bardofcodes
commited on
Delete analogy_projector.py
Browse files- analogy_projector.py +0 -227
analogy_projector.py
DELETED
@@ -1,227 +0,0 @@
|
|
1 |
-
"""
|
2 |
-
ADOBE CONFIDENTIAL
|
3 |
-
Copyright 2024 Adobe
|
4 |
-
All Rights Reserved.
|
5 |
-
NOTICE: All information contained herein is, and remains
|
6 |
-
the property of Adobe and its suppliers, if any. The intellectual
|
7 |
-
and technical concepts contained herein are proprietary to Adobe
|
8 |
-
and its suppliers and are protected by all applicable intellectual
|
9 |
-
property laws, including trade secret and copyright laws.
|
10 |
-
Dissemination of this information or reproduction of this material
|
11 |
-
is strictly forbidden unless prior written permission is obtained
|
12 |
-
from Adobe.
|
13 |
-
"""
|
14 |
-
|
15 |
-
import einops
|
16 |
-
import numpy as np
|
17 |
-
import torch as th
|
18 |
-
import torch.nn as nn
|
19 |
-
from diffusers import ModelMixin
|
20 |
-
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
21 |
-
# REf: https://github.com/tatp22/multidim-positional-encoding/tree/master
|
22 |
-
|
23 |
-
|
24 |
-
OUT_SIZE = 768
|
25 |
-
IN_SIZE = 2048
|
26 |
-
|
27 |
-
|
28 |
-
def get_emb(sin_inp):
|
29 |
-
"""
|
30 |
-
Gets a base embedding for one dimension with sin and cos intertwined
|
31 |
-
"""
|
32 |
-
emb = th.stack((sin_inp.sin(), sin_inp.cos()), dim=-1)
|
33 |
-
return th.flatten(emb, -2, -1)
|
34 |
-
|
35 |
-
|
36 |
-
class PositionalEncoding1D(nn.Module):
|
37 |
-
def __init__(self, channels):
|
38 |
-
"""
|
39 |
-
:param channels: The last dimension of the tensor you want to apply pos emb to.
|
40 |
-
"""
|
41 |
-
super(PositionalEncoding1D, self).__init__()
|
42 |
-
self.org_channels = channels
|
43 |
-
channels = int(np.ceil(channels / 2) * 2)
|
44 |
-
self.channels = channels
|
45 |
-
inv_freq = 1.0 / (10000 ** (th.arange(0, channels, 2).float() / channels))
|
46 |
-
self.register_buffer("inv_freq", inv_freq)
|
47 |
-
self.register_buffer("cached_penc", None, persistent=False)
|
48 |
-
|
49 |
-
def forward(self, tensor):
|
50 |
-
"""
|
51 |
-
:param tensor: A 3d tensor of size (batch_size, x, ch)
|
52 |
-
:return: Positional Encoding Matrix of size (batch_size, x, ch)
|
53 |
-
"""
|
54 |
-
if len(tensor.shape) != 3:
|
55 |
-
raise RuntimeError("The input tensor has to be 3d!")
|
56 |
-
|
57 |
-
if self.cached_penc is not None and self.cached_penc.shape == tensor.shape:
|
58 |
-
return self.cached_penc
|
59 |
-
|
60 |
-
self.cached_penc = None
|
61 |
-
batch_size, x, orig_ch = tensor.shape
|
62 |
-
pos_x = th.arange(x, device=tensor.device, dtype=self.inv_freq.dtype)
|
63 |
-
sin_inp_x = th.einsum("i,j->ij", pos_x, self.inv_freq)
|
64 |
-
emb_x = get_emb(sin_inp_x)
|
65 |
-
emb = th.zeros((x, self.channels), device=tensor.device, dtype=tensor.dtype)
|
66 |
-
emb[:, : self.channels] = emb_x
|
67 |
-
|
68 |
-
self.cached_penc = emb[None, :, :orig_ch].repeat(batch_size, 1, 1)
|
69 |
-
return self.cached_penc
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
class PositionalEncoding3D(nn.Module):
|
74 |
-
def __init__(self, channels):
|
75 |
-
"""
|
76 |
-
:param channels: The last dimension of the tensor you want to apply pos emb to.
|
77 |
-
"""
|
78 |
-
super(PositionalEncoding3D, self).__init__()
|
79 |
-
self.org_channels = channels
|
80 |
-
channels = int(np.ceil(channels / 6) * 2)
|
81 |
-
if channels % 2:
|
82 |
-
channels += 1
|
83 |
-
self.channels = channels
|
84 |
-
inv_freq = 1.0 / (10000 ** (th.arange(0, channels, 2).float() / channels))
|
85 |
-
self.register_buffer("inv_freq", inv_freq)
|
86 |
-
self.register_buffer("cached_penc", None, persistent=False)
|
87 |
-
|
88 |
-
def forward(self, tensor):
|
89 |
-
"""
|
90 |
-
:param tensor: A 5d tensor of size (batch_size, x, y, z, ch)
|
91 |
-
:return: Positional Encoding Matrix of size (batch_size, x, y, z, ch)
|
92 |
-
"""
|
93 |
-
if len(tensor.shape) != 5:
|
94 |
-
raise RuntimeError("The input tensor has to be 5d!")
|
95 |
-
|
96 |
-
if self.cached_penc is not None and self.cached_penc.shape == tensor.shape:
|
97 |
-
return self.cached_penc
|
98 |
-
|
99 |
-
self.cached_penc = None
|
100 |
-
batch_size, x, y, z, orig_ch = tensor.shape
|
101 |
-
pos_x = th.arange(x, device=tensor.device, dtype=self.inv_freq.dtype)
|
102 |
-
pos_y = th.arange(y, device=tensor.device, dtype=self.inv_freq.dtype)
|
103 |
-
pos_z = th.arange(z, device=tensor.device, dtype=self.inv_freq.dtype)
|
104 |
-
sin_inp_x = th.einsum("i,j->ij", pos_x, self.inv_freq)
|
105 |
-
sin_inp_y = th.einsum("i,j->ij", pos_y, self.inv_freq)
|
106 |
-
sin_inp_z = th.einsum("i,j->ij", pos_z, self.inv_freq)
|
107 |
-
emb_x = get_emb(sin_inp_x).unsqueeze(1).unsqueeze(1)
|
108 |
-
emb_y = get_emb(sin_inp_y).unsqueeze(1)
|
109 |
-
emb_z = get_emb(sin_inp_z)
|
110 |
-
emb = th.zeros(
|
111 |
-
(x, y, z, self.channels * 3),
|
112 |
-
device=tensor.device,
|
113 |
-
dtype=tensor.dtype,
|
114 |
-
)
|
115 |
-
emb[:, :, :, : self.channels] = emb_x
|
116 |
-
emb[:, :, :, self.channels : 2 * self.channels] = emb_y
|
117 |
-
emb[:, :, :, 2 * self.channels :] = emb_z
|
118 |
-
|
119 |
-
self.cached_penc = emb[None, :, :, :, :orig_ch].repeat(batch_size, 1, 1, 1, 1)
|
120 |
-
return self.cached_penc
|
121 |
-
|
122 |
-
class AnalogyProjector(ModelMixin, ConfigMixin):
|
123 |
-
|
124 |
-
@register_to_config
|
125 |
-
def __init__(self):
|
126 |
-
super(AnalogyProjector, self).__init__()
|
127 |
-
self.projector = DinoSiglipMixer()
|
128 |
-
self.pos_embd_1D = PositionalEncoding1D(OUT_SIZE)
|
129 |
-
self.pos_embd_3D = PositionalEncoding3D(OUT_SIZE)
|
130 |
-
|
131 |
-
|
132 |
-
def forward(self, dino_in, siglip_in, batch_size):
|
133 |
-
|
134 |
-
image_embeddings = self.projector(dino_in, siglip_in)
|
135 |
-
|
136 |
-
image_embeddings = einops.rearrange(image_embeddings, '(k b) t d -> b k t d', b=batch_size)
|
137 |
-
image_embeddings = self.position_embd(image_embeddings)
|
138 |
-
return image_embeddings
|
139 |
-
|
140 |
-
def position_embd(self, image_embeddings, concat=False):
|
141 |
-
canvas_embd = image_embeddings[:, :, 1:, :]
|
142 |
-
batch_size = canvas_embd.shape[0]
|
143 |
-
type_size = canvas_embd.shape[1]
|
144 |
-
xy_size = canvas_embd.shape[2]
|
145 |
-
|
146 |
-
x_size = int(xy_size ** 0.5)
|
147 |
-
|
148 |
-
canvas_embd = canvas_embd.reshape(batch_size, type_size, x_size, x_size, -1)
|
149 |
-
if concat:
|
150 |
-
canvas_embd = th.cat([canvas_embd, self.pos_embd_3D(canvas_embd)], -1)
|
151 |
-
else:
|
152 |
-
canvas_embd = self.pos_embd_3D(canvas_embd) + canvas_embd
|
153 |
-
canvas_embd = canvas_embd.reshape(batch_size, type_size, xy_size, -1)
|
154 |
-
|
155 |
-
class_embd = image_embeddings[:, :, 0, :]
|
156 |
-
if concat:
|
157 |
-
class_embd = th.cat([class_embd, self.pos_embd_1D(class_embd)], -1)
|
158 |
-
else:
|
159 |
-
class_embd = self.pos_embd_1D(class_embd) + class_embd
|
160 |
-
all_embd_list = []
|
161 |
-
for i in range(type_size):
|
162 |
-
all_embd_list.append(class_embd[:, i:i+1])
|
163 |
-
all_embd_list.append(canvas_embd[:, i])
|
164 |
-
image_embeddings = th.cat(all_embd_list, 1)
|
165 |
-
return image_embeddings
|
166 |
-
|
167 |
-
|
168 |
-
class HighLowMixer(th.nn.Module):
|
169 |
-
def __init__(self, in_size=IN_SIZE, out_size=OUT_SIZE):
|
170 |
-
super().__init__()
|
171 |
-
mid_size = (in_size + out_size) // 2
|
172 |
-
|
173 |
-
self.lower_projector = th.nn.Sequential(
|
174 |
-
th.nn.LayerNorm(IN_SIZE//2),
|
175 |
-
th.nn.SiLU()
|
176 |
-
)
|
177 |
-
self.upper_projector = th.nn.Sequential(
|
178 |
-
th.nn.LayerNorm(IN_SIZE//2),
|
179 |
-
th.nn.SiLU()
|
180 |
-
)
|
181 |
-
self.projectors = th.nn.ModuleList([
|
182 |
-
# add layer norm
|
183 |
-
th.nn.Linear(in_size, mid_size),
|
184 |
-
th.nn.SiLU(),
|
185 |
-
th.nn.Linear(mid_size, out_size)
|
186 |
-
])
|
187 |
-
# initialize
|
188 |
-
for proj in self.projectors:
|
189 |
-
if isinstance(proj, th.nn.Linear):
|
190 |
-
th.nn.init.xavier_uniform_(proj.weight)
|
191 |
-
th.nn.init.zeros_(proj.bias)
|
192 |
-
|
193 |
-
def forward(self, lower_in, upper_in, ):
|
194 |
-
# ALso format lower_in
|
195 |
-
lower_in = self.lower_projector(lower_in)
|
196 |
-
upper_in = self.upper_projector(upper_in)
|
197 |
-
x = th.cat([lower_in, upper_in], -1)
|
198 |
-
for proj in self.projectors:
|
199 |
-
x = proj(x)
|
200 |
-
return x
|
201 |
-
|
202 |
-
class DinoSiglipMixer(th.nn.Module):
|
203 |
-
def __init__(self, in_size=OUT_SIZE * 2, out_size=OUT_SIZE):
|
204 |
-
super().__init__()
|
205 |
-
self.dino_projector = HighLowMixer()
|
206 |
-
self.siglip_projector = HighLowMixer()
|
207 |
-
self.projectors = th.nn.Sequential(
|
208 |
-
th.nn.SiLU(),
|
209 |
-
th.nn.Linear(in_size, out_size),
|
210 |
-
)
|
211 |
-
# initialize
|
212 |
-
for proj in self.projectors:
|
213 |
-
if isinstance(proj, th.nn.Linear):
|
214 |
-
th.nn.init.xavier_uniform_(proj.weight)
|
215 |
-
th.nn.init.zeros_(proj.bias)
|
216 |
-
|
217 |
-
|
218 |
-
def forward(self, dino_in, siglip_in):
|
219 |
-
# ALso format lower_in
|
220 |
-
lower, upper = th.chunk(dino_in, 2, -1)
|
221 |
-
dino_out = self.dino_projector(lower, upper)
|
222 |
-
lower, upper = th.chunk(siglip_in, 2, -1)
|
223 |
-
siglip_out = self.siglip_projector(lower, upper)
|
224 |
-
x = th.cat([dino_out, siglip_out], -1)
|
225 |
-
for proj in self.projectors:
|
226 |
-
x = proj(x)
|
227 |
-
return x
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|