Default
+server_name: ''
+server_port: 0
+server_share: true
+video_quality: 14

installer/installer.py

@@ -0,0 +1,87 @@
+import argparse
+import glob
+import os
+import shutil
+import site
+import subprocess
+import sys
+
+
+script_dir = os.getcwd()
+
+
+def run_cmd(cmd, capture_output=False, env=None):
+    # Run shell commands
+    return subprocess.run(cmd, shell=True, capture_output=capture_output, env=env)
+
+
+def check_env():
+    # If we have access to conda, we are probably in an environment
+    conda_not_exist = run_cmd("conda", capture_output=True).returncode
+    if conda_not_exist:
+        print("Conda is not installed. Exiting...")
+        sys.exit()
+    
+    # Ensure this is a new environment and not the base environment
+    if os.environ["CONDA_DEFAULT_ENV"] == "base":
+        print("Create an environment for this project and activate it. Exiting...")
+        sys.exit()
+
+
+def install_dependencies():
+    global MY_PATH
+
+    # Install Git and clone repo
+    run_cmd("conda install -y -k git")
+    run_cmd("git clone https://github.com/C0untFloyd/roop-unleashed.git")
+    os.chdir(MY_PATH)
+    run_cmd("git checkout 5bfafdc97a0c47b46ec83e6530a57399aaad75d7")
+    # Installs dependencies from requirements.txt
+    run_cmd("python -m pip install -r requirements.txt")
+
+
+
+def update_dependencies():
+    global MY_PATH
+    
+    os.chdir(MY_PATH)
+	# do a hard reset for to update even if there are local changes
+    run_cmd("git fetch --all")
+    run_cmd("git reset --hard origin/main")
+    run_cmd("git pull")
+    # Installs/Updates dependencies from all requirements.txt
+    run_cmd("python -m pip install -r requirements.txt")
+
+
+def start_app():
+    global MY_PATH
+    
+    os.chdir(MY_PATH)
+    # forward commandline arguments
+    sys.argv.pop(0)
+    args = ' '.join(sys.argv)
+    print("Launching App")
+    run_cmd(f'python run.py {args}')
+
+
+if __name__ == "__main__":
+    global MY_PATH
+    
+    MY_PATH = "roop-unleashed"
+
+    
+    # Verifies we are in a conda environment
+    check_env()
+
+    # If webui has already been installed, skip and run
+    if not os.path.exists(MY_PATH):
+        install_dependencies()
+    else:
+        # moved update from batch to here, because of batch limitations
+        updatechoice = input("Check for Updates? [y/n]").lower()
+        if updatechoice == "y":
+           update_dependencies()
+
+    # Run the model with webui
+    os.chdir(script_dir)
+    start_app()

installer/macOSinstaller.sh

@@ -0,0 +1,73 @@
+#!/bin/bash
+
+# This script checks and installs all dependencies which are needed to run roop-unleashed. After that, it clones the repo.
+# Execute this easily with /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/PJF16/roop-unleashed/master/installer/macOSinstaller.sh)
+
+# Function to check if a command exists
+command_exists() {
+    command -v "$1" >/dev/null 2>&1
+}
+
+echo "Starting check and installation process of dependencies for roop-unleashed"
+
+# Check if Homebrew is installed
+if ! command_exists brew; then
+    echo "Homebrew is not installed. Starting installation..."
+    /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
+else
+    echo "Homebrew is already installed."
+fi
+
+# Update Homebrew
+echo "Updating Homebrew..."
+brew update
+
+# Check if Python 3.11 is installed
+if brew list --versions [email protected] >/dev/null; then
+    echo "Python 3.11 is already installed."
+else
+    echo "Python 3.11 is not installed. Installing it now..."
+    brew install [email protected]
+fi
+
+# Check if [email protected] is installed
+if brew list --versions [email protected] >/dev/null; then
+    echo "[email protected] is already installed."
+else
+    echo "[email protected] is not installed. Installing it now..."
+    brew install [email protected]
+fi
+
+# Check if ffmpeg is installed
+if command_exists ffmpeg; then
+    echo "ffmpeg is already installed."
+else
+    echo "ffmpeg is not installed. Installing it now..."
+    brew install ffmpeg
+fi
+
+# Check if git is installed
+if command_exists git; then
+    echo "git is already installed."
+else
+    echo "git is not installed. Installing it now..."
+    brew install git
+fi
+
+# Clone the repository
+REPO_URL="https://github.com/C0untFloyd/roop-unleashed.git"
+REPO_NAME="roop-unleashed"
+
+echo "Cloning the repository $REPO_URL..."
+git clone $REPO_URL
+
+# Check if the repository was cloned successfully
+if [ -d "$REPO_NAME" ]; then
+    echo "Repository cloned successfully. Changing into directory $REPO_NAME..."
+    cd "$REPO_NAME"
+else
+    echo "Failed to clone the repository."
+fi
+
+echo "Check and installation process completed."
+

installer/windows_run.bat

@@ -0,0 +1,95 @@
+@echo off
+
+REM No CLI arguments supported anymore
+set COMMANDLINE_ARGS=
+
+cd /D "%~dp0"
+
+echo "%CD%"| findstr /C:" " >nul && echo This script relies on Miniconda which can not be silently installed under a path with spaces. && goto end
+
+set PATH=%PATH%;%SystemRoot%\system32
+
+@rem config
+set INSTALL_DIR=%cd%\installer_files
+set CONDA_ROOT_PREFIX=%cd%\installer_files\conda
+set INSTALL_ENV_DIR=%cd%\installer_files\env
+set MINICONDA_DOWNLOAD_URL=https://repo.anaconda.com/miniconda/Miniconda3-latest-Windows-x86_64.exe
+set FFMPEG_DOWNLOAD_URL=https://github.com/GyanD/codexffmpeg/releases/download/7.1/ffmpeg-7.1-essentials_build.zip
+set INSTALL_FFMPEG_DIR=%cd%\installer_files\ffmpeg
+set INSIGHTFACE_PACKAGE_URL=https://github.com/C0untFloyd/roop-unleashed/releases/download/3.6.6/insightface-0.7.3-cp310-cp310-win_amd64.whl
+set INSIGHTFACE_PACKAGE_PATH=%INSTALL_DIR%\insightface-0.7.3-cp310-cp310-win_amd64.whl
+
+set conda_exists=F
+set ffmpeg_exists=F
+
+@rem figure out whether git and conda needs to be installed
+call "%CONDA_ROOT_PREFIX%\_conda.exe" --version >nul 2>&1
+if "%ERRORLEVEL%" EQU "0" set conda_exists=T
+
+@rem Check if FFmpeg is already in PATH
+where ffmpeg >nul 2>&1
+if "%ERRORLEVEL%" EQU "0" (
+    echo FFmpeg is already installed.
+    set ffmpeg_exists=T
+)
+
+@rem (if necessary) install git and conda into a contained environment
+
+@rem download conda
+if "%conda_exists%" == "F" (
+    echo Downloading Miniconda from %MINICONDA_DOWNLOAD_URL% to %INSTALL_DIR%\miniconda_installer.exe
+    mkdir "%INSTALL_DIR%"
+    call curl -Lk "%MINICONDA_DOWNLOAD_URL%" > "%INSTALL_DIR%\miniconda_installer.exe" || ( echo. && echo Miniconda failed to download. && goto end )
+    echo Installing Miniconda to %CONDA_ROOT_PREFIX%
+    start /wait "" "%INSTALL_DIR%\miniconda_installer.exe" /InstallationType=JustMe /NoShortcuts=1 /AddToPath=0 /RegisterPython=0 /NoRegistry=1 /S /D=%CONDA_ROOT_PREFIX%
+
+    @rem test the conda binary
+    echo Miniconda version:
+    call "%CONDA_ROOT_PREFIX%\_conda.exe" --version || ( echo. && echo Miniconda not found. && goto end )
+)
+
+@rem create the installer env
+if not exist "%INSTALL_ENV_DIR%" (
+    echo Creating Conda Environment
+    call "%CONDA_ROOT_PREFIX%\_conda.exe" create --no-shortcuts -y -k --prefix "%INSTALL_ENV_DIR%" python=3.10 || ( echo. && echo ERROR: Conda environment creation failed. && goto end )
+    @rem check if conda environment was actually created
+    if not exist "%INSTALL_ENV_DIR%\python.exe" ( echo. && echo ERROR: Conda environment is empty. && goto end )
+    @rem activate installer env
+    call "%CONDA_ROOT_PREFIX%\condabin\conda.bat" activate "%INSTALL_ENV_DIR%" || ( echo. && echo ERROR: Miniconda hook not found. && goto end )
+    @rem Download insightface package
+    echo Downloading insightface package from %INSIGHTFACE_PACKAGE_URL% to %INSIGHTFACE_PACKAGE_PATH%
+    call curl -Lk "%INSIGHTFACE_PACKAGE_URL%" > "%INSIGHTFACE_PACKAGE_PATH%" || ( echo. && echo ERROR: Insightface package failed to download. && goto end )
+    @rem install insightface package using pip
+    echo Installing insightface package
+    call pip install "%INSIGHTFACE_PACKAGE_PATH%" || ( echo. && echo ERROR: Insightface package installation failed. && goto end )
+)
+
+@rem Download and install FFmpeg if not already installed
+if "%ffmpeg_exists%" == "F" (
+    if not exist "%INSTALL_FFMPEG_DIR%" (
+        echo Downloading ffmpeg from %FFMPEG_DOWNLOAD_URL% to %INSTALL_DIR%
+        call curl -Lk "%FFMPEG_DOWNLOAD_URL%" > "%INSTALL_DIR%\ffmpeg.zip" || ( echo. && echo ffmpeg failed to download. && goto end )
+        call powershell -command "Expand-Archive -Force '%INSTALL_DIR%\ffmpeg.zip' '%INSTALL_DIR%\'"
+        cd "%INSTALL_DIR%"
+        move ffmpeg-* ffmpeg
+        setx PATH "%INSTALL_FFMPEG_DIR%\bin\;%PATH%"
+        echo To use videos, you need to restart roop after this installation.
+        cd ..
+    )
+) else (
+    echo Skipping FFmpeg installation as it is already available.
+)
+
+@rem setup installer env
+@rem check if conda environment was actually created
+if not exist "%INSTALL_ENV_DIR%\python.exe" ( echo. && echo ERROR: Conda environment is empty. && goto end )
+@rem activate installer env
+call "%CONDA_ROOT_PREFIX%\condabin\conda.bat" activate "%INSTALL_ENV_DIR%" || ( echo. && echo ERROR: Miniconda hook not found. && goto end )
+echo Launching roop unleashed
+call python installer.py %COMMANDLINE_ARGS%
+
+echo.
+echo Done!
+
+:end
+pause

mypy.ini

@@ -0,0 +1,7 @@
+[mypy]
+check_untyped_defs = True
+disallow_any_generics = True
+disallow_untyped_calls = True
+disallow_untyped_defs = True
+ignore_missing_imports = True
+strict_optional = False

requirements.txt

@@ -0,0 +1,15 @@
+numpy==1.26.4
+gradio==4.44.0
+fastapi<0.113.0
+opencv-python-headless==4.9.0.80
+onnx==1.16.0
+insightface==0.7.3
+albucore==0.0.16
+psutil==5.9.6
+torch==2.1.2
+torchvision==0.16.2
+onnxruntime==1.17.1
+tqdm==4.66.4
+ftfy
+regex
+pyvirtualcam

roop-unleashed-main/.flake8

@@ -0,0 +1,3 @@
+[flake8]
+select = E3, E4, F
+per-file-ignores = roop/core.py:E402

roop-unleashed-main/.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,37 @@
+---
+name: Bug report
+about: Create a report to help us improve
+title: ''
+labels: ''
+assignees: ''
+
+---
+
+**Describe the bug**
+A clear and concise description of what the bug is.
+
+**To Reproduce**
+Steps to reproduce the behavior:
+1. Go to '...'
+2. Click on '....'
+3. Scroll down to '....'
+4. See error
+
+**Details**
+What OS are you using?
+- [ ] Linux
+- [ ] Linux in WSL
+- [ ] Windows 
+- [ ] Mac
+
+Are you using a GPU?
+- [ ] No. CPU FTW
+- [ ] NVIDIA
+- [ ] AMD
+- [ ] Intel
+- [ ] Mac
+
+**Which version of roop unleashed are you using?**
+
+**Screenshots**
+If applicable, add screenshots to help explain your problem.

roop-unleashed-main/.github/workflows/stale.yml

@@ -0,0 +1,29 @@
+# This workflow warns and then closes issues and PRs that have had no activity for a specified amount of time.
+#
+# You can adjust the behavior by modifying this file.
+# For more information, see:
+# https://github.com/actions/stale
+name: Mark stale issues and pull requests
+
+on:
+  schedule:
+  - cron: '32 0 * * *'
+
+jobs:
+  stale:
+
+    runs-on: ubuntu-latest
+    permissions:
+      issues: write
+      pull-requests: write
+
+    steps:
+    - uses: actions/stale@v5
+      with:
+        repo-token: ${{ secrets.GITHUB_TOKEN }}
+        stale-issue-message: 'This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.'
+        stale-pr-message: 'This PR is stale because it has been open 45 days with no activity. Remove stale label or comment or this will be closed in 10 days.'
+        close-issue-message: 'This issue was closed because it has been stalled for 5 days with no activity.'
+        days-before-stale: 30
+        days-before-close: 5
+        days-before-pr-close: -1

roop-unleashed-main/.gitignore

@@ -0,0 +1,15 @@
+.vs
+.idea
+models
+temp
+__pycache__
+*.pth
+/start.bat
+/env
+.vscode
+output
+temp
+config.yaml
+run.bat
+venv
+start.sh

roop-unleashed-main/Dockerfile

@@ -0,0 +1,18 @@
+FROM python:3.11
+
+# making app folder
+WORKDIR /app
+
+# copying files
+COPY . .
+
+# installing requirements
+RUN apt-get update 
+RUN apt-get install ffmpeg -y
+RUN pip install --upgrade pip 
+RUN pip install -r ./requirements.txt 
+
+# launching gradio app 
+ENV GRADIO_SERVER_NAME="0.0.0.0"
+EXPOSE 7860
+ENTRYPOINT python ./run.py

roop-unleashed-main/LICENSE

@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU Affero General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero General Public License from time to time.  Such new versions
+will be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU Affero General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU Affero General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU Affero General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License as published
+    by the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+<https://www.gnu.org/licenses/>.

roop-unleashed-main/README.md

@@ -0,0 +1,253 @@
+# roop-unleashed
+
+[Changelog](#changelog) • [Usage](#usage) • [Wiki](https://github.com/C0untFloyd/roop-unleashed/wiki)
+
+
+Uncensored Deepfakes for images and videos without training and an easy-to-use GUI.
+
+
+![Screen](https://github.com/C0untFloyd/roop-unleashed/assets/131583554/6ee6860d-efbe-4337-8c62-a67598863637)
+
+### Features
+
+- Platform-independant Browser GUI
+- Selection of multiple input/output faces in one go
+- Many different swapping modes, first detected, face selections, by gender
+- Batch processing of images/videos
+- Masking of face occluders using text prompts or automatically
+- Optional Face Upscaler/Restoration using different enhancers
+- Preview swapping from different video frames
+- Live Fake Cam using your webcam
+- Extras Tab for cutting videos etc.
+- Settings - storing configuration for next session
+- Theme Support
+
+and lots more...
+
+
+## Disclaimer
+
+This project is for technical and academic use only.
+Users of this software are expected to use this software responsibly while abiding the local law. If a face of a real person is being used, users are suggested to get consent from the concerned person and clearly mention that it is a deepfake when posting content online. Developers of this software will not be responsible for actions of end-users.
+**Please do not apply it to illegal and unethical scenarios.**
+
+In the event of violation of the legal and ethical requirements of the user's country or region, this code repository is exempt from liability
+
+### Installation
+
+Please refer to the [wiki](https://github.com/C0untFloyd/roop-unleashed/wiki).
+
+#### macOS Installation
+Simply run the following command. It will check and install all dependencies if necessary.
+
+`/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/PJF16/roop-unleashed/master/installer/macOSinstaller.sh)"`
+
+
+
+### Usage
+
+- Windows: run the `windows_run.bat` from the Installer.
+- Linux: `python run.py`
+- macOS: `sh runMacOS.sh`
+- Dockerfile:
+  ```shell
+  docker build -t roop-unleashed . && docker run -t \
+    -p 7860:7860 \
+    -v ./config.yaml:/app/config.yaml \
+    -v ./models:/app/models \
+    -v ./temp:/app/temp \
+    -v ./output:/app/output \
+    roop-unleashed
+  ```
+
+<a target="_blank" href="https://colab.research.google.com/github/C0untFloyd/roop-unleashed/blob/main/roop-unleashed.ipynb">
+  <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
+</a>
+  
+
+Additional commandline arguments are currently unsupported and settings should be done via the UI.
+
+> Note: When you run this program for the first time, it will download some models roughly ~2Gb in size.
+
+
+
+
+### Changelog
+
+**31.12.2024** v4.4.0 Hotfix
+
+Bugfix: Updated Colab to use present Cuda Drivers
+Bugfix: Live-Cam not working because of new face swapper
+Set default swapping model back to Insightface
+
+Happy New Year!
+
+
+**30.12.2024** v4.4.0
+
+- Added random face selection mode
+- Added alternative face swapping model with 128px & 256 px output ([ReSwapper](https://github.com/somanchiu/ReSwapper/tree/main))
+- Video repair added to Extras Tab
+- Updated most packages to newer versions. CUDA >= 12.4 now required!
+- Several minor bugfixes and QoL Changes
+
+
+**28.9.2024** v4.3.1
+
+- Bugfix: Several possible memory leaks
+- Added different output modes, e.g. to virtual cam stream
+- New swapping mode "All input faces"
+- Average total fps displayed and setting for autorun
+
+
+**16.9.2024** v4.2.8
+
+- Bugfix: Starting roop-unleashed without NVIDIA gpu but cuda option enabled
+- Bugfix: Target Faces couldn't be moved left/right
+- Bugfix: Enhancement and upscaling working again in virtual cam
+- Corrupt videos caught when adding to target files, displaying warning msg
+- Source Files Component cleared after face detection to release temp files
+- Added masking and mouth restore options to virtual cam
+
+
+**9.9.2024** v4.2.3
+
+- Hotfix for gradio pydantic issue with fastapi
+- Upgraded to Gradio 4.43 hoping it will fix remaining issues
+- Added new action when no face detected -> use last swapped
+- Specified image format for image controls - opening new tabs on preview images possible again!
+- Hardcoded image output format for livecam to jpeg - might be faster than previous webp
+- Chain events to be only executed if previous was a success
+
+
+**5.9.2024** v4.2.0
+
+- Added ability to move input & target faces order
+- New CLI Arguments override settings
+- Small UI changes to faceswapping tab
+- Added mask option and code for restoration of original mouth area
+- Updated gradio to v4.42.0
+- Added CLI Arguments --server_share and --cuda_device_id
+- Added webp image support
+
+
+**15.07.2024** v4.1.1
+
+- Bugfix: Post-processing after swapping
+
+
+**14.07.2024** v4.1.0
+
+- Added subsample upscaling to increase swap resolution
+- Upgraded gradio
+
+
+**12.05.2024** v4.0.0
+
+- Bugfix: Unnecessary init every frame in live-cam
+- Bugfix: Installer downloading insightface package each run
+- Added xseg masking to live-cam
+- Added realesrganx2 to frame processors
+- Upgraded some requirements
+- Added subtypes and different model support to frame processors
+- Allow frame processors to change resolutions of videos
+- Different OpenCV Cap for MacOS Virtual Cam
+- Added complete frame processing to extras tab
+- Colorize, upscale and misc filters added
+
+
+**22.04.2024** v3.9.0
+
+- Bugfix: Face detection bounding box corrupt values at weird angles
+- Rewrote mask previewing to work with every model
+- Switching mask engines toggles text interactivity
+- Clearing target files, resets face selection dropdown
+- Massive rewrite of swapping architecture, needed for xseg implementation
+- Added DFL Xseg Support for partial face occlusion
+- Face masking only runs when there is a face detected
+- Removed unnecessary toggle checkbox for text masking
+
+
+**22.03.2024** v3.6.5
+
+- Bugfix: Installer pulling latest update on first installation
+- Bugfix: Regression issue, blurring/erosion missing from face swap
+- Exposed erosion and blur amounts to UI
+- Using same values for manual masking too
+
+
+**20.03.2024** v3.6.3
+
+- Bugfix: Workaround for Gradio Slider Change Bug
+- Bugfix: CSS Styling to fix Gradio Image Height Bug
+- Made face swapping mask offsets resolution independant
+- Show offset mask as overlay
+- Changed layout for masking
+
+
+**18.03.2024** v3.6.0
+
+- Updated to Gradio 4.21.0 - requiring many changes under the hood
+- New manual masking (draw the mask yourself)
+- Extras Tab, streamlined cutting/joining videos
+- Re-added face selection by gender (on-demand loading, default turned off)
+- Removed unnecessary activate live-cam option
+- Added time info to preview frame and changed frame slider event to allow faster changes
+
+
+**10.03.2024** v3.5.5
+
+- Bugfix: Installer Path Env
+- Bugfix: file attributes
+- Video processing checks for presence of ffmpeg and displays warning if not found
+- Removed gender + age detection to speed up processing. Option removed from UI
+- Replaced restoreformer with restoreformer++
+- Live Cam recoded to run separate from virtual cam and without blocking controls
+- Swapping with only 1 target face allows selecting from several input faces
+
+
+
+**08.01.2024** v3.5.0
+
+- Bugfix: wrong access options when creating folders
+- New auto rotation of horizontal faces, fixing bad landmark positions (expanded on ![PR 364](https://github.com/C0untFloyd/roop-unleashed/pull/364))
+- Simple VR Option for stereo Images/Movies, best used in selected face mode
+- Added RestoreFormer Enhancer - https://github.com/wzhouxiff/RestoreFormer
+- Bumped up package versions for onnx/Torch etc.   
+
+
+**16.10.2023** v3.3.4
+
+**11.8.2023** v2.7.0
+
+Initial Gradio Version - old TkInter Version now deprecated
+
+- Re-added unified padding to face enhancers
+- Fixed DMDNet for all resolutions
+- Selecting target face now automatically switches swapping mode to selected
+- GPU providers are correctly set using the GUI (needs restart currently)
+- Local output folder can be opened from page
+- Unfinished extras functions disabled for now
+- Installer checks out specific commit, allowing to go back to first install
+- Updated readme for new gradio version
+- Updated Colab
+
+
+# Acknowledgements
+
+Lots of ideas, code or pre-trained models borrowed from the following projects:
+
+https://github.com/deepinsight/insightface<br />
+https://github.com/s0md3v/roop<br />
+https://github.com/AUTOMATIC1111/stable-diffusion-webui<br /> 
+https://github.com/Hillobar/Rope<br />
+https://github.com/TencentARC/GFPGAN<br />   
+https://github.com/kadirnar/codeformer-pip<br />
+https://github.com/csxmli2016/DMDNet<br />
+https://github.com/glucauze/sd-webui-faceswaplab<br />
+https://github.com/ykk648/face_power<br />
+
+<br />
+<br />
+Thanks to all developers!
+

roop-unleashed-main/clip/__init__.py

@@ -0,0 +1 @@
+from .clip import *

roop-unleashed-main/clip/bpe_simple_vocab_16e6.txt.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a
+size 1356917

roop-unleashed-main/clip/clip.py

@@ -0,0 +1,241 @@
+import hashlib
+import os
+import urllib
+import warnings
+from typing import Any, Union, List
+
+import torch
+from PIL import Image
+from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize
+from tqdm import tqdm
+
+from .model import build_model
+from .simple_tokenizer import SimpleTokenizer as _Tokenizer
+
+try:
+    from torchvision.transforms import InterpolationMode
+    BICUBIC = InterpolationMode.BICUBIC
+except ImportError:
+    BICUBIC = Image.BICUBIC
+
+
+
+__all__ = ["available_models", "load", "tokenize"]
+_tokenizer = _Tokenizer()
+
+_MODELS = {
+    "RN50": "https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
+    "RN101": "https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
+    "RN50x4": "https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt",
+    "RN50x16": "https://openaipublic.azureedge.net/clip/models/52378b407f34354e150460fe41077663dd5b39c54cd0bfd2b27167a4a06ec9aa/RN50x16.pt",
+    "RN50x64": "https://openaipublic.azureedge.net/clip/models/be1cfb55d75a9666199fb2206c106743da0f6468c9d327f3e0d0a543a9919d9c/RN50x64.pt",
+    "ViT-B/32": "https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
+    "ViT-B/16": "https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt",
+    "ViT-L/14": "https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt",
+    "ViT-L/14@336px": "https://openaipublic.azureedge.net/clip/models/3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02/ViT-L-14-336px.pt",
+}
+
+
+def _download(url: str, root: str):
+    os.makedirs(root, exist_ok=True)
+    filename = os.path.basename(url)
+
+    expected_sha256 = url.split("/")[-2]
+    download_target = os.path.join(root, filename)
+
+    if os.path.exists(download_target) and not os.path.isfile(download_target):
+        raise RuntimeError(f"{download_target} exists and is not a regular file")
+
+    if os.path.isfile(download_target):
+        if hashlib.sha256(open(download_target, "rb").read()).hexdigest() == expected_sha256:
+            return download_target
+        else:
+            warnings.warn(f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file")
+
+    with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
+        with tqdm(total=int(source.info().get("Content-Length")), ncols=80, unit='iB', unit_scale=True, unit_divisor=1024) as loop:
+            while True:
+                buffer = source.read(8192)
+                if not buffer:
+                    break
+
+                output.write(buffer)
+                loop.update(len(buffer))
+
+    if hashlib.sha256(open(download_target, "rb").read()).hexdigest() != expected_sha256:
+        raise RuntimeError("Model has been downloaded but the SHA256 checksum does not not match")
+
+    return download_target
+
+
+def _convert_image_to_rgb(image):
+    return image.convert("RGB")
+
+
+def _transform(n_px):
+    return Compose([
+        Resize(n_px, interpolation=BICUBIC),
+        CenterCrop(n_px),
+        _convert_image_to_rgb,
+        ToTensor(),
+        Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
+    ])
+
+
+def available_models() -> List[str]:
+    """Returns the names of available CLIP models"""
+    return list(_MODELS.keys())
+
+
+def load(name: str, device: Union[str, torch.device] = "cuda" if torch.cuda.is_available() else "cpu", jit: bool = False, download_root: str = None):
+    """Load a CLIP model
+
+    Parameters
+    ----------
+    name : str
+        A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict
+
+    device : Union[str, torch.device]
+        The device to put the loaded model
+
+    jit : bool
+        Whether to load the optimized JIT model or more hackable non-JIT model (default).
+
+    download_root: str
+        path to download the model files; by default, it uses "~/.cache/clip"
+
+    Returns
+    -------
+    model : torch.nn.Module
+        The CLIP model
+
+    preprocess : Callable[[PIL.Image], torch.Tensor]
+        A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input
+    """
+    if name in _MODELS:
+        model_path = _download(_MODELS[name], download_root or os.path.expanduser("~/.cache/clip"))
+    elif os.path.isfile(name):
+        model_path = name
+    else:
+        raise RuntimeError(f"Model {name} not found; available models = {available_models()}")
+
+    with open(model_path, 'rb') as opened_file:
+        try:
+            # loading JIT archive
+            model = torch.jit.load(opened_file, map_location=device if jit else "cpu").eval()
+            state_dict = None
+        except RuntimeError:
+            # loading saved state dict
+            if jit:
+                warnings.warn(f"File {model_path} is not a JIT archive. Loading as a state dict instead")
+                jit = False
+            state_dict = torch.load(opened_file, map_location="cpu")
+
+    if not jit:
+        model = build_model(state_dict or model.state_dict()).to(device)
+        if str(device) == "cpu":
+            model.float()
+        return model, _transform(model.visual.input_resolution)
+
+    # patch the device names
+    device_holder = torch.jit.trace(lambda: torch.ones([]).to(torch.device(device)), example_inputs=[])
+    device_node = [n for n in device_holder.graph.findAllNodes("prim::Constant") if "Device" in repr(n)][-1]
+
+    def _node_get(node: torch._C.Node, key: str):
+        """Gets attributes of a node which is polymorphic over return type.
+        
+        From https://github.com/pytorch/pytorch/pull/82628
+        """
+        sel = node.kindOf(key)
+        return getattr(node, sel)(key)
+
+    def patch_device(module):
+        try:
+            graphs = [module.graph] if hasattr(module, "graph") else []
+        except RuntimeError:
+            graphs = []
+
+        if hasattr(module, "forward1"):
+            graphs.append(module.forward1.graph)
+
+        for graph in graphs:
+            for node in graph.findAllNodes("prim::Constant"):
+                if "value" in node.attributeNames() and str(_node_get(node, "value")).startswith("cuda"):
+                    node.copyAttributes(device_node)
+
+    model.apply(patch_device)
+    patch_device(model.encode_image)
+    patch_device(model.encode_text)
+
+    # patch dtype to float32 on CPU
+    if str(device) == "cpu":
+        float_holder = torch.jit.trace(lambda: torch.ones([]).float(), example_inputs=[])
+        float_input = list(float_holder.graph.findNode("aten::to").inputs())[1]
+        float_node = float_input.node()
+
+        def patch_float(module):
+            try:
+                graphs = [module.graph] if hasattr(module, "graph") else []
+            except RuntimeError:
+                graphs = []
+
+            if hasattr(module, "forward1"):
+                graphs.append(module.forward1.graph)
+
+            for graph in graphs:
+                for node in graph.findAllNodes("aten::to"):
+                    inputs = list(node.inputs())
+                    for i in [1, 2]:  # dtype can be the second or third argument to aten::to()
+                        if _node_get(inputs[i].node(), "value") == 5:
+                            inputs[i].node().copyAttributes(float_node)
+
+        model.apply(patch_float)
+        patch_float(model.encode_image)
+        patch_float(model.encode_text)
+
+        model.float()
+
+    return model, _transform(model.input_resolution.item())
+
+
+def tokenize(texts: Union[str, List[str]], context_length: int = 77, truncate: bool = False) -> Union[torch.IntTensor, torch.LongTensor]:
+    """
+    Returns the tokenized representation of given input string(s)
+
+    Parameters
+    ----------
+    texts : Union[str, List[str]]
+        An input string or a list of input strings to tokenize
+
+    context_length : int
+        The context length to use; all CLIP models use 77 as the context length
+
+    truncate: bool
+        Whether to truncate the text in case its encoding is longer than the context length
+
+    Returns
+    -------
+    A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length].
+    We return LongTensor when torch version is <1.8.0, since older index_select requires indices to be long.
+    """
+    if isinstance(texts, str):
+        texts = [texts]
+
+    sot_token = _tokenizer.encoder["<|startoftext|>"]
+    eot_token = _tokenizer.encoder["<|endoftext|>"]
+    all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]
+    #if packaging.version.parse(torch.__version__) < packaging.version.parse("1.8.0"):
+    #    result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+    #else:
+    result = torch.zeros(len(all_tokens), context_length, dtype=torch.int)
+
+    for i, tokens in enumerate(all_tokens):
+        if len(tokens) > context_length:
+            if truncate:
+                tokens = tokens[:context_length]
+                tokens[-1] = eot_token
+            else:
+                raise RuntimeError(f"Input {texts[i]} is too long for context length {context_length}")
+        result[i, :len(tokens)] = torch.tensor(tokens)
+
+    return result

roop-unleashed-main/clip/clipseg.py

@@ -0,0 +1,538 @@
+import math
+from os.path import basename, dirname, join, isfile
+import torch
+from torch import nn
+from torch.nn import functional as nnf
+from torch.nn.modules.activation import ReLU
+
+
+def get_prompt_list(prompt):
+    if prompt == 'plain':
+        return ['{}']    
+    elif prompt == 'fixed':
+        return ['a photo of a {}.']
+    elif prompt == 'shuffle':
+        return ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.']
+    elif prompt == 'shuffle+':
+        return ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.',
+                            'a cropped photo of a {}.', 'a good photo of a {}.', 'a photo of one {}.',
+                            'a bad photo of a {}.', 'a photo of the {}.']
+    else:
+        raise ValueError('Invalid value for prompt')        
+
+
+def forward_multihead_attention(x, b, with_aff=False, attn_mask=None):
+    """ 
+    Simplified version of multihead attention (taken from torch source code but without tons of if clauses). 
+    The mlp and layer norm come from CLIP.
+    x: input.
+    b: multihead attention module. 
+    """
+
+    x_ = b.ln_1(x)
+    q, k, v = nnf.linear(x_, b.attn.in_proj_weight, b.attn.in_proj_bias).chunk(3, dim=-1)
+    tgt_len, bsz, embed_dim = q.size()
+
+    head_dim = embed_dim // b.attn.num_heads
+    scaling = float(head_dim) ** -0.5
+
+    q = q.contiguous().view(tgt_len, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
+    k = k.contiguous().view(-1, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
+    v = v.contiguous().view(-1, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
+
+    q = q * scaling
+
+    attn_output_weights = torch.bmm(q, k.transpose(1, 2)) #  n_heads * batch_size, tokens^2, tokens^2
+    if attn_mask is not None:
+
+
+        attn_mask_type, attn_mask = attn_mask
+        n_heads = attn_output_weights.size(0) // attn_mask.size(0)
+        attn_mask = attn_mask.repeat(n_heads, 1)
+        
+        if attn_mask_type == 'cls_token':
+            # the mask only affects similarities compared to the readout-token.
+            attn_output_weights[:, 0, 1:] = attn_output_weights[:, 0, 1:] * attn_mask[None,...]
+            # attn_output_weights[:, 0, 0] = 0*attn_output_weights[:, 0, 0]
+
+        if attn_mask_type == 'all':
+            # print(attn_output_weights.shape, attn_mask[:, None].shape)
+            attn_output_weights[:, 1:, 1:] = attn_output_weights[:, 1:, 1:] * attn_mask[:, None]
+        
+    
+    attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
+
+    attn_output = torch.bmm(attn_output_weights, v)
+    attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+    attn_output = b.attn.out_proj(attn_output)
+
+    x = x + attn_output
+    x = x + b.mlp(b.ln_2(x))
+
+    if with_aff:
+        return x, attn_output_weights
+    else:
+        return x
+
+
+class CLIPDenseBase(nn.Module):
+
+    def __init__(self, version, reduce_cond, reduce_dim, prompt, n_tokens):
+        super().__init__()
+
+        import clip
+
+        # prec = torch.FloatTensor
+        self.clip_model, _ = clip.load(version, device='cpu', jit=False)
+        self.model = self.clip_model.visual
+
+        # if not None, scale conv weights such that we obtain n_tokens.
+        self.n_tokens = n_tokens
+
+        for p in self.clip_model.parameters():
+            p.requires_grad_(False)
+
+        # conditional
+        if reduce_cond is not None:
+            self.reduce_cond = nn.Linear(512, reduce_cond)
+            for p in self.reduce_cond.parameters():
+                p.requires_grad_(False)
+        else:
+            self.reduce_cond = None        
+
+        self.film_mul = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        self.film_add = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        
+        self.reduce = nn.Linear(768, reduce_dim)
+
+        self.prompt_list = get_prompt_list(prompt)     
+
+        # precomputed prompts
+        import pickle
+        if isfile('precomputed_prompt_vectors.pickle'):
+            precomp = pickle.load(open('precomputed_prompt_vectors.pickle', 'rb'))
+            self.precomputed_prompts = {k: torch.from_numpy(v) for k, v in precomp.items()}        
+        else:
+            self.precomputed_prompts = dict()
+    
+    def rescaled_pos_emb(self, new_size):
+        assert len(new_size) == 2
+
+        a = self.model.positional_embedding[1:].T.view(1, 768, *self.token_shape)
+        b = nnf.interpolate(a, new_size, mode='bicubic', align_corners=False).squeeze(0).view(768, new_size[0]*new_size[1]).T
+        return torch.cat([self.model.positional_embedding[:1], b])
+
+    def visual_forward(self, x_inp, extract_layers=(), skip=False, mask=None):
+        
+
+        with torch.no_grad():
+
+            inp_size = x_inp.shape[2:]
+
+            if self.n_tokens is not None:
+                stride2 = x_inp.shape[2] // self.n_tokens
+                conv_weight2 = nnf.interpolate(self.model.conv1.weight, (stride2, stride2), mode='bilinear', align_corners=True)
+                x = nnf.conv2d(x_inp, conv_weight2, bias=self.model.conv1.bias, stride=stride2, dilation=self.model.conv1.dilation)
+            else:
+                x = self.model.conv1(x_inp)  # shape = [*, width, grid, grid]
+
+            x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+            x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+
+            x = torch.cat([self.model.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+
+            standard_n_tokens = 50 if self.model.conv1.kernel_size[0] == 32 else 197
+
+            if x.shape[1] != standard_n_tokens:
+                new_shape = int(math.sqrt(x.shape[1]-1))
+                x = x + self.rescaled_pos_emb((new_shape, new_shape)).to(x.dtype)[None,:,:]
+            else:
+                x = x + self.model.positional_embedding.to(x.dtype)
+
+            x = self.model.ln_pre(x)
+
+            x = x.permute(1, 0, 2)  # NLD -> LND
+
+            activations, affinities = [], []
+            for i, res_block in enumerate(self.model.transformer.resblocks):
+                
+                if mask is not None:
+                    mask_layer, mask_type, mask_tensor = mask
+                    if mask_layer == i or mask_layer == 'all':
+                        # import ipdb; ipdb.set_trace()
+                        size = int(math.sqrt(x.shape[0] - 1))
+                        
+                        attn_mask = (mask_type, nnf.interpolate(mask_tensor.unsqueeze(1).float(), (size, size)).view(mask_tensor.shape[0], size * size))
+                        
+                    else:
+                        attn_mask = None
+                else:
+                    attn_mask = None
+
+                x, aff_per_head = forward_multihead_attention(x, res_block, with_aff=True, attn_mask=attn_mask)
+
+                if i in extract_layers:
+                    affinities += [aff_per_head]
+
+                    #if self.n_tokens is not None:
+                    #    activations += [nnf.interpolate(x, inp_size, mode='bilinear', align_corners=True)]
+                    #else:
+                    activations += [x]
+
+                if len(extract_layers) > 0 and i == max(extract_layers) and skip:
+                    print('early skip')
+                    break
+                
+            x = x.permute(1, 0, 2)  # LND -> NLD
+            x = self.model.ln_post(x[:, 0, :])
+
+            if self.model.proj is not None:
+                x = x @ self.model.proj
+
+            return x, activations, affinities
+
+    def sample_prompts(self, words, prompt_list=None):
+
+        prompt_list = prompt_list if prompt_list is not None else self.prompt_list
+
+        prompt_indices = torch.multinomial(torch.ones(len(prompt_list)), len(words), replacement=True)
+        prompts = [prompt_list[i] for i in prompt_indices]
+        return [promt.format(w) for promt, w in zip(prompts, words)]
+
+    def get_cond_vec(self, conditional, batch_size):
+        # compute conditional from a single string
+        if conditional is not None and type(conditional) == str:
+            cond = self.compute_conditional(conditional)
+            cond = cond.repeat(batch_size, 1)
+
+        # compute conditional from string list/tuple
+        elif conditional is not None and type(conditional) in {list, tuple} and type(conditional[0]) == str:
+            assert len(conditional) == batch_size
+            cond = self.compute_conditional(conditional)
+
+        # use conditional directly
+        elif conditional is not None and type(conditional) == torch.Tensor and conditional.ndim == 2:
+            cond = conditional
+
+        # compute conditional from image
+        elif conditional is not None and type(conditional) == torch.Tensor:
+            with torch.no_grad():
+                cond, _, _ = self.visual_forward(conditional)
+        else:
+            raise ValueError('invalid conditional')
+        return cond   
+
+    def compute_conditional(self, conditional):
+        import clip
+
+        dev = next(self.parameters()).device
+
+        if type(conditional) in {list, tuple}:
+            text_tokens = clip.tokenize(conditional).to(dev)
+            cond = self.clip_model.encode_text(text_tokens)
+        else:
+            if conditional in self.precomputed_prompts:
+                cond = self.precomputed_prompts[conditional].float().to(dev)
+            else:
+                text_tokens = clip.tokenize([conditional]).to(dev)
+                cond = self.clip_model.encode_text(text_tokens)[0]
+        
+        if self.shift_vector is not None:
+            return cond + self.shift_vector
+        else:
+            return cond
+
+
+def clip_load_untrained(version):
+    assert version == 'ViT-B/16'
+    from clip.model import CLIP
+    from clip.clip import _MODELS, _download
+    model = torch.jit.load(_download(_MODELS['ViT-B/16'])).eval()
+    state_dict = model.state_dict()
+
+    vision_width = state_dict["visual.conv1.weight"].shape[0]
+    vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
+    vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
+    grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
+    image_resolution = vision_patch_size * grid_size
+    embed_dim = state_dict["text_projection"].shape[1]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith(f"transformer.resblocks")))
+
+    return CLIP(embed_dim, image_resolution, vision_layers, vision_width, vision_patch_size, 
+        context_length, vocab_size, transformer_width, transformer_heads, transformer_layers)    
+
+
+class CLIPDensePredT(CLIPDenseBase):
+
+    def __init__(self, version='ViT-B/32', extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4, prompt='fixed', 
+                 extra_blocks=0, reduce_cond=None, fix_shift=False,
+                 learn_trans_conv_only=False,  limit_to_clip_only=False, upsample=False, 
+                 add_calibration=False, rev_activations=False, trans_conv=None, n_tokens=None, complex_trans_conv=False):
+        
+        super().__init__(version, reduce_cond, reduce_dim, prompt, n_tokens)
+        # device = 'cpu'
+
+        self.extract_layers = extract_layers
+        self.cond_layer = cond_layer
+        self.limit_to_clip_only = limit_to_clip_only
+        self.process_cond = None
+        self.rev_activations = rev_activations
+        
+        depth = len(extract_layers)
+
+        if add_calibration:
+            self.calibration_conds = 1
+
+        self.upsample_proj = nn.Conv2d(reduce_dim, 1, kernel_size=1) if upsample else None
+
+        self.add_activation1 = True
+
+        self.version = version
+        
+        self.token_shape = {'ViT-B/32': (7, 7), 'ViT-B/16': (14, 14)}[version]
+
+        if fix_shift:
+            # self.shift_vector = nn.Parameter(torch.load(join(dirname(basename(__file__)), 'clip_text_shift_vector.pth')), requires_grad=False)
+            self.shift_vector = nn.Parameter(torch.load(join(dirname(basename(__file__)), 'shift_text_to_vis.pth')), requires_grad=False)
+            # self.shift_vector = nn.Parameter(-1*torch.load(join(dirname(basename(__file__)), 'shift2.pth')), requires_grad=False)
+        else:
+            self.shift_vector = None
+
+        if trans_conv is None:
+            trans_conv_ks = {'ViT-B/32': (32, 32), 'ViT-B/16': (16, 16)}[version]
+        else:
+            # explicitly define transposed conv kernel size
+            trans_conv_ks = (trans_conv, trans_conv)
+
+        if not complex_trans_conv:
+            self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+        else:
+            assert trans_conv_ks[0] == trans_conv_ks[1]
+
+            tp_kernels = (trans_conv_ks[0] // 4, trans_conv_ks[0] // 4)
+
+            self.trans_conv = nn.Sequential(
+                nn.Conv2d(reduce_dim, reduce_dim, kernel_size=3, padding=1),
+                nn.ReLU(),
+                nn.ConvTranspose2d(reduce_dim, reduce_dim // 2, kernel_size=tp_kernels[0], stride=tp_kernels[0]),
+                nn.ReLU(),
+                nn.ConvTranspose2d(reduce_dim // 2, 1, kernel_size=tp_kernels[1], stride=tp_kernels[1]),               
+            )
+
+#        self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+        
+        assert len(self.extract_layers) == depth
+
+        self.reduces = nn.ModuleList([nn.Linear(768, reduce_dim) for _ in range(depth)])
+        self.blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(len(self.extract_layers))])
+        self.extra_blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(extra_blocks)])
+        
+        # refinement and trans conv
+
+        if learn_trans_conv_only:
+            for p in self.parameters():
+                p.requires_grad_(False)
+            
+            for p in self.trans_conv.parameters():
+                p.requires_grad_(True)
+
+        self.prompt_list = get_prompt_list(prompt)
+
+
+    def forward(self, inp_image, conditional=None, return_features=False, mask=None):
+
+        assert type(return_features) == bool
+
+        inp_image = inp_image.to(self.model.positional_embedding.device)
+
+        if mask is not None:
+            raise ValueError('mask not supported')
+
+        # x_inp = normalize(inp_image)
+        x_inp = inp_image
+
+        bs, dev = inp_image.shape[0], x_inp.device
+
+        cond = self.get_cond_vec(conditional, bs)
+
+        visual_q, activations, _ = self.visual_forward(x_inp, extract_layers=[0] + list(self.extract_layers))
+
+        activation1 = activations[0]
+        activations = activations[1:]
+
+        _activations = activations[::-1] if not self.rev_activations else activations
+
+        a = None
+        for i, (activation, block, reduce) in enumerate(zip(_activations, self.blocks, self.reduces)):
+            
+            if a is not None:
+                a = reduce(activation) + a
+            else:
+                a = reduce(activation)
+
+            if i == self.cond_layer:
+                if self.reduce_cond is not None:
+                    cond = self.reduce_cond(cond)
+                
+                a = self.film_mul(cond) * a + self.film_add(cond)
+
+            a = block(a)
+
+        for block in self.extra_blocks:
+            a = a + block(a)
+
+        a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
+
+        size = int(math.sqrt(a.shape[2]))
+
+        a = a.view(bs, a.shape[1], size, size)
+
+        a = self.trans_conv(a)
+
+        if self.n_tokens is not None:
+            a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear', align_corners=True) 
+
+        if self.upsample_proj is not None:
+            a = self.upsample_proj(a)
+            a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear')
+
+        if return_features:
+            return a, visual_q, cond, [activation1] + activations
+        else:
+            return a,
+
+
+
+class CLIPDensePredTMasked(CLIPDensePredT):
+
+    def __init__(self, version='ViT-B/32', extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4, 
+                 prompt='fixed', extra_blocks=0, reduce_cond=None, fix_shift=False, learn_trans_conv_only=False, 
+                 refine=None, limit_to_clip_only=False, upsample=False, add_calibration=False, n_tokens=None):
+
+        super().__init__(version=version, extract_layers=extract_layers, cond_layer=cond_layer, reduce_dim=reduce_dim, 
+                         n_heads=n_heads, prompt=prompt, extra_blocks=extra_blocks, reduce_cond=reduce_cond, 
+                         fix_shift=fix_shift, learn_trans_conv_only=learn_trans_conv_only,
+                         limit_to_clip_only=limit_to_clip_only, upsample=upsample, add_calibration=add_calibration,
+                         n_tokens=n_tokens)
+
+    def visual_forward_masked(self, img_s, seg_s):
+        return super().visual_forward(img_s, mask=('all', 'cls_token', seg_s))
+
+    def forward(self, img_q, cond_or_img_s, seg_s=None, return_features=False):
+
+        if seg_s is None:
+            cond = cond_or_img_s
+        else:
+            img_s = cond_or_img_s
+
+            with torch.no_grad():
+                cond, _, _ = self.visual_forward_masked(img_s, seg_s)
+
+        return super().forward(img_q, cond, return_features=return_features)
+
+
+
+class CLIPDenseBaseline(CLIPDenseBase):
+
+    def __init__(self, version='ViT-B/32', cond_layer=0, 
+                extract_layer=9, reduce_dim=128, reduce2_dim=None, prompt='fixed', 
+                 reduce_cond=None, limit_to_clip_only=False, n_tokens=None):
+        
+        super().__init__(version, reduce_cond, reduce_dim, prompt, n_tokens)
+        device = 'cpu'
+
+        # self.cond_layer = cond_layer
+        self.extract_layer = extract_layer
+        self.limit_to_clip_only = limit_to_clip_only
+        self.shift_vector = None
+
+        self.token_shape = {'ViT-B/32': (7, 7), 'ViT-B/16': (14, 14)}[version]
+        
+        assert reduce2_dim is not None
+
+        self.reduce2 = nn.Sequential(
+            nn.Linear(reduce_dim, reduce2_dim),
+            nn.ReLU(),
+            nn.Linear(reduce2_dim, reduce_dim)
+        )
+        
+        trans_conv_ks = {'ViT-B/32': (32, 32), 'ViT-B/16': (16, 16)}[version]
+        self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+
+
+    def forward(self, inp_image, conditional=None, return_features=False):
+
+        inp_image = inp_image.to(self.model.positional_embedding.device)
+
+        # x_inp = normalize(inp_image)
+        x_inp = inp_image
+
+        bs, dev = inp_image.shape[0], x_inp.device
+
+        cond = self.get_cond_vec(conditional, bs)
+
+        visual_q, activations, affinities = self.visual_forward(x_inp, extract_layers=[self.extract_layer])
+
+        a = activations[0]
+        a = self.reduce(a)
+        a = self.film_mul(cond) * a + self.film_add(cond)
+
+        if self.reduce2 is not None:
+            a = self.reduce2(a)
+
+        # the original model would execute a transformer block here
+
+        a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
+
+        size = int(math.sqrt(a.shape[2]))
+
+        a = a.view(bs, a.shape[1], size, size)
+        a = self.trans_conv(a)
+
+        if return_features:
+            return a, visual_q, cond, activations
+        else:
+            return a,
+
+
+class CLIPSegMultiLabel(nn.Module):
+
+    def __init__(self, model) -> None:
+        super().__init__()
+
+        from third_party.JoEm.data_loader import get_seen_idx, get_unseen_idx, VOC
+
+        self.pascal_classes = VOC
+
+        from clip.clipseg import CLIPDensePredT
+        from general_utils import load_model
+        # self.clipseg = load_model('rd64-vit16-neg0.2-phrasecut', strict=False)
+        self.clipseg = load_model(model, strict=False)
+        
+        self.clipseg.eval()
+
+    def forward(self, x):
+
+        bs = x.shape[0]
+        out = torch.ones(21, bs, 352, 352).to(x.device) * -10
+
+        for class_id, class_name in enumerate(self.pascal_classes):
+        
+            fac = 3 if class_name == 'background' else 1
+
+            with torch.no_grad():
+                pred = torch.sigmoid(self.clipseg(x, class_name)[0][:,0]) * fac
+
+            out[class_id] += pred
+
+
+        out = out.permute(1, 0, 2, 3)
+
+        return out
+
+        # construct output tensor
+                    

roop-unleashed-main/clip/model.py

@@ -0,0 +1,436 @@
+from collections import OrderedDict
+from typing import Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1):
+        super().__init__()
+
+        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
+        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.relu1 = nn.ReLU(inplace=True)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.relu2 = nn.ReLU(inplace=True)
+
+        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
+
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+        self.relu3 = nn.ReLU(inplace=True)
+
+        self.downsample = None
+        self.stride = stride
+
+        if stride > 1 or inplanes != planes * Bottleneck.expansion:
+            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
+            self.downsample = nn.Sequential(OrderedDict([
+                ("-1", nn.AvgPool2d(stride)),
+                ("0", nn.Conv2d(inplanes, planes * self.expansion, 1, stride=1, bias=False)),
+                ("1", nn.BatchNorm2d(planes * self.expansion))
+            ]))
+
+    def forward(self, x: torch.Tensor):
+        identity = x
+
+        out = self.relu1(self.bn1(self.conv1(x)))
+        out = self.relu2(self.bn2(self.conv2(out)))
+        out = self.avgpool(out)
+        out = self.bn3(self.conv3(out))
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu3(out)
+        return out
+
+
+class AttentionPool2d(nn.Module):
+    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2 + 1, embed_dim) / embed_dim ** 0.5)
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.flatten(start_dim=2).permute(2, 0, 1)  # NCHW -> (HW)NC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x[:1], key=x, value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False
+        )
+        return x.squeeze(0)
+
+
+class ModifiedResNet(nn.Module):
+    """
+    A ResNet class that is similar to torchvision's but contains the following changes:
+    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
+    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
+    - The final pooling layer is a QKV attention instead of an average pool
+    """
+
+    def __init__(self, layers, output_dim, heads, input_resolution=224, width=64):
+        super().__init__()
+        self.output_dim = output_dim
+        self.input_resolution = input_resolution
+
+        # the 3-layer stem
+        self.conv1 = nn.Conv2d(3, width // 2, kernel_size=3, stride=2, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(width // 2)
+        self.relu1 = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.relu2 = nn.ReLU(inplace=True)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.relu3 = nn.ReLU(inplace=True)
+        self.avgpool = nn.AvgPool2d(2)
+
+        # residual layers
+        self._inplanes = width  # this is a *mutable* variable used during construction
+        self.layer1 = self._make_layer(width, layers[0])
+        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
+        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
+        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
+
+        embed_dim = width * 32  # the ResNet feature dimension
+        self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim, heads, output_dim)
+
+    def _make_layer(self, planes, blocks, stride=1):
+        layers = [Bottleneck(self._inplanes, planes, stride)]
+
+        self._inplanes = planes * Bottleneck.expansion
+        for _ in range(1, blocks):
+            layers.append(Bottleneck(self._inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        def stem(x):
+            x = self.relu1(self.bn1(self.conv1(x)))
+            x = self.relu2(self.bn2(self.conv2(x)))
+            x = self.relu3(self.bn3(self.conv3(x)))
+            x = self.avgpool(x)
+            return x
+
+        x = x.type(self.conv1.weight.dtype)
+        x = stem(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.attnpool(x)
+
+        return x
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        ret = super().forward(x.type(torch.float32))
+        return ret.type(orig_type)
+
+
+class QuickGELU(nn.Module):
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = LayerNorm(d_model)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, d_model * 4)),
+            ("gelu", QuickGELU()),
+            ("c_proj", nn.Linear(d_model * 4, d_model))
+        ]))
+        self.ln_2 = LayerNorm(d_model)
+        self.attn_mask = attn_mask
+
+    def attention(self, x: torch.Tensor):
+        self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
+        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attention(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])
+
+    def forward(self, x: torch.Tensor):
+        return self.resblocks(x)
+
+
+class VisionTransformer(nn.Module):
+    def __init__(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.output_dim = output_dim
+        self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)
+
+        scale = width ** -0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width))
+        self.ln_pre = LayerNorm(width)
+
+        self.transformer = Transformer(width, layers, heads)
+
+        self.ln_post = LayerNorm(width)
+        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+    def forward(self, x: torch.Tensor):
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+        x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        x = self.ln_post(x[:, 0, :])
+
+        if self.proj is not None:
+            x = x @ self.proj
+
+        return x
+
+
+class CLIP(nn.Module):
+    def __init__(self,
+                 embed_dim: int,
+                 # vision
+                 image_resolution: int,
+                 vision_layers: Union[Tuple[int, int, int, int], int],
+                 vision_width: int,
+                 vision_patch_size: int,
+                 # text
+                 context_length: int,
+                 vocab_size: int,
+                 transformer_width: int,
+                 transformer_heads: int,
+                 transformer_layers: int
+                 ):
+        super().__init__()
+
+        self.context_length = context_length
+
+        if isinstance(vision_layers, (tuple, list)):
+            vision_heads = vision_width * 32 // 64
+            self.visual = ModifiedResNet(
+                layers=vision_layers,
+                output_dim=embed_dim,
+                heads=vision_heads,
+                input_resolution=image_resolution,
+                width=vision_width
+            )
+        else:
+            vision_heads = vision_width // 64
+            self.visual = VisionTransformer(
+                input_resolution=image_resolution,
+                patch_size=vision_patch_size,
+                width=vision_width,
+                layers=vision_layers,
+                heads=vision_heads,
+                output_dim=embed_dim
+            )
+
+        self.transformer = Transformer(
+            width=transformer_width,
+            layers=transformer_layers,
+            heads=transformer_heads,
+            attn_mask=self.build_attention_mask()
+        )
+
+        self.vocab_size = vocab_size
+        self.token_embedding = nn.Embedding(vocab_size, transformer_width)
+        self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width))
+        self.ln_final = LayerNorm(transformer_width)
+
+        self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim))
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+        self.initialize_parameters()
+
+    def initialize_parameters(self):
+        nn.init.normal_(self.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.positional_embedding, std=0.01)
+
+        if isinstance(self.visual, ModifiedResNet):
+            if self.visual.attnpool is not None:
+                std = self.visual.attnpool.c_proj.in_features ** -0.5
+                nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std)
+
+            for resnet_block in [self.visual.layer1, self.visual.layer2, self.visual.layer3, self.visual.layer4]:
+                for name, param in resnet_block.named_parameters():
+                    if name.endswith("bn3.weight"):
+                        nn.init.zeros_(param)
+
+        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
+        attn_std = self.transformer.width ** -0.5
+        fc_std = (2 * self.transformer.width) ** -0.5
+        for block in self.transformer.resblocks:
+            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+
+        if self.text_projection is not None:
+            nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.context_length, self.context_length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    @property
+    def dtype(self):
+        return self.visual.conv1.weight.dtype
+
+    def encode_image(self, image):
+        return self.visual(image.type(self.dtype))
+
+    def encode_text(self, text):
+        x = self.token_embedding(text).type(self.dtype)  # [batch_size, n_ctx, d_model]
+
+        x = x + self.positional_embedding.type(self.dtype)
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.ln_final(x).type(self.dtype)
+
+        # x.shape = [batch_size, n_ctx, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
+
+        return x
+
+    def forward(self, image, text):
+        image_features = self.encode_image(image)
+        text_features = self.encode_text(text)
+
+        # normalized features
+        image_features = image_features / image_features.norm(dim=1, keepdim=True)
+        text_features = text_features / text_features.norm(dim=1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale = self.logit_scale.exp()
+        logits_per_image = logit_scale * image_features @ text_features.t()
+        logits_per_text = logits_per_image.t()
+
+        # shape = [global_batch_size, global_batch_size]
+        return logits_per_image, logits_per_text
+
+
+def convert_weights(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+        if isinstance(l, nn.MultiheadAttention):
+            for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]:
+                tensor = getattr(l, attr)
+                if tensor is not None:
+                    tensor.data = tensor.data.half()
+
+        for name in ["text_projection", "proj"]:
+            if hasattr(l, name):
+                attr = getattr(l, name)
+                if attr is not None:
+                    attr.data = attr.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+
+def build_model(state_dict: dict):
+    vit = "visual.proj" in state_dict
+
+    if vit:
+        vision_width = state_dict["visual.conv1.weight"].shape[0]
+        vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
+        vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
+        grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
+        image_resolution = vision_patch_size * grid_size
+    else:
+        counts: list = [len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]]
+        vision_layers = tuple(counts)
+        vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
+        output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5)
+        vision_patch_size = None
+        assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0]
+        image_resolution = output_width * 32
+
+    embed_dim = state_dict["text_projection"].shape[1]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith("transformer.resblocks")))
+
+    model = CLIP(
+        embed_dim,
+        image_resolution, vision_layers, vision_width, vision_patch_size,
+        context_length, vocab_size, transformer_width, transformer_heads, transformer_layers
+    )
+
+    for key in ["input_resolution", "context_length", "vocab_size"]:
+        if key in state_dict:
+            del state_dict[key]
+
+    convert_weights(model)
+    model.load_state_dict(state_dict)
+    return model.eval()

roop-unleashed-main/clip/simple_tokenizer.py

@@ -0,0 +1,132 @@
+import gzip
+import html
+import os
+from functools import lru_cache
+
+import ftfy
+import regex as re
+
+
+@lru_cache()
+def default_bpe():
+    return os.path.join(os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz")
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r'\s+', ' ', text)
+    text = text.strip()
+    return text
+
+
+class SimpleTokenizer(object):
+    def __init__(self, bpe_path: str = default_bpe()):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        merges = gzip.open(bpe_path).read().decode("utf-8").split('\n')
+        merges = merges[1:49152-256-2+1]
+        merges = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v+'</w>' for v in vocab]
+        for merge in merges:
+            vocab.append(''.join(merge))
+        vocab.extend(['<|startoftext|>', '<|endoftext|>'])
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
+        self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE)
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + ( token[-1] + '</w>',)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token+'</w>'
+
+        while True:
+            bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+
+                if word[i] == first and i < len(word)-1 and word[i+1] == second:
+                    new_word.append(first+second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = ' '.join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = whitespace_clean(basic_clean(text)).lower()
+        for token in re.findall(self.pat, text):
+            token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
+            bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = ''.join([self.decoder[token] for token in tokens])
+        text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ')
+        return text

roop-unleashed-main/clip/vitseg.py

@@ -0,0 +1,286 @@
+import math
+from posixpath import basename, dirname, join
+# import clip
+from clip.model import convert_weights
+import torch
+import json
+from torch import nn
+from torch.nn import functional as nnf
+from torch.nn.modules import activation
+from torch.nn.modules.activation import ReLU
+from torchvision import transforms
+
+normalize = transforms.Normalize(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711))
+
+from torchvision.models import ResNet
+
+
+def process_prompts(conditional, prompt_list, conditional_map):
+    # DEPRECATED
+            
+    # randomly sample a synonym
+    words = [conditional_map[int(i)] for i in conditional]
+    words = [syns[torch.multinomial(torch.ones(len(syns)), 1, replacement=True).item()] for syns in words]
+    words = [w.replace('_', ' ') for w in words]
+
+    if prompt_list is not None:
+        prompt_indices = torch.multinomial(torch.ones(len(prompt_list)), len(words), replacement=True)
+        prompts = [prompt_list[i] for i in prompt_indices]
+    else:
+        prompts = ['a photo of {}'] * (len(words))
+
+    return [promt.format(w) for promt, w in zip(prompts, words)]
+
+
+class VITDenseBase(nn.Module):
+    
+    def rescaled_pos_emb(self, new_size):
+        assert len(new_size) == 2
+
+        a = self.model.positional_embedding[1:].T.view(1, 768, *self.token_shape)
+        b = nnf.interpolate(a, new_size, mode='bicubic', align_corners=False).squeeze(0).view(768, new_size[0]*new_size[1]).T
+        return torch.cat([self.model.positional_embedding[:1], b])
+
+    def visual_forward(self, x_inp, extract_layers=(), skip=False, mask=None):
+        
+        with torch.no_grad():
+
+            x_inp = nnf.interpolate(x_inp, (384, 384))
+
+            x = self.model.patch_embed(x_inp)
+            cls_token = self.model.cls_token.expand(x.shape[0], -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+            if self.model.dist_token is None:
+                x = torch.cat((cls_token, x), dim=1)
+            else:
+                x = torch.cat((cls_token, self.model.dist_token.expand(x.shape[0], -1, -1), x), dim=1)
+            x = self.model.pos_drop(x + self.model.pos_embed)
+
+            activations = []
+            for i, block in enumerate(self.model.blocks):
+                x = block(x)
+
+                if i in extract_layers:
+                    # permute to be compatible with CLIP
+                    activations += [x.permute(1,0,2)]                
+
+            x = self.model.norm(x)
+            x = self.model.head(self.model.pre_logits(x[:, 0]))
+
+            # again for CLIP compatibility
+            # x = x.permute(1, 0, 2)
+
+        return x, activations, None
+
+    def sample_prompts(self, words, prompt_list=None):
+
+        prompt_list = prompt_list if prompt_list is not None else self.prompt_list
+
+        prompt_indices = torch.multinomial(torch.ones(len(prompt_list)), len(words), replacement=True)
+        prompts = [prompt_list[i] for i in prompt_indices]
+        return [promt.format(w) for promt, w in zip(prompts, words)]
+
+    def get_cond_vec(self, conditional, batch_size):
+        # compute conditional from a single string
+        if conditional is not None and type(conditional) == str:
+            cond = self.compute_conditional(conditional)
+            cond = cond.repeat(batch_size, 1)
+
+        # compute conditional from string list/tuple
+        elif conditional is not None and type(conditional) in {list, tuple} and type(conditional[0]) == str:
+            assert len(conditional) == batch_size
+            cond = self.compute_conditional(conditional)
+
+        # use conditional directly
+        elif conditional is not None and type(conditional) == torch.Tensor and conditional.ndim == 2:
+            cond = conditional
+
+        # compute conditional from image
+        elif conditional is not None and type(conditional) == torch.Tensor:
+            with torch.no_grad():
+                cond, _, _ = self.visual_forward(conditional)
+        else:
+            raise ValueError('invalid conditional')
+        return cond   
+
+    def compute_conditional(self, conditional):
+        import clip
+
+        dev = next(self.parameters()).device
+
+        if type(conditional) in {list, tuple}:
+            text_tokens = clip.tokenize(conditional).to(dev)
+            cond = self.clip_model.encode_text(text_tokens)
+        else:
+            if conditional in self.precomputed_prompts:
+                cond = self.precomputed_prompts[conditional].float().to(dev)
+            else:
+                text_tokens = clip.tokenize([conditional]).to(dev)
+                cond = self.clip_model.encode_text(text_tokens)[0]
+        
+        return cond
+
+
+class VITDensePredT(VITDenseBase):
+
+    def __init__(self, extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4, prompt='fixed', 
+                 depth=3, extra_blocks=0, reduce_cond=None, fix_shift=False,
+                 learn_trans_conv_only=False, refine=None, limit_to_clip_only=False, upsample=False, 
+                 add_calibration=False, process_cond=None, not_pretrained=False):
+        super().__init__()
+        # device = 'cpu'
+
+        self.extract_layers = extract_layers
+        self.cond_layer = cond_layer
+        self.limit_to_clip_only = limit_to_clip_only
+        self.process_cond = None
+        
+        if add_calibration:
+            self.calibration_conds = 1
+
+        self.upsample_proj = nn.Conv2d(reduce_dim, 1, kernel_size=1) if upsample else None
+
+        self.add_activation1 = True
+
+        import timm 
+        self.model = timm.create_model('vit_base_patch16_384', pretrained=True)
+        self.model.head = nn.Linear(768, 512 if reduce_cond is None else reduce_cond)
+
+        for p in self.model.parameters():
+            p.requires_grad_(False)
+
+        import clip
+        self.clip_model, _ = clip.load('ViT-B/16', device='cpu', jit=False)
+        # del self.clip_model.visual
+        
+        
+        self.token_shape = (14, 14)
+
+        # conditional
+        if reduce_cond is not None:
+            self.reduce_cond = nn.Linear(512, reduce_cond)
+            for p in self.reduce_cond.parameters():
+                p.requires_grad_(False)
+        else:
+            self.reduce_cond = None
+
+        # self.film = AVAILABLE_BLOCKS['film'](512, 128)
+        self.film_mul = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        self.film_add = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        
+        # DEPRECATED
+        # self.conditional_map = {c['id']: c['synonyms'] for c in json.load(open(cond_map))}
+        
+        assert len(self.extract_layers) == depth
+
+        self.reduces = nn.ModuleList([nn.Linear(768, reduce_dim) for _ in range(depth)])
+        self.blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(len(self.extract_layers))])
+        self.extra_blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(extra_blocks)])
+
+        trans_conv_ks = (16, 16)
+        self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+
+        # refinement and trans conv
+
+        if learn_trans_conv_only:
+            for p in self.parameters():
+                p.requires_grad_(False)
+            
+            for p in self.trans_conv.parameters():
+                p.requires_grad_(True)
+
+        if prompt == 'fixed':
+            self.prompt_list = ['a photo of a {}.']
+        elif prompt == 'shuffle':
+            self.prompt_list = ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.']
+        elif prompt == 'shuffle+':
+            self.prompt_list = ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.',
+                                'a cropped photo of a {}.', 'a good photo of a {}.', 'a photo of one {}.',
+                                'a bad photo of a {}.', 'a photo of the {}.']
+        elif prompt == 'shuffle_clip':
+            from models.clip_prompts import imagenet_templates
+            self.prompt_list = imagenet_templates
+
+        if process_cond is not None:
+            if process_cond == 'clamp' or process_cond[0] == 'clamp':
+
+                val = process_cond[1] if type(process_cond) in {list, tuple} else 0.2
+
+                def clamp_vec(x):
+                    return torch.clamp(x, -val, val)
+
+                self.process_cond = clamp_vec
+
+            elif process_cond.endswith('.pth'):
+                
+                shift = torch.load(process_cond)
+                def add_shift(x):
+                    return x + shift.to(x.device)
+
+                self.process_cond = add_shift
+
+        import pickle
+        precomp = pickle.load(open('precomputed_prompt_vectors.pickle', 'rb'))
+        self.precomputed_prompts = {k: torch.from_numpy(v) for k, v in precomp.items()}
+
+
+    def forward(self, inp_image, conditional=None, return_features=False, mask=None):
+
+        assert type(return_features) == bool
+
+        # inp_image = inp_image.to(self.model.positional_embedding.device)
+
+        if mask is not None:
+            raise ValueError('mask not supported')
+
+        # x_inp = normalize(inp_image)
+        x_inp = inp_image
+
+        bs, dev = inp_image.shape[0], x_inp.device
+
+        inp_image_size = inp_image.shape[2:]
+
+        cond = self.get_cond_vec(conditional, bs)
+
+        visual_q, activations, _ = self.visual_forward(x_inp, extract_layers=[0] + list(self.extract_layers))
+
+        activation1 = activations[0]
+        activations = activations[1:]
+
+        a = None
+        for i, (activation, block, reduce) in enumerate(zip(activations[::-1], self.blocks, self.reduces)):
+            
+            if a is not None:
+                a = reduce(activation) + a
+            else:
+                a = reduce(activation)
+
+            if i == self.cond_layer:
+                if self.reduce_cond is not None:
+                    cond = self.reduce_cond(cond)
+                
+                a = self.film_mul(cond) * a + self.film_add(cond)
+
+            a = block(a)
+
+        for block in self.extra_blocks:
+            a = a + block(a)
+
+        a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
+
+        size = int(math.sqrt(a.shape[2]))
+
+        a = a.view(bs, a.shape[1], size, size)
+
+        if self.trans_conv is not None:
+            a = self.trans_conv(a)
+
+        if self.upsample_proj is not None:
+            a = self.upsample_proj(a)
+            a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear')
+
+        a = nnf.interpolate(a, inp_image_size)
+
+        if return_features:
+            return a, visual_q, cond, [activation1] + activations
+        else:
+            return a,

roop-unleashed-main/config_colab.yaml

@@ -0,0 +1,14 @@
+clear_output: true
+force_cpu: false
+max_threads: 3
+memory_limit: 0
+output_image_format: png
+output_template: '{file}_{time}'
+output_video_codec: libx264
+output_video_format: mp4
+provider: cuda
+selected_theme: Default
+server_name: ''
+server_port: 0
+server_share: true
+video_quality: 14

roop-unleashed-main/installer/installer.py

@@ -0,0 +1,87 @@
+import argparse
+import glob
+import os
+import shutil
+import site
+import subprocess
+import sys
+
+
+script_dir = os.getcwd()
+
+
+def run_cmd(cmd, capture_output=False, env=None):
+    # Run shell commands
+    return subprocess.run(cmd, shell=True, capture_output=capture_output, env=env)
+
+
+def check_env():
+    # If we have access to conda, we are probably in an environment
+    conda_not_exist = run_cmd("conda", capture_output=True).returncode
+    if conda_not_exist:
+        print("Conda is not installed. Exiting...")
+        sys.exit()
+    
+    # Ensure this is a new environment and not the base environment
+    if os.environ["CONDA_DEFAULT_ENV"] == "base":
+        print("Create an environment for this project and activate it. Exiting...")
+        sys.exit()
+
+
+def install_dependencies():
+    global MY_PATH
+
+    # Install Git and clone repo
+    run_cmd("conda install -y -k git")
+    run_cmd("git clone https://github.com/C0untFloyd/roop-unleashed.git")
+    os.chdir(MY_PATH)
+    run_cmd("git checkout 5bfafdc97a0c47b46ec83e6530a57399aaad75d7")
+    # Installs dependencies from requirements.txt
+    run_cmd("python -m pip install -r requirements.txt")
+
+
+
+def update_dependencies():
+    global MY_PATH
+    
+    os.chdir(MY_PATH)
+	# do a hard reset for to update even if there are local changes
+    run_cmd("git fetch --all")
+    run_cmd("git reset --hard origin/main")
+    run_cmd("git pull")
+    # Installs/Updates dependencies from all requirements.txt
+    run_cmd("python -m pip install -r requirements.txt")
+
+
+def start_app():
+    global MY_PATH
+    
+    os.chdir(MY_PATH)
+    # forward commandline arguments
+    sys.argv.pop(0)
+    args = ' '.join(sys.argv)
+    print("Launching App")
+    run_cmd(f'python run.py {args}')
+
+
+if __name__ == "__main__":
+    global MY_PATH
+    
+    MY_PATH = "roop-unleashed"
+
+    
+    # Verifies we are in a conda environment
+    check_env()
+
+    # If webui has already been installed, skip and run
+    if not os.path.exists(MY_PATH):
+        install_dependencies()
+    else:
+        # moved update from batch to here, because of batch limitations
+        updatechoice = input("Check for Updates? [y/n]").lower()
+        if updatechoice == "y":
+           update_dependencies()
+
+    # Run the model with webui
+    os.chdir(script_dir)
+    start_app()

roop-unleashed-main/installer/macOSinstaller.sh

@@ -0,0 +1,73 @@
+#!/bin/bash
+
+# This script checks and installs all dependencies which are needed to run roop-unleashed. After that, it clones the repo.
+# Execute this easily with /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/PJF16/roop-unleashed/master/installer/macOSinstaller.sh)
+
+# Function to check if a command exists
+command_exists() {
+    command -v "$1" >/dev/null 2>&1
+}
+
+echo "Starting check and installation process of dependencies for roop-unleashed"
+
+# Check if Homebrew is installed
+if ! command_exists brew; then
+    echo "Homebrew is not installed. Starting installation..."
+    /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
+else
+    echo "Homebrew is already installed."
+fi
+
+# Update Homebrew
+echo "Updating Homebrew..."
+brew update
+
+# Check if Python 3.11 is installed
+if brew list --versions [email protected] >/dev/null; then
+    echo "Python 3.11 is already installed."
+else
+    echo "Python 3.11 is not installed. Installing it now..."
+    brew install [email protected]
+fi
+
+# Check if [email protected] is installed
+if brew list --versions [email protected] >/dev/null; then
+    echo "[email protected] is already installed."
+else
+    echo "[email protected] is not installed. Installing it now..."
+    brew install [email protected]
+fi
+
+# Check if ffmpeg is installed
+if command_exists ffmpeg; then
+    echo "ffmpeg is already installed."
+else
+    echo "ffmpeg is not installed. Installing it now..."
+    brew install ffmpeg
+fi
+
+# Check if git is installed
+if command_exists git; then
+    echo "git is already installed."
+else
+    echo "git is not installed. Installing it now..."
+    brew install git
+fi
+
+# Clone the repository
+REPO_URL="https://github.com/C0untFloyd/roop-unleashed.git"
+REPO_NAME="roop-unleashed"
+
+echo "Cloning the repository $REPO_URL..."
+git clone $REPO_URL
+
+# Check if the repository was cloned successfully
+if [ -d "$REPO_NAME" ]; then
+    echo "Repository cloned successfully. Changing into directory $REPO_NAME..."
+    cd "$REPO_NAME"
+else
+    echo "Failed to clone the repository."
+fi
+
+echo "Check and installation process completed."
+

roop-unleashed-main/installer/windows_run.bat

@@ -0,0 +1,95 @@
+@echo off
+
+REM No CLI arguments supported anymore
+set COMMANDLINE_ARGS=
+
+cd /D "%~dp0"
+
+echo "%CD%"| findstr /C:" " >nul && echo This script relies on Miniconda which can not be silently installed under a path with spaces. && goto end
+
+set PATH=%PATH%;%SystemRoot%\system32
+
+@rem config
+set INSTALL_DIR=%cd%\installer_files
+set CONDA_ROOT_PREFIX=%cd%\installer_files\conda
+set INSTALL_ENV_DIR=%cd%\installer_files\env
+set MINICONDA_DOWNLOAD_URL=https://repo.anaconda.com/miniconda/Miniconda3-latest-Windows-x86_64.exe
+set FFMPEG_DOWNLOAD_URL=https://github.com/GyanD/codexffmpeg/releases/download/7.1/ffmpeg-7.1-essentials_build.zip
+set INSTALL_FFMPEG_DIR=%cd%\installer_files\ffmpeg
+set INSIGHTFACE_PACKAGE_URL=https://github.com/C0untFloyd/roop-unleashed/releases/download/3.6.6/insightface-0.7.3-cp310-cp310-win_amd64.whl
+set INSIGHTFACE_PACKAGE_PATH=%INSTALL_DIR%\insightface-0.7.3-cp310-cp310-win_amd64.whl
+
+set conda_exists=F
+set ffmpeg_exists=F
+
+@rem figure out whether git and conda needs to be installed
+call "%CONDA_ROOT_PREFIX%\_conda.exe" --version >nul 2>&1
+if "%ERRORLEVEL%" EQU "0" set conda_exists=T
+
+@rem Check if FFmpeg is already in PATH
+where ffmpeg >nul 2>&1
+if "%ERRORLEVEL%" EQU "0" (
+    echo FFmpeg is already installed.
+    set ffmpeg_exists=T
+)
+
+@rem (if necessary) install git and conda into a contained environment
+
+@rem download conda
+if "%conda_exists%" == "F" (
+    echo Downloading Miniconda from %MINICONDA_DOWNLOAD_URL% to %INSTALL_DIR%\miniconda_installer.exe
+    mkdir "%INSTALL_DIR%"
+    call curl -Lk "%MINICONDA_DOWNLOAD_URL%" > "%INSTALL_DIR%\miniconda_installer.exe" || ( echo. && echo Miniconda failed to download. && goto end )
+    echo Installing Miniconda to %CONDA_ROOT_PREFIX%
+    start /wait "" "%INSTALL_DIR%\miniconda_installer.exe" /InstallationType=JustMe /NoShortcuts=1 /AddToPath=0 /RegisterPython=0 /NoRegistry=1 /S /D=%CONDA_ROOT_PREFIX%
+
+    @rem test the conda binary
+    echo Miniconda version:
+    call "%CONDA_ROOT_PREFIX%\_conda.exe" --version || ( echo. && echo Miniconda not found. && goto end )
+)
+
+@rem create the installer env
+if not exist "%INSTALL_ENV_DIR%" (
+    echo Creating Conda Environment
+    call "%CONDA_ROOT_PREFIX%\_conda.exe" create --no-shortcuts -y -k --prefix "%INSTALL_ENV_DIR%" python=3.10 || ( echo. && echo ERROR: Conda environment creation failed. && goto end )
+    @rem check if conda environment was actually created
+    if not exist "%INSTALL_ENV_DIR%\python.exe" ( echo. && echo ERROR: Conda environment is empty. && goto end )
+    @rem activate installer env
+    call "%CONDA_ROOT_PREFIX%\condabin\conda.bat" activate "%INSTALL_ENV_DIR%" || ( echo. && echo ERROR: Miniconda hook not found. && goto end )
+    @rem Download insightface package
+    echo Downloading insightface package from %INSIGHTFACE_PACKAGE_URL% to %INSIGHTFACE_PACKAGE_PATH%
+    call curl -Lk "%INSIGHTFACE_PACKAGE_URL%" > "%INSIGHTFACE_PACKAGE_PATH%" || ( echo. && echo ERROR: Insightface package failed to download. && goto end )
+    @rem install insightface package using pip
+    echo Installing insightface package
+    call pip install "%INSIGHTFACE_PACKAGE_PATH%" || ( echo. && echo ERROR: Insightface package installation failed. && goto end )
+)
+
+@rem Download and install FFmpeg if not already installed
+if "%ffmpeg_exists%" == "F" (
+    if not exist "%INSTALL_FFMPEG_DIR%" (
+        echo Downloading ffmpeg from %FFMPEG_DOWNLOAD_URL% to %INSTALL_DIR%
+        call curl -Lk "%FFMPEG_DOWNLOAD_URL%" > "%INSTALL_DIR%\ffmpeg.zip" || ( echo. && echo ffmpeg failed to download. && goto end )
+        call powershell -command "Expand-Archive -Force '%INSTALL_DIR%\ffmpeg.zip' '%INSTALL_DIR%\'"
+        cd "%INSTALL_DIR%"
+        move ffmpeg-* ffmpeg
+        setx PATH "%INSTALL_FFMPEG_DIR%\bin\;%PATH%"
+        echo To use videos, you need to restart roop after this installation.
+        cd ..
+    )
+) else (
+    echo Skipping FFmpeg installation as it is already available.
+)
+
+@rem setup installer env
+@rem check if conda environment was actually created
+if not exist "%INSTALL_ENV_DIR%\python.exe" ( echo. && echo ERROR: Conda environment is empty. && goto end )
+@rem activate installer env
+call "%CONDA_ROOT_PREFIX%\condabin\conda.bat" activate "%INSTALL_ENV_DIR%" || ( echo. && echo ERROR: Miniconda hook not found. && goto end )
+echo Launching roop unleashed
+call python installer.py %COMMANDLINE_ARGS%
+
+echo.
+echo Done!
+
+:end
+pause

roop-unleashed-main/mypy.ini

@@ -0,0 +1,7 @@
+[mypy]
+check_untyped_defs = True
+disallow_any_generics = True
+disallow_untyped_calls = True
+disallow_untyped_defs = True
+ignore_missing_imports = True
+strict_optional = False

roop-unleashed-main/requirements.txt

@@ -0,0 +1,19 @@
+--extra-index-url https://download.pytorch.org/whl/cu124
+numpy==1.26.4
+gradio==5.9.1
+opencv-python-headless==4.10.0.84
+onnx==1.16.1
+insightface==0.7.3
+albucore==0.0.16
+psutil==5.9.6
+torch==2.5.1+cu124; sys_platform != 'darwin'
+torch==2.5.1; sys_platform == 'darwin'
+torchvision==0.20.1+cu124; sys_platform != 'darwin'
+torchvision==0.20.1; sys_platform == 'darwin'
+onnxruntime==1.20.1; sys_platform == 'darwin' and platform_machine != 'arm64'
+onnxruntime-silicon==1.20.1; sys_platform == 'darwin' and platform_machine == 'arm64'
+onnxruntime-gpu==1.20.1; sys_platform != 'darwin'
+tqdm==4.66.4
+ftfy
+regex
+pyvirtualcam

roop-unleashed-main/roop-unleashed.ipynb

@@ -0,0 +1,166 @@
+{
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "G9BdiCppV6AS"
+      },
+      "source": [
+        "# Colab for roop-unleashed - Gradio version\n",
+        "https://github.com/C0untFloyd/roop-unleashed\n"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "CanIXgLJgaOj"
+      },
+      "source": [
+        "Install CUDA 12.6 on Google Cloud Compute\n",
+        "(currently unnecessary because the latest 12.x should be already installed)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "96GE4UgYg3Ej"
+      },
+      "outputs": [],
+      "source": [
+        "# don't run this cell if you know that there is at least Cuda 12.4 installed\n",
+        "!apt-get -y update\n",
+        "!apt-get -y install cuda-toolkit-12-6\n",
+        "import os\n",
+        "os.environ[\"LD_LIBRARY_PATH\"] += \":\" + \"/usr/local/cuda-12/lib64\"\n",
+        "os.environ[\"LD_LIBRARY_PATH\"] += \":\" + \"/usr/local/cuda-12.6/lib64\""
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "0ZYRNb0AWLLW"
+      },
+      "source": [
+        "Installing & preparing requirements"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "t1yPuhdySqCq"
+      },
+      "outputs": [],
+      "source": [
+        "!git clone https://github.com/C0untFloyd/roop-unleashed.git\n",
+        "%cd roop-unleashed\n",
+        "!mv config_colab.yaml config.yaml\n",
+        "!pip install -r requirements.txt"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "u_4JQiSlV9Fi"
+      },
+      "source": [
+        "Running roop-unleashed with default config"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "Is6U2huqSzLE"
+      },
+      "outputs": [],
+      "source": [
+        "!python run.py"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "UdQ1VHdI8lCf"
+      },
+      "source": [
+        "### Download generated images folder\n",
+        "(only needed if you want to zip the generated output)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/",
+          "height": 17
+        },
+        "id": "oYjWveAmw10X",
+        "outputId": "5b4c3650-f951-434a-c650-5525a8a70c1e"
+      },
+      "outputs": [
+        {
+          "data": {
+            "application/javascript": "\n    async function download(id, filename, size) {\n      if (!google.colab.kernel.accessAllowed) {\n        return;\n      }\n      const div = document.createElement('div');\n      const label = document.createElement('label');\n      label.textContent = `Downloading \"${filename}\": `;\n      div.appendChild(label);\n      const progress = document.createElement('progress');\n      progress.max = size;\n      div.appendChild(progress);\n      document.body.appendChild(div);\n\n      const buffers = [];\n      let downloaded = 0;\n\n      const channel = await google.colab.kernel.comms.open(id);\n      // Send a message to notify the kernel that we're ready.\n      channel.send({})\n\n      for await (const message of channel.messages) {\n        // Send a message to notify the kernel that we're ready.\n        channel.send({})\n        if (message.buffers) {\n          for (const buffer of message.buffers) {\n            buffers.push(buffer);\n            downloaded += buffer.byteLength;\n            progress.value = downloaded;\n          }\n        }\n      }\n      const blob = new Blob(buffers, {type: 'application/binary'});\n      const a = document.createElement('a');\n      a.href = window.URL.createObjectURL(blob);\n      a.download = filename;\n      div.appendChild(a);\n      a.click();\n      div.remove();\n    }\n  ",
+            "text/plain": [
+              "<IPython.core.display.Javascript object>"
+            ]
+          },
+          "metadata": {},
+          "output_type": "display_data"
+        },
+        {
+          "data": {
+            "application/javascript": "download(\"download_789eab11-93d2-4880-adf3-6aceee0cc5f9\", \"fake_output.zip.zip\", 80125)",
+            "text/plain": [
+              "<IPython.core.display.Javascript object>"
+            ]
+          },
+          "metadata": {},
+          "output_type": "display_data"
+        }
+      ],
+      "source": [
+        "import shutil\n",
+        "import os\n",
+        "from google.colab import files\n",
+        "\n",
+        "def zip_directory(directory_path, zip_path):\n",
+        "    shutil.make_archive(zip_path, 'zip', directory_path)\n",
+        "\n",
+        "# Set the directory path you want to download\n",
+        "directory_path = '/content/roop-unleashed/output'\n",
+        "\n",
+        "# Set the zip file name\n",
+        "zip_filename = 'fake_output.zip'\n",
+        "\n",
+        "# Zip the directory\n",
+        "zip_directory(directory_path, zip_filename)\n",
+        "\n",
+        "# Download the zip file\n",
+        "files.download(zip_filename+'.zip')\n"
+      ]
+    }
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "collapsed_sections": [
+        "UdQ1VHdI8lCf"
+      ],
+      "gpuType": "T4",
+      "provenance": []
+    },
+    "kernelspec": {
+      "display_name": "Python 3",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
+}

roop-unleashed-main/roop/FaceSet.py

@@ -0,0 +1,20 @@
+import numpy as np
+
+class FaceSet:
+    faces = []
+    ref_images = []
+    embedding_average = 'None'
+    embeddings_backup = None
+
+    def __init__(self):
+        self.faces = []
+        self.ref_images = []
+        self.embeddings_backup = None
+
+    def AverageEmbeddings(self):
+        if len(self.faces) > 1 and self.embeddings_backup is None:
+            self.embeddings_backup = self.faces[0]['embedding']
+            embeddings = [face.embedding for face in self.faces]
+
+            self.faces[0]['embedding'] = np.mean(embeddings, axis=0)
+            # try median too?

roop-unleashed-main/roop/ProcessEntry.py

@@ -0,0 +1,7 @@
+class ProcessEntry:
+    def __init__(self, filename: str, start: int, end: int, fps: float):
+        self.filename = filename
+        self.finalname = None
+        self.startframe = start
+        self.endframe = end
+        self.fps = fps

roop-unleashed-main/roop/ProcessMgr.py

@@ -0,0 +1,911 @@
+import os
+import cv2 
+import numpy as np
+import psutil
+
+from roop.ProcessOptions import ProcessOptions
+
+from roop.face_util import get_first_face, get_all_faces, rotate_anticlockwise, rotate_clockwise, clamp_cut_values
+from roop.utilities import compute_cosine_distance, get_device, str_to_class, shuffle_array
+import roop.vr_util as vr
+
+from typing import Any, List, Callable
+from roop.typing import Frame, Face
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from threading import Thread, Lock
+from queue import Queue
+from tqdm import tqdm
+from roop.ffmpeg_writer import FFMPEG_VideoWriter
+from roop.StreamWriter import StreamWriter
+import roop.globals
+
+
+
+# Poor man's enum to be able to compare to int
+class eNoFaceAction():
+    USE_ORIGINAL_FRAME = 0
+    RETRY_ROTATED = 1
+    SKIP_FRAME = 2
+    SKIP_FRAME_IF_DISSIMILAR = 3,
+    USE_LAST_SWAPPED = 4
+
+
+
+def create_queue(temp_frame_paths: List[str]) -> Queue[str]:
+    queue: Queue[str] = Queue()
+    for frame_path in temp_frame_paths:
+        queue.put(frame_path)
+    return queue
+
+
+def pick_queue(queue: Queue[str], queue_per_future: int) -> List[str]:
+    queues = []
+    for _ in range(queue_per_future):
+        if not queue.empty():
+            queues.append(queue.get())
+    return queues
+
+
+
+class ProcessMgr():
+    input_face_datas = []
+    target_face_datas = []
+
+    imagemask = None
+
+    processors = []
+    options : ProcessOptions = None
+    
+    num_threads = 1
+    current_index = 0
+    processing_threads = 1
+    buffer_wait_time = 0.1
+
+    lock = Lock()
+
+    frames_queue = None
+    processed_queue = None
+
+    videowriter= None
+    streamwriter = None
+
+    progress_gradio = None
+    total_frames = 0
+
+    num_frames_no_face = 0
+    last_swapped_frame = None
+
+    output_to_file = None
+    output_to_cam = None
+
+
+    plugins =  { 
+    'faceswap'          : 'FaceSwapInsightFace',
+    'mask_clip2seg'     : 'Mask_Clip2Seg',
+    'mask_xseg'         : 'Mask_XSeg',
+    'codeformer'        : 'Enhance_CodeFormer',
+    'gfpgan'            : 'Enhance_GFPGAN',
+    'dmdnet'            : 'Enhance_DMDNet',
+    'gpen'              : 'Enhance_GPEN',
+    'restoreformer++'   : 'Enhance_RestoreFormerPPlus',
+    'colorizer'         : 'Frame_Colorizer',
+    'filter_generic'    : 'Frame_Filter',
+    'removebg'          : 'Frame_Masking',
+    'upscale'           : 'Frame_Upscale'
+    }
+
+    def __init__(self, progress):
+        if progress is not None:
+            self.progress_gradio = progress
+
+    def reuseOldProcessor(self, name:str):
+        for p in self.processors:
+            if p.processorname == name:
+                return p
+            
+        return None
+
+
+    def initialize(self, input_faces, target_faces, options):
+        self.input_face_datas = input_faces
+        self.target_face_datas = target_faces
+        self.num_frames_no_face = 0
+        self.last_swapped_frame = None
+        self.options = options
+        devicename = get_device()
+
+        roop.globals.g_desired_face_analysis=["landmark_3d_68", "landmark_2d_106","detection","recognition"]
+        if options.swap_mode == "all_female" or options.swap_mode == "all_male":
+            roop.globals.g_desired_face_analysis.append("genderage")
+        elif options.swap_mode == "all_random":
+            # don't modify original list
+            self.input_face_datas = input_faces.copy()
+            shuffle_array(self.input_face_datas)
+
+
+        for p in self.processors:
+            newp = next((x for x in options.processors.keys() if x == p.processorname), None)
+            if newp is None:
+                p.Release()
+                del p
+
+        newprocessors = []
+        for key, extoption in options.processors.items():
+            p = self.reuseOldProcessor(key)
+            if p is None:
+                classname = self.plugins[key]
+                module = 'roop.processors.' + classname
+                p = str_to_class(module, classname)
+            if p is not None:
+                extoption.update({"devicename": devicename})
+                if p.type == "swap":
+                    if self.options.swap_modelname == "InSwapper 128":
+                        extoption.update({"modelname": "inswapper_128.onnx"})
+                    elif self.options.swap_modelname == "ReSwapper 128":
+                        extoption.update({"modelname": "reswapper_128.onnx"})
+                    elif self.options.swap_modelname == "ReSwapper 256":
+                        extoption.update({"modelname": "reswapper_256.onnx"})
+
+                p.Initialize(extoption)
+                newprocessors.append(p)
+            else:
+                print(f"Not using {module}")
+        self.processors = newprocessors
+
+
+
+        if isinstance(self.options.imagemask, dict) and self.options.imagemask.get("layers") and len(self.options.imagemask["layers"]) > 0:
+            self.options.imagemask  = self.options.imagemask.get("layers")[0]
+            # Get rid of alpha
+            self.options.imagemask = cv2.cvtColor(self.options.imagemask, cv2.COLOR_RGBA2GRAY)
+            if np.any(self.options.imagemask):
+                mo = self.input_face_datas[0].faces[0].mask_offsets
+                self.options.imagemask = self.blur_area(self.options.imagemask, mo[4], mo[5])
+                self.options.imagemask = self.options.imagemask.astype(np.float32) / 255
+                self.options.imagemask = cv2.cvtColor(self.options.imagemask, cv2.COLOR_GRAY2RGB)
+            else:
+                self.options.imagemask = None
+
+        self.options.frame_processing = False
+        for p in self.processors:
+            if p.type.startswith("frame_"):
+                self.options.frame_processing = True
+
+            
+ 
+
+
+
+    def run_batch(self, source_files, target_files, threads:int = 1):
+        progress_bar_format = '{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}<{remaining}, {rate_fmt}{postfix}]'
+        self.total_frames = len(source_files)
+        self.num_threads = threads
+        with tqdm(total=self.total_frames, desc='Processing', unit='frame', dynamic_ncols=True, bar_format=progress_bar_format) as progress:
+            with ThreadPoolExecutor(max_workers=threads) as executor:
+                futures = []
+                queue = create_queue(source_files)
+                queue_per_future = max(len(source_files) // threads, 1)
+                while not queue.empty():
+                    future = executor.submit(self.process_frames, source_files, target_files, pick_queue(queue, queue_per_future), lambda: self.update_progress(progress))
+                    futures.append(future)
+                for future in as_completed(futures):
+                    future.result()
+
+
+    def process_frames(self, source_files: List[str], target_files: List[str], current_files, update: Callable[[], None]) -> None:
+        for f in current_files:
+            if not roop.globals.processing:
+                return
+            
+            # Decode the byte array into an OpenCV image
+            temp_frame = cv2.imdecode(np.fromfile(f, dtype=np.uint8), cv2.IMREAD_COLOR)
+            if temp_frame is not None:
+                if self.options.frame_processing:
+                    for p in self.processors:
+                        frame = p.Run(temp_frame)
+                    resimg = frame
+                else:
+                    resimg = self.process_frame(temp_frame)
+                if resimg is not None:
+                    i = source_files.index(f)
+                    # Also let numpy write the file to support utf-8/16 filenames
+                    cv2.imencode(f'.{roop.globals.CFG.output_image_format}',resimg)[1].tofile(target_files[i])
+            if update:
+                update()
+
+
+
+    def read_frames_thread(self, cap, frame_start, frame_end, num_threads):
+        num_frame = 0
+        total_num = frame_end - frame_start
+        if frame_start > 0:
+            cap.set(cv2.CAP_PROP_POS_FRAMES,frame_start)
+
+        while True and roop.globals.processing:
+            ret, frame = cap.read()
+            if not ret:
+                break
+                
+            self.frames_queue[num_frame % num_threads].put(frame, block=True)
+            num_frame += 1
+            if num_frame == total_num:
+                break
+
+        for i in range(num_threads):
+            self.frames_queue[i].put(None)
+
+
+
+    def process_videoframes(self, threadindex, progress) -> None:
+        while True:
+            frame = self.frames_queue[threadindex].get()
+            if frame is None:
+                self.processing_threads -= 1
+                self.processed_queue[threadindex].put((False, None))
+                return
+            else:
+                if self.options.frame_processing:
+                    for p in self.processors:
+                        frame = p.Run(frame)
+                    resimg = frame
+                else:                            
+                    resimg = self.process_frame(frame)
+                self.processed_queue[threadindex].put((True, resimg))
+                del frame
+                progress()
+
+
+    def write_frames_thread(self):
+        nextindex = 0
+        num_producers = self.num_threads
+        
+        while True:
+            process, frame = self.processed_queue[nextindex % self.num_threads].get()
+            nextindex += 1
+            if frame is not None:
+                if self.output_to_file:
+                    self.videowriter.write_frame(frame)
+                if self.output_to_cam:
+                    self.streamwriter.WriteToStream(frame)
+                del frame
+            elif process == False:
+                num_producers -= 1
+                if num_producers < 1:
+                    return
+            
+
+
+    def run_batch_inmem(self, output_method, source_video, target_video, frame_start, frame_end, fps, threads:int = 1):
+        if len(self.processors) < 1:
+            print("No processor defined!")
+            return
+
+        cap = cv2.VideoCapture(source_video)
+        # frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+        frame_count = (frame_end - frame_start) + 1
+        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+        processed_resolution = None
+        for p in self.processors:
+            if hasattr(p, 'getProcessedResolution'):
+                processed_resolution = p.getProcessedResolution(width, height)
+                print(f"Processed resolution: {processed_resolution}")
+        if processed_resolution is not None:
+            width = processed_resolution[0]
+            height = processed_resolution[1]
+
+
+        self.total_frames = frame_count
+        self.num_threads = threads
+
+        self.processing_threads = self.num_threads
+        self.frames_queue = []
+        self.processed_queue = []
+        for _ in range(threads):
+            self.frames_queue.append(Queue(1))
+            self.processed_queue.append(Queue(1))
+
+        self.output_to_file = output_method != "Virtual Camera"
+        self.output_to_cam = output_method == "Virtual Camera" or output_method == "Both"
+
+        if self.output_to_file:
+            self.videowriter = FFMPEG_VideoWriter(target_video, (width, height), fps, codec=roop.globals.video_encoder, crf=roop.globals.video_quality, audiofile=None)
+        if self.output_to_cam:
+            self.streamwriter = StreamWriter((width, height), int(fps))
+
+        readthread = Thread(target=self.read_frames_thread, args=(cap, frame_start, frame_end, threads))
+        readthread.start()
+
+        writethread = Thread(target=self.write_frames_thread)
+        writethread.start()
+
+        progress_bar_format = '{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}<{remaining}, {rate_fmt}{postfix}]'
+        with tqdm(total=self.total_frames, desc='Processing', unit='frames', dynamic_ncols=True, bar_format=progress_bar_format) as progress:
+            with ThreadPoolExecutor(thread_name_prefix='swap_proc', max_workers=self.num_threads) as executor:
+                futures = []
+                
+                for threadindex in range(threads):
+                    future = executor.submit(self.process_videoframes, threadindex, lambda: self.update_progress(progress))
+                    futures.append(future)
+                
+                for future in as_completed(futures):
+                    future.result()
+        # wait for the task to complete
+        readthread.join()
+        writethread.join()
+        cap.release()
+        if self.output_to_file:
+            self.videowriter.close()
+        if self.output_to_cam:
+            self.streamwriter.Close()
+
+        self.frames_queue.clear()
+        self.processed_queue.clear()
+
+
+
+
+    def update_progress(self, progress: Any = None) -> None:
+        process = psutil.Process(os.getpid())
+        memory_usage = process.memory_info().rss / 1024 / 1024 / 1024
+        progress.set_postfix({
+            'memory_usage': '{:.2f}'.format(memory_usage).zfill(5) + 'GB',
+            'execution_threads': self.num_threads
+        })
+        progress.update(1)
+        if self.progress_gradio is not None:
+            self.progress_gradio((progress.n, self.total_frames), desc='Processing', total=self.total_frames, unit='frames')
+
+
+
+    def process_frame(self, frame:Frame):
+        if len(self.input_face_datas) < 1 and not self.options.show_face_masking:
+            return frame
+        temp_frame = frame.copy()
+        num_swapped, temp_frame = self.swap_faces(frame, temp_frame)
+        if num_swapped > 0:
+            if roop.globals.no_face_action == eNoFaceAction.SKIP_FRAME_IF_DISSIMILAR:
+                if len(self.input_face_datas) > num_swapped:
+                    return None
+            self.num_frames_no_face = 0
+            self.last_swapped_frame = temp_frame.copy()
+            return temp_frame
+        if roop.globals.no_face_action == eNoFaceAction.USE_LAST_SWAPPED:
+            if self.last_swapped_frame is not None and self.num_frames_no_face < self.options.max_num_reuse_frame:
+                self.num_frames_no_face += 1
+                return self.last_swapped_frame.copy()
+            return frame
+
+        elif roop.globals.no_face_action == eNoFaceAction.USE_ORIGINAL_FRAME:
+            return frame
+        if roop.globals.no_face_action == eNoFaceAction.SKIP_FRAME:
+            #This only works with in-mem processing, as it simply skips the frame.
+            #For 'extract frames' it simply leaves the unprocessed frame unprocessed and it gets used in the final output by ffmpeg.
+            #If we could delete that frame here, that'd work but that might cause ffmpeg to fail unless the frames are renamed, and I don't think we have the info on what frame it actually is?????
+            #alternatively, it could mark all the necessary frames for deletion, delete them at the end, then rename the remaining frames that might work?
+            return None
+        else:
+            return self.retry_rotated(frame)
+
+    def retry_rotated(self, frame):
+        copyframe = frame.copy()
+        copyframe = rotate_clockwise(copyframe)
+        temp_frame = copyframe.copy()
+        num_swapped, temp_frame = self.swap_faces(copyframe, temp_frame)
+        if num_swapped > 0:
+            return rotate_anticlockwise(temp_frame)
+        
+        copyframe = frame.copy()
+        copyframe = rotate_anticlockwise(copyframe)
+        temp_frame = copyframe.copy()
+        num_swapped, temp_frame = self.swap_faces(copyframe, temp_frame)
+        if num_swapped > 0:
+            return rotate_clockwise(temp_frame)
+        del copyframe
+        return frame
+        
+
+
+    def swap_faces(self, frame, temp_frame):
+        num_faces_found = 0
+
+        if self.options.swap_mode == "first":
+            face = get_first_face(frame)
+
+            if face is None:
+                return num_faces_found, frame
+            
+            num_faces_found += 1
+            temp_frame = self.process_face(self.options.selected_index, face, temp_frame)
+            del face
+
+        else:
+            faces = get_all_faces(frame)
+            if faces is None:
+                return num_faces_found, frame
+            
+            if self.options.swap_mode == "all":
+                for face in faces:
+                    num_faces_found += 1
+                    temp_frame = self.process_face(self.options.selected_index, face, temp_frame)
+
+            elif self.options.swap_mode == "all_input" or self.options.swap_mode == "all_random":
+                for i,face in enumerate(faces):
+                    num_faces_found += 1
+                    if i < len(self.input_face_datas):
+                        temp_frame = self.process_face(i, face, temp_frame)
+                    else:
+                        break
+            
+            elif self.options.swap_mode == "selected":
+                num_targetfaces = len(self.target_face_datas) 
+                use_index = num_targetfaces == 1
+                for i,tf in enumerate(self.target_face_datas):
+                    for face in faces:
+                        if compute_cosine_distance(tf.embedding, face.embedding) <= self.options.face_distance_threshold:
+                            if i < len(self.input_face_datas):
+                                if use_index:
+                                    temp_frame = self.process_face(self.options.selected_index, face, temp_frame)
+                                else:
+                                    temp_frame = self.process_face(i, face, temp_frame)
+                                num_faces_found += 1
+                            if not roop.globals.vr_mode and num_faces_found == num_targetfaces:
+                                break
+            elif self.options.swap_mode == "all_female" or self.options.swap_mode == "all_male":
+                gender = 'F' if self.options.swap_mode == "all_female" else 'M'
+                for face in faces:
+                    if face.sex == gender:
+                        num_faces_found += 1
+                        temp_frame = self.process_face(self.options.selected_index, face, temp_frame)
+            
+            # might be slower but way more clean to release everything here
+            for face in faces:
+                del face
+            faces.clear()
+
+
+
+        if roop.globals.vr_mode and num_faces_found % 2 > 0:
+            # stereo image, there has to be an even number of faces
+            num_faces_found = 0
+            return num_faces_found, frame
+        if num_faces_found == 0:
+            return num_faces_found, frame
+
+        #maskprocessor = next((x for x in self.processors if x.type == 'mask'), None)
+
+        if self.options.imagemask is not None and self.options.imagemask.shape == frame.shape:
+            temp_frame = self.simple_blend_with_mask(temp_frame, frame, self.options.imagemask)
+        return num_faces_found, temp_frame
+
+
+    def rotation_action(self, original_face:Face, frame:Frame):
+        (height, width) = frame.shape[:2]
+
+        bounding_box_width = original_face.bbox[2] - original_face.bbox[0]
+        bounding_box_height = original_face.bbox[3] - original_face.bbox[1]
+        horizontal_face = bounding_box_width > bounding_box_height
+
+        center_x = width // 2.0
+        start_x = original_face.bbox[0]
+        end_x = original_face.bbox[2]
+        bbox_center_x = start_x + (bounding_box_width // 2.0)
+
+        # need to leverage the array of landmarks as decribed here:
+        # https://github.com/deepinsight/insightface/tree/master/alignment/coordinate_reg
+        # basically, we should be able to check for the relative position of eyes and nose
+        # then use that to determine which way the face is actually facing when in a horizontal position
+        # and use that to determine the correct rotation_action
+
+        forehead_x = original_face.landmark_2d_106[72][0]
+        chin_x = original_face.landmark_2d_106[0][0]
+
+        if horizontal_face:
+            if chin_x < forehead_x:
+                # this is someone lying down with their face like this (:
+                return "rotate_anticlockwise"
+            elif forehead_x < chin_x:
+                # this is someone lying down with their face like this :)
+                return "rotate_clockwise"
+            if bbox_center_x >= center_x:
+                # this is someone lying down with their face in the right hand side of the frame
+                return "rotate_anticlockwise"
+            if bbox_center_x < center_x:
+                # this is someone lying down with their face in the left hand side of the frame
+                return "rotate_clockwise"
+
+        return None
+
+
+    def auto_rotate_frame(self, original_face, frame:Frame):
+        target_face = original_face
+        original_frame = frame
+
+        rotation_action = self.rotation_action(original_face, frame)
+
+        if rotation_action == "rotate_anticlockwise":
+            #face is horizontal, rotating frame anti-clockwise and getting face bounding box from rotated frame
+            frame = rotate_anticlockwise(frame)
+        elif rotation_action == "rotate_clockwise":
+            #face is horizontal, rotating frame clockwise and getting face bounding box from rotated frame
+            frame = rotate_clockwise(frame)
+
+        return target_face, frame, rotation_action
+    
+
+    def auto_unrotate_frame(self, frame:Frame, rotation_action):
+        if rotation_action == "rotate_anticlockwise":
+            return rotate_clockwise(frame)
+        elif rotation_action == "rotate_clockwise":
+            return rotate_anticlockwise(frame)
+        
+        return frame
+
+
+
+    def process_face(self,face_index, target_face:Face, frame:Frame):
+        from roop.face_util import align_crop
+
+        enhanced_frame = None
+        if(len(self.input_face_datas) > 0):
+            inputface = self.input_face_datas[face_index].faces[0]
+        else:
+            inputface = None
+
+        rotation_action = None
+        if roop.globals.autorotate_faces:
+            # check for sideways rotation of face
+            rotation_action = self.rotation_action(target_face, frame)
+            if rotation_action is not None:
+                (startX, startY, endX, endY) = target_face["bbox"].astype("int")
+                width = endX - startX
+                height = endY - startY
+                offs = int(max(width,height) * 0.25)
+                rotcutframe,startX, startY, endX, endY = self.cutout(frame, startX - offs, startY - offs, endX + offs, endY + offs)
+                if rotation_action == "rotate_anticlockwise":
+                    rotcutframe = rotate_anticlockwise(rotcutframe)
+                elif rotation_action == "rotate_clockwise":
+                    rotcutframe = rotate_clockwise(rotcutframe)
+                # rotate image and re-detect face to correct wonky landmarks
+                rotface = get_first_face(rotcutframe)
+                if rotface is None:
+                    rotation_action = None
+                else:
+                    saved_frame = frame.copy()
+                    frame = rotcutframe
+                    target_face = rotface
+
+
+
+        # if roop.globals.vr_mode:
+            # bbox = target_face.bbox
+            # [orig_width, orig_height, _] = frame.shape
+
+            # # Convert bounding box to ints
+            # x1, y1, x2, y2 = map(int, bbox)
+
+            # # Determine the center of the bounding box
+            # x_center = (x1 + x2) / 2
+            # y_center = (y1 + y2) / 2
+
+            # # Normalize coordinates to range [-1, 1]
+            # x_center_normalized = x_center / (orig_width / 2) - 1
+            # y_center_normalized = y_center / (orig_width / 2) - 1
+
+            # # Convert normalized coordinates to spherical (theta, phi)
+            # theta = x_center_normalized * 180  # Theta ranges from -180 to 180 degrees
+            # phi = -y_center_normalized * 90  # Phi ranges from -90 to 90 degrees
+
+            # img = vr.GetPerspective(frame, 90, theta, phi, 1280, 1280)  # Generate perspective image
+
+
+        """ Code ported/adapted from Facefusion which borrowed the idea from Rope:
+            Kind of subsampling the cutout and aligned face image and faceswapping slices of it up to
+            the desired output resolution. This works around the current resolution limitations without using enhancers.
+        """
+        model_output_size = self.options.swap_output_size
+        subsample_size = max(self.options.subsample_size, model_output_size)
+        subsample_total = subsample_size // model_output_size
+        aligned_img, M = align_crop(frame, target_face.kps, subsample_size)
+
+        fake_frame = aligned_img
+        target_face.matrix = M
+
+        for p in self.processors:
+            if p.type == 'swap':
+                swap_result_frames = []
+                subsample_frames = self.implode_pixel_boost(aligned_img, model_output_size, subsample_total)
+                for sliced_frame in subsample_frames:
+                    for _ in range(0,self.options.num_swap_steps):
+                        sliced_frame = self.prepare_crop_frame(sliced_frame)
+                        sliced_frame = p.Run(inputface, target_face, sliced_frame)
+                        sliced_frame = self.normalize_swap_frame(sliced_frame)
+                    swap_result_frames.append(sliced_frame)
+                fake_frame = self.explode_pixel_boost(swap_result_frames, model_output_size, subsample_total, subsample_size)
+                fake_frame = fake_frame.astype(np.uint8)
+                scale_factor = 0.0
+            elif p.type == 'mask':
+                fake_frame = self.process_mask(p, aligned_img, fake_frame)
+            else:
+                enhanced_frame, scale_factor = p.Run(self.input_face_datas[face_index], target_face, fake_frame)
+
+        upscale = 512
+        orig_width = fake_frame.shape[1]
+        if orig_width != upscale:
+            fake_frame = cv2.resize(fake_frame, (upscale, upscale), cv2.INTER_CUBIC)
+        mask_offsets = (0,0,0,0,1,20) if inputface is None else inputface.mask_offsets
+
+        
+        if enhanced_frame is None:
+            scale_factor = int(upscale / orig_width)
+            result = self.paste_upscale(fake_frame, fake_frame, target_face.matrix, frame, scale_factor, mask_offsets)
+        else:
+            result = self.paste_upscale(fake_frame, enhanced_frame, target_face.matrix, frame, scale_factor, mask_offsets)
+
+        # Restore mouth before unrotating
+        if self.options.restore_original_mouth:
+            mouth_cutout, mouth_bb = self.create_mouth_mask(target_face, frame)
+            result = self.apply_mouth_area(result, mouth_cutout, mouth_bb)
+
+        if rotation_action is not None:
+            fake_frame = self.auto_unrotate_frame(result, rotation_action)
+            result = self.paste_simple(fake_frame, saved_frame, startX, startY)
+        
+        return result
+
+        
+
+
+    def cutout(self, frame:Frame, start_x, start_y, end_x, end_y):
+        if start_x < 0:
+            start_x = 0
+        if start_y < 0:
+            start_y = 0
+        if end_x > frame.shape[1]:
+            end_x = frame.shape[1]
+        if end_y > frame.shape[0]:
+            end_y = frame.shape[0]
+        return frame[start_y:end_y, start_x:end_x], start_x, start_y, end_x, end_y
+
+    def paste_simple(self, src:Frame, dest:Frame, start_x, start_y):
+        end_x = start_x + src.shape[1]
+        end_y = start_y + src.shape[0]
+
+        start_x, end_x, start_y, end_y = clamp_cut_values(start_x, end_x, start_y, end_y, dest)
+        dest[start_y:end_y, start_x:end_x] = src
+        return dest
+        
+    def simple_blend_with_mask(self, image1, image2, mask):
+        # Blend the images
+        blended_image = image1.astype(np.float32) * (1.0 - mask) + image2.astype(np.float32) * mask
+        return blended_image.astype(np.uint8)
+
+
+    def paste_upscale(self, fake_face, upsk_face, M, target_img, scale_factor, mask_offsets):
+        M_scale = M * scale_factor
+        IM = cv2.invertAffineTransform(M_scale)
+
+        face_matte = np.full((target_img.shape[0],target_img.shape[1]), 255, dtype=np.uint8)
+        # Generate white square sized as a upsk_face
+        img_matte = np.zeros((upsk_face.shape[0],upsk_face.shape[1]), dtype=np.uint8)
+
+        w = img_matte.shape[1]
+        h = img_matte.shape[0]
+
+        top = int(mask_offsets[0] * h)
+        bottom = int(h - (mask_offsets[1] * h))
+        left = int(mask_offsets[2] * w)
+        right = int(w - (mask_offsets[3] * w))
+        img_matte[top:bottom,left:right] = 255
+
+        # Transform white square back to target_img
+        img_matte = cv2.warpAffine(img_matte, IM, (target_img.shape[1], target_img.shape[0]), flags=cv2.INTER_NEAREST, borderValue=0.0) 
+        ##Blacken the edges of face_matte by 1 pixels (so the mask in not expanded on the image edges)
+        img_matte[:1,:] = img_matte[-1:,:] = img_matte[:,:1] = img_matte[:,-1:] = 0
+
+        img_matte = self.blur_area(img_matte, mask_offsets[4], mask_offsets[5])
+        #Normalize images to float values and reshape
+        img_matte = img_matte.astype(np.float32)/255
+        face_matte = face_matte.astype(np.float32)/255
+        img_matte = np.minimum(face_matte, img_matte)
+        if self.options.show_face_area_overlay:
+            # Additional steps for green overlay
+            green_overlay = np.zeros_like(target_img)
+            green_color = [0, 255, 0]  # RGB for green
+            for i in range(3):  # Apply green color where img_matte is not zero
+                green_overlay[:, :, i] = np.where(img_matte > 0, green_color[i], 0)        ##Transform upcaled face back to target_img
+        img_matte = np.reshape(img_matte, [img_matte.shape[0],img_matte.shape[1],1]) 
+        paste_face = cv2.warpAffine(upsk_face, IM, (target_img.shape[1], target_img.shape[0]), borderMode=cv2.BORDER_REPLICATE)
+        if upsk_face is not fake_face:
+            fake_face = cv2.warpAffine(fake_face, IM, (target_img.shape[1], target_img.shape[0]), borderMode=cv2.BORDER_REPLICATE)
+            paste_face = cv2.addWeighted(paste_face, self.options.blend_ratio, fake_face, 1.0 - self.options.blend_ratio, 0)
+
+        # Re-assemble image
+        paste_face = img_matte * paste_face
+        paste_face = paste_face + (1-img_matte) * target_img.astype(np.float32)
+        if self.options.show_face_area_overlay:
+            # Overlay the green overlay on the final image
+            paste_face = cv2.addWeighted(paste_face.astype(np.uint8), 1 - 0.5, green_overlay, 0.5, 0)
+        return paste_face.astype(np.uint8)
+
+
+    def blur_area(self, img_matte, num_erosion_iterations, blur_amount):
+        # Detect the affine transformed white area
+        mask_h_inds, mask_w_inds = np.where(img_matte==255) 
+        # Calculate the size (and diagonal size) of transformed white area width and height boundaries
+        mask_h = np.max(mask_h_inds) - np.min(mask_h_inds) 
+        mask_w = np.max(mask_w_inds) - np.min(mask_w_inds)
+        mask_size = int(np.sqrt(mask_h*mask_w))
+        # Calculate the kernel size for eroding img_matte by kernel (insightface empirical guess for best size was max(mask_size//10,10))
+        # k = max(mask_size//12, 8)
+        k = max(mask_size//(blur_amount // 2) , blur_amount // 2)
+        kernel = np.ones((k,k),np.uint8)
+        img_matte = cv2.erode(img_matte,kernel,iterations = num_erosion_iterations)
+        #Calculate the kernel size for blurring img_matte by blur_size (insightface empirical guess for best size was max(mask_size//20, 5))
+        # k = max(mask_size//24, 4) 
+        k = max(mask_size//blur_amount, blur_amount//5) 
+        kernel_size = (k, k)
+        blur_size = tuple(2*i+1 for i in kernel_size)
+        return cv2.GaussianBlur(img_matte, blur_size, 0)
+
+
+    def prepare_crop_frame(self, swap_frame):
+        model_type = 'inswapper'
+        model_mean = [0.0, 0.0, 0.0]
+        model_standard_deviation = [1.0, 1.0, 1.0]
+
+        if model_type == 'ghost':
+            swap_frame = swap_frame[:, :, ::-1] / 127.5 - 1
+        else:
+            swap_frame = swap_frame[:, :, ::-1] / 255.0
+        swap_frame = (swap_frame - model_mean) / model_standard_deviation
+        swap_frame = swap_frame.transpose(2, 0, 1)
+        swap_frame = np.expand_dims(swap_frame, axis = 0).astype(np.float32)
+        return swap_frame
+
+
+    def normalize_swap_frame(self, swap_frame):
+        model_type = 'inswapper'
+        swap_frame = swap_frame.transpose(1, 2, 0)
+
+        if model_type == 'ghost':
+            swap_frame = (swap_frame * 127.5 + 127.5).round()
+        else:
+            swap_frame = (swap_frame * 255.0).round()
+        swap_frame = swap_frame[:, :, ::-1]
+        return swap_frame
+
+    def implode_pixel_boost(self, aligned_face_frame, model_size, pixel_boost_total : int):
+        subsample_frame = aligned_face_frame.reshape(model_size, pixel_boost_total, model_size, pixel_boost_total, 3)
+        subsample_frame = subsample_frame.transpose(1, 3, 0, 2, 4).reshape(pixel_boost_total ** 2, model_size, model_size, 3)
+        return subsample_frame
+
+
+    def explode_pixel_boost(self, subsample_frame, model_size, pixel_boost_total, pixel_boost_size):
+        final_frame = np.stack(subsample_frame, axis = 0).reshape(pixel_boost_total, pixel_boost_total, model_size, model_size, 3)
+        final_frame = final_frame.transpose(2, 0, 3, 1, 4).reshape(pixel_boost_size, pixel_boost_size, 3)
+        return final_frame
+
+    def process_mask(self, processor, frame:Frame, target:Frame):
+        img_mask = processor.Run(frame, self.options.masking_text)
+        img_mask = cv2.resize(img_mask, (target.shape[1], target.shape[0]))
+        img_mask = np.reshape(img_mask, [img_mask.shape[0],img_mask.shape[1],1])
+
+        if self.options.show_face_masking:
+            result = (1 - img_mask) * frame.astype(np.float32)
+            return np.uint8(result)
+
+
+        target = target.astype(np.float32)
+        result = (1-img_mask) * target
+        result += img_mask * frame.astype(np.float32)
+        return np.uint8(result)
+
+
+    # Code for mouth restoration adapted from https://github.com/iVideoGameBoss/iRoopDeepFaceCam
+    
+    def create_mouth_mask(self, face: Face, frame: Frame):
+        mouth_cutout = None
+        
+        landmarks = face.landmark_2d_106
+        if landmarks is not None:
+            # Get mouth landmarks (indices 52 to 71 typically represent the outer mouth)
+            mouth_points = landmarks[52:71].astype(np.int32)
+            
+            # Add padding to mouth area
+            min_x, min_y = np.min(mouth_points, axis=0)
+            max_x, max_y = np.max(mouth_points, axis=0)
+            min_x = max(0, min_x - (15*6))
+            min_y = max(0, min_y - 22)
+            max_x = min(frame.shape[1], max_x + (15*6))
+            max_y = min(frame.shape[0], max_y + (90*6))
+            
+            # Extract the mouth area from the frame using the calculated bounding box
+            mouth_cutout = frame[min_y:max_y, min_x:max_x].copy()
+
+        return mouth_cutout, (min_x, min_y, max_x, max_y)
+
+
+
+    def create_feathered_mask(self, shape, feather_amount=30):
+        mask = np.zeros(shape[:2], dtype=np.float32)
+        center = (shape[1] // 2, shape[0] // 2)
+        cv2.ellipse(mask, center, (shape[1] // 2 - feather_amount, shape[0] // 2 - feather_amount), 
+                    0, 0, 360, 1, -1)
+        mask = cv2.GaussianBlur(mask, (feather_amount*2+1, feather_amount*2+1), 0)
+        return mask / np.max(mask)
+
+    def apply_mouth_area(self, frame: np.ndarray, mouth_cutout: np.ndarray, mouth_box: tuple) -> np.ndarray:
+        min_x, min_y, max_x, max_y = mouth_box
+        box_width = max_x - min_x
+        box_height = max_y - min_y
+        
+
+        # Resize the mouth cutout to match the mouth box size
+        if mouth_cutout is None or box_width is None or box_height is None:
+            return frame
+        try:
+            resized_mouth_cutout = cv2.resize(mouth_cutout, (box_width, box_height))
+            
+            # Extract the region of interest (ROI) from the target frame
+            roi = frame[min_y:max_y, min_x:max_x]
+            
+            # Ensure the ROI and resized_mouth_cutout have the same shape
+            if roi.shape != resized_mouth_cutout.shape:
+                resized_mouth_cutout = cv2.resize(resized_mouth_cutout, (roi.shape[1], roi.shape[0]))
+            
+            # Apply color transfer from ROI to mouth cutout
+            color_corrected_mouth = self.apply_color_transfer(resized_mouth_cutout, roi)
+            
+            # Create a feathered mask with increased feather amount
+            feather_amount = min(30, box_width // 15, box_height // 15)
+            mask = self.create_feathered_mask(resized_mouth_cutout.shape, feather_amount)
+            
+            # Blend the color-corrected mouth cutout with the ROI using the feathered mask
+            mask = mask[:,:,np.newaxis]  # Add channel dimension to mask
+            blended = (color_corrected_mouth * mask + roi * (1 - mask)).astype(np.uint8)
+            
+            # Place the blended result back into the frame
+            frame[min_y:max_y, min_x:max_x] = blended
+        except Exception as e:
+            print(f'Error {e}')
+            pass
+
+        return frame
+
+    def apply_color_transfer(self, source, target):
+        """
+        Apply color transfer from target to source image
+        """
+        source = cv2.cvtColor(source, cv2.COLOR_BGR2LAB).astype("float32")
+        target = cv2.cvtColor(target, cv2.COLOR_BGR2LAB).astype("float32")
+
+        source_mean, source_std = cv2.meanStdDev(source)
+        target_mean, target_std = cv2.meanStdDev(target)
+
+        # Reshape mean and std to be broadcastable
+        source_mean = source_mean.reshape(1, 1, 3)
+        source_std = source_std.reshape(1, 1, 3)
+        target_mean = target_mean.reshape(1, 1, 3)
+        target_std = target_std.reshape(1, 1, 3)
+
+        # Perform the color transfer
+        source = (source - source_mean) * (target_std / source_std) + target_mean
+        return cv2.cvtColor(np.clip(source, 0, 255).astype("uint8"), cv2.COLOR_LAB2BGR)
+
+
+
+    def unload_models():
+        pass
+
+
+    def release_resources(self):
+        for p in self.processors:
+            p.Release()
+        self.processors.clear()
+        if self.videowriter is not None:
+            self.videowriter.close()
+        if self.streamwriter is not None:
+            self.streamwriter.Close()
+

roop-unleashed-main/roop/ProcessOptions.py

@@ -0,0 +1,18 @@
+class ProcessOptions:
+
+    def __init__(self, swap_model, processordefines:dict, face_distance,  blend_ratio, swap_mode, selected_index, masking_text, imagemask, num_steps, subsample_size, show_face_area, restore_original_mouth, show_mask=False):
+        self.swap_modelname = swap_model
+        self.swap_output_size = int(swap_model.split()[-1])
+        self.processors = processordefines
+        self.face_distance_threshold = face_distance
+        self.blend_ratio = blend_ratio
+        self.swap_mode = swap_mode
+        self.selected_index = selected_index
+        self.masking_text = masking_text
+        self.imagemask = imagemask
+        self.num_swap_steps = num_steps
+        self.show_face_area_overlay = show_face_area
+        self.show_face_masking = show_mask
+        self.subsample_size = subsample_size
+        self.restore_original_mouth = restore_original_mouth
+        self.max_num_reuse_frame = 15

roop-unleashed-main/roop/StreamWriter.py

@@ -0,0 +1,60 @@
+import threading
+import time
+import pyvirtualcam
+
+
+class StreamWriter():
+    FPS = 30
+    VCam = None
+    Active = False
+    THREAD_LOCK_STREAM = threading.Lock()
+    time_last_process = None
+    timespan_min = 0.0
+    
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        self.Close()
+
+    def __init__(self, size, fps):
+        self.time_last_process = time.perf_counter()
+        self.FPS = fps
+        self.timespan_min = 1.0 / fps
+        print('Detecting virtual cam devices')
+        self.VCam = pyvirtualcam.Camera(width=size[0], height=size[1], fps=fps, fmt=pyvirtualcam.PixelFormat.BGR, print_fps=False)
+        if self.VCam is None:
+             print("No virtual camera found!")
+             return
+        print(f'Using virtual camera: {self.VCam.device}')
+        print(f'Using {self.VCam.native_fmt}')
+        self.Active = True
+         
+
+    def LimitFrames(self):
+        while True:
+            current_time = time.perf_counter()
+            time_passed = current_time - self.time_last_process
+            if time_passed >= self.timespan_min:
+                break
+
+    # First version used a queue and threading. Surprisingly this
+    # totally simple, blocking version is 10 times faster!
+    def WriteToStream(self, frame):
+        if self.VCam is None:
+             return
+        with self.THREAD_LOCK_STREAM:
+            self.LimitFrames()
+            self.VCam.send(frame)
+            self.time_last_process = time.perf_counter()
+             
+
+    def Close(self):
+        self.Active = False
+        if self.VCam is None:
+            self.VCam.close()
+            self.VCam = None
+
+
+
+