Update to Diffusers
Browse files- .gitignore +2 -1
- .gitmodules +0 -3
- .pre-commit-config.yaml +2 -0
- README.md +4 -3
- app.py +11 -228
- app_inference.py +150 -0
- app_training.py +128 -0
- app_upload.py +95 -0
- constants.py +6 -0
- inference.py +45 -48
- lora +0 -1
- requirements.txt +9 -7
- train_dreambooth_lora.py +1018 -0
- trainer.py +67 -54
- uploader.py +35 -16
- utils.py +18 -0
.gitignore
CHANGED
@@ -1,5 +1,6 @@
|
|
1 |
training_data/
|
2 |
-
|
|
|
3 |
|
4 |
|
5 |
# Byte-compiled / optimized / DLL files
|
|
|
1 |
training_data/
|
2 |
+
experiments/
|
3 |
+
wandb/
|
4 |
|
5 |
|
6 |
# Byte-compiled / optimized / DLL files
|
.gitmodules
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
[submodule "lora"]
|
2 |
-
path = lora
|
3 |
-
url = https://github.com/cloneofsimo/lora
|
|
|
|
|
|
|
|
.pre-commit-config.yaml
CHANGED
@@ -1,3 +1,4 @@
|
|
|
|
1 |
repos:
|
2 |
- repo: https://github.com/pre-commit/pre-commit-hooks
|
3 |
rev: v4.2.0
|
@@ -28,6 +29,7 @@ repos:
|
|
28 |
hooks:
|
29 |
- id: mypy
|
30 |
args: ['--ignore-missing-imports']
|
|
|
31 |
- repo: https://github.com/google/yapf
|
32 |
rev: v0.32.0
|
33 |
hooks:
|
|
|
1 |
+
exclude: train_dreambooth_lora.py
|
2 |
repos:
|
3 |
- repo: https://github.com/pre-commit/pre-commit-hooks
|
4 |
rev: v4.2.0
|
|
|
29 |
hooks:
|
30 |
- id: mypy
|
31 |
args: ['--ignore-missing-imports']
|
32 |
+
additional_dependencies: ['types-python-slugify']
|
33 |
- repo: https://github.com/google/yapf
|
34 |
rev: v0.32.0
|
35 |
hooks:
|
README.md
CHANGED
@@ -1,10 +1,11 @@
|
|
1 |
---
|
2 |
-
title: LoRA
|
3 |
-
emoji:
|
4 |
colorFrom: red
|
5 |
colorTo: purple
|
6 |
sdk: gradio
|
7 |
-
sdk_version: 3.
|
|
|
8 |
app_file: app.py
|
9 |
pinned: false
|
10 |
license: mit
|
|
|
1 |
---
|
2 |
+
title: LoRA DreamBooth Training UI
|
3 |
+
emoji: ⚡
|
4 |
colorFrom: red
|
5 |
colorTo: purple
|
6 |
sdk: gradio
|
7 |
+
sdk_version: 3.16.2
|
8 |
+
python_version: 3.10.9
|
9 |
app_file: app.py
|
10 |
pinned: false
|
11 |
license: mit
|
app.py
CHANGED
@@ -1,27 +1,21 @@
|
|
1 |
#!/usr/bin/env python
|
2 |
-
"""Unofficial demo app for https://github.com/cloneofsimo/lora.
|
3 |
-
|
4 |
-
The code in this repo is partly adapted from the following repository:
|
5 |
-
https://huggingface.co/spaces/multimodalart/dreambooth-training/tree/a00184917aa273c6d8adab08d5deb9b39b997938
|
6 |
-
The license of the original code is MIT, which is specified in the README.md.
|
7 |
-
"""
|
8 |
|
9 |
from __future__ import annotations
|
10 |
|
11 |
import os
|
12 |
-
import pathlib
|
13 |
|
14 |
import gradio as gr
|
15 |
import torch
|
16 |
|
|
|
|
|
|
|
17 |
from inference import InferencePipeline
|
18 |
from trainer import Trainer
|
19 |
-
from uploader import upload
|
20 |
|
21 |
-
TITLE = '# LoRA
|
22 |
-
DESCRIPTION = 'This is an unofficial demo for [https://github.com/cloneofsimo/lora](https://github.com/cloneofsimo/lora).'
|
23 |
|
24 |
-
ORIGINAL_SPACE_ID = 'hysts/LoRA-
|
25 |
SPACE_ID = os.getenv('SPACE_ID', ORIGINAL_SPACE_ID)
|
26 |
SHARED_UI_WARNING = f'''# Attention - This Space doesn't work in this shared UI. You can duplicate and use it with a paid private T4 GPU.
|
27 |
|
@@ -29,7 +23,6 @@ SHARED_UI_WARNING = f'''# Attention - This Space doesn't work in this shared UI.
|
|
29 |
'''
|
30 |
if os.getenv('SYSTEM') == 'spaces' and SPACE_ID != ORIGINAL_SPACE_ID:
|
31 |
SETTINGS = f'<a href="https://huggingface.co/spaces/{SPACE_ID}/settings">Settings</a>'
|
32 |
-
|
33 |
else:
|
34 |
SETTINGS = 'Settings'
|
35 |
CUDA_NOT_AVAILABLE_WARNING = f'''# Attention - Running on CPU.
|
@@ -39,6 +32,8 @@ You can assign a GPU in the {SETTINGS} tab if you are running this on HF Spaces.
|
|
39 |
</center>
|
40 |
'''
|
41 |
|
|
|
|
|
42 |
|
43 |
def show_warning(warning_text: str) -> gr.Blocks:
|
44 |
with gr.Blocks() as demo:
|
@@ -47,217 +42,7 @@ def show_warning(warning_text: str) -> gr.Blocks:
|
|
47 |
return demo
|
48 |
|
49 |
|
50 |
-
|
51 |
-
paths = sorted(pathlib.Path('results').glob('*.pt'))
|
52 |
-
paths = [path.as_posix() for path in paths] # type: ignore
|
53 |
-
return gr.update(value=paths or None)
|
54 |
-
|
55 |
-
|
56 |
-
def create_training_demo(trainer: Trainer,
|
57 |
-
pipe: InferencePipeline) -> gr.Blocks:
|
58 |
-
with gr.Blocks() as demo:
|
59 |
-
base_model = gr.Dropdown(
|
60 |
-
choices=['stabilityai/stable-diffusion-2-1-base'],
|
61 |
-
value='stabilityai/stable-diffusion-2-1-base',
|
62 |
-
label='Base Model',
|
63 |
-
visible=False)
|
64 |
-
resolution = gr.Dropdown(choices=['512'],
|
65 |
-
value='512',
|
66 |
-
label='Resolution',
|
67 |
-
visible=False)
|
68 |
-
|
69 |
-
with gr.Row():
|
70 |
-
with gr.Box():
|
71 |
-
gr.Markdown('Training Data')
|
72 |
-
concept_images = gr.Files(label='Images for your concept')
|
73 |
-
concept_prompt = gr.Textbox(label='Concept Prompt',
|
74 |
-
max_lines=1)
|
75 |
-
gr.Markdown('''
|
76 |
-
- Upload images of the style you are planning on training on.
|
77 |
-
- For a concept prompt, use a unique, made up word to avoid collisions.
|
78 |
-
''')
|
79 |
-
with gr.Box():
|
80 |
-
gr.Markdown('Training Parameters')
|
81 |
-
num_training_steps = gr.Number(
|
82 |
-
label='Number of Training Steps', value=1000, precision=0)
|
83 |
-
learning_rate = gr.Number(label='Learning Rate', value=0.0001)
|
84 |
-
train_text_encoder = gr.Checkbox(label='Train Text Encoder',
|
85 |
-
value=True)
|
86 |
-
learning_rate_text = gr.Number(
|
87 |
-
label='Learning Rate for Text Encoder', value=0.00005)
|
88 |
-
gradient_accumulation = gr.Number(
|
89 |
-
label='Number of Gradient Accumulation',
|
90 |
-
value=1,
|
91 |
-
precision=0)
|
92 |
-
fp16 = gr.Checkbox(label='FP16', value=True)
|
93 |
-
use_8bit_adam = gr.Checkbox(label='Use 8bit Adam', value=True)
|
94 |
-
gr.Markdown('''
|
95 |
-
- It will take about 8 minutes to train for 1000 steps with a T4 GPU.
|
96 |
-
- You may want to try a small number of steps first, like 1, to see if everything works fine in your environment.
|
97 |
-
- Note that your trained models will be deleted when the second training is started. You can upload your trained model in the "Upload" tab.
|
98 |
-
''')
|
99 |
-
|
100 |
-
run_button = gr.Button('Start Training')
|
101 |
-
with gr.Box():
|
102 |
-
with gr.Row():
|
103 |
-
check_status_button = gr.Button('Check Training Status')
|
104 |
-
with gr.Column():
|
105 |
-
with gr.Box():
|
106 |
-
gr.Markdown('Message')
|
107 |
-
training_status = gr.Markdown()
|
108 |
-
output_files = gr.Files(label='Trained Weight Files')
|
109 |
-
|
110 |
-
run_button.click(fn=pipe.clear)
|
111 |
-
run_button.click(fn=trainer.run,
|
112 |
-
inputs=[
|
113 |
-
base_model,
|
114 |
-
resolution,
|
115 |
-
concept_images,
|
116 |
-
concept_prompt,
|
117 |
-
num_training_steps,
|
118 |
-
learning_rate,
|
119 |
-
train_text_encoder,
|
120 |
-
learning_rate_text,
|
121 |
-
gradient_accumulation,
|
122 |
-
fp16,
|
123 |
-
use_8bit_adam,
|
124 |
-
],
|
125 |
-
outputs=[
|
126 |
-
training_status,
|
127 |
-
output_files,
|
128 |
-
],
|
129 |
-
queue=False)
|
130 |
-
check_status_button.click(fn=trainer.check_if_running,
|
131 |
-
inputs=None,
|
132 |
-
outputs=training_status,
|
133 |
-
queue=False)
|
134 |
-
check_status_button.click(fn=update_output_files,
|
135 |
-
inputs=None,
|
136 |
-
outputs=output_files,
|
137 |
-
queue=False)
|
138 |
-
return demo
|
139 |
-
|
140 |
-
|
141 |
-
def find_weight_files() -> list[str]:
|
142 |
-
curr_dir = pathlib.Path(__file__).parent
|
143 |
-
paths = sorted(curr_dir.rglob('*.pt'))
|
144 |
-
paths = [path for path in paths if not path.stem.endswith('.text_encoder')]
|
145 |
-
return [path.relative_to(curr_dir).as_posix() for path in paths]
|
146 |
-
|
147 |
-
|
148 |
-
def reload_lora_weight_list() -> dict:
|
149 |
-
return gr.update(choices=find_weight_files())
|
150 |
-
|
151 |
-
|
152 |
-
def create_inference_demo(pipe: InferencePipeline) -> gr.Blocks:
|
153 |
-
with gr.Blocks() as demo:
|
154 |
-
with gr.Row():
|
155 |
-
with gr.Column():
|
156 |
-
base_model = gr.Dropdown(
|
157 |
-
choices=['stabilityai/stable-diffusion-2-1-base'],
|
158 |
-
value='stabilityai/stable-diffusion-2-1-base',
|
159 |
-
label='Base Model',
|
160 |
-
visible=False)
|
161 |
-
reload_button = gr.Button('Reload Weight List')
|
162 |
-
lora_weight_name = gr.Dropdown(choices=find_weight_files(),
|
163 |
-
value='lora/lora_disney.pt',
|
164 |
-
label='LoRA Weight File')
|
165 |
-
prompt = gr.Textbox(
|
166 |
-
label='Prompt',
|
167 |
-
max_lines=1,
|
168 |
-
placeholder='Example: "style of sks, baby lion"')
|
169 |
-
alpha = gr.Slider(label='Alpha',
|
170 |
-
minimum=0,
|
171 |
-
maximum=2,
|
172 |
-
step=0.05,
|
173 |
-
value=1)
|
174 |
-
alpha_for_text = gr.Slider(label='Alpha for Text Encoder',
|
175 |
-
minimum=0,
|
176 |
-
maximum=2,
|
177 |
-
step=0.05,
|
178 |
-
value=1)
|
179 |
-
seed = gr.Slider(label='Seed',
|
180 |
-
minimum=0,
|
181 |
-
maximum=100000,
|
182 |
-
step=1,
|
183 |
-
value=1)
|
184 |
-
with gr.Accordion('Other Parameters', open=False):
|
185 |
-
num_steps = gr.Slider(label='Number of Steps',
|
186 |
-
minimum=0,
|
187 |
-
maximum=100,
|
188 |
-
step=1,
|
189 |
-
value=50)
|
190 |
-
guidance_scale = gr.Slider(label='CFG Scale',
|
191 |
-
minimum=0,
|
192 |
-
maximum=50,
|
193 |
-
step=0.1,
|
194 |
-
value=7)
|
195 |
-
|
196 |
-
run_button = gr.Button('Generate')
|
197 |
-
|
198 |
-
gr.Markdown('''
|
199 |
-
- Models with names starting with "lora/" are the pretrained models provided in the [original repo](https://github.com/cloneofsimo/lora), and the ones with names starting with "results/" are your trained models.
|
200 |
-
- After training, you can press "Reload Weight List" button to load your trained model names.
|
201 |
-
- The pretrained models for "disney", "illust" and "pop" are trained with the concept prompt "style of sks".
|
202 |
-
- The pretrained model for "kiriko" is trained with the concept prompt "game character bnha". For this model, the text encoder is also trained.
|
203 |
-
''')
|
204 |
-
with gr.Column():
|
205 |
-
result = gr.Image(label='Result')
|
206 |
-
|
207 |
-
reload_button.click(fn=reload_lora_weight_list,
|
208 |
-
inputs=None,
|
209 |
-
outputs=lora_weight_name)
|
210 |
-
prompt.submit(fn=pipe.run,
|
211 |
-
inputs=[
|
212 |
-
base_model,
|
213 |
-
lora_weight_name,
|
214 |
-
prompt,
|
215 |
-
alpha,
|
216 |
-
alpha_for_text,
|
217 |
-
seed,
|
218 |
-
num_steps,
|
219 |
-
guidance_scale,
|
220 |
-
],
|
221 |
-
outputs=result,
|
222 |
-
queue=False)
|
223 |
-
run_button.click(fn=pipe.run,
|
224 |
-
inputs=[
|
225 |
-
base_model,
|
226 |
-
lora_weight_name,
|
227 |
-
prompt,
|
228 |
-
alpha,
|
229 |
-
alpha_for_text,
|
230 |
-
seed,
|
231 |
-
num_steps,
|
232 |
-
guidance_scale,
|
233 |
-
],
|
234 |
-
outputs=result,
|
235 |
-
queue=False)
|
236 |
-
return demo
|
237 |
-
|
238 |
-
|
239 |
-
def create_upload_demo() -> gr.Blocks:
|
240 |
-
with gr.Blocks() as demo:
|
241 |
-
model_name = gr.Textbox(label='Model Name')
|
242 |
-
hf_token = gr.Textbox(
|
243 |
-
label='Hugging Face Token (with write permission)')
|
244 |
-
upload_button = gr.Button('Upload')
|
245 |
-
with gr.Box():
|
246 |
-
gr.Markdown('Message')
|
247 |
-
result = gr.Markdown()
|
248 |
-
gr.Markdown('''
|
249 |
-
- You can upload your trained model to your private Model repo (i.e. https://huggingface.co/{your_username}/{model_name}).
|
250 |
-
- You can find your Hugging Face token [here](https://huggingface.co/settings/tokens).
|
251 |
-
''')
|
252 |
-
|
253 |
-
upload_button.click(fn=upload,
|
254 |
-
inputs=[model_name, hf_token],
|
255 |
-
outputs=result)
|
256 |
-
|
257 |
-
return demo
|
258 |
-
|
259 |
-
|
260 |
-
pipe = InferencePipeline()
|
261 |
trainer = Trainer()
|
262 |
|
263 |
with gr.Blocks(css='style.css') as demo:
|
@@ -267,14 +52,12 @@ with gr.Blocks(css='style.css') as demo:
|
|
267 |
show_warning(CUDA_NOT_AVAILABLE_WARNING)
|
268 |
|
269 |
gr.Markdown(TITLE)
|
270 |
-
gr.Markdown(DESCRIPTION)
|
271 |
-
|
272 |
with gr.Tabs():
|
273 |
with gr.TabItem('Train'):
|
274 |
create_training_demo(trainer, pipe)
|
275 |
with gr.TabItem('Test'):
|
276 |
-
create_inference_demo(pipe)
|
277 |
with gr.TabItem('Upload'):
|
278 |
-
create_upload_demo()
|
279 |
|
280 |
-
demo.queue(
|
|
|
1 |
#!/usr/bin/env python
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
from __future__ import annotations
|
4 |
|
5 |
import os
|
|
|
6 |
|
7 |
import gradio as gr
|
8 |
import torch
|
9 |
|
10 |
+
from app_inference import create_inference_demo
|
11 |
+
from app_training import create_training_demo
|
12 |
+
from app_upload import create_upload_demo
|
13 |
from inference import InferencePipeline
|
14 |
from trainer import Trainer
|
|
|
15 |
|
16 |
+
TITLE = '# LoRA DreamBooth Training UI'
|
|
|
17 |
|
18 |
+
ORIGINAL_SPACE_ID = 'hysts/LoRA-DreamBooth-Training-UI'
|
19 |
SPACE_ID = os.getenv('SPACE_ID', ORIGINAL_SPACE_ID)
|
20 |
SHARED_UI_WARNING = f'''# Attention - This Space doesn't work in this shared UI. You can duplicate and use it with a paid private T4 GPU.
|
21 |
|
|
|
23 |
'''
|
24 |
if os.getenv('SYSTEM') == 'spaces' and SPACE_ID != ORIGINAL_SPACE_ID:
|
25 |
SETTINGS = f'<a href="https://huggingface.co/spaces/{SPACE_ID}/settings">Settings</a>'
|
|
|
26 |
else:
|
27 |
SETTINGS = 'Settings'
|
28 |
CUDA_NOT_AVAILABLE_WARNING = f'''# Attention - Running on CPU.
|
|
|
32 |
</center>
|
33 |
'''
|
34 |
|
35 |
+
HF_TOKEN = os.getenv('HF_TOKEN')
|
36 |
+
|
37 |
|
38 |
def show_warning(warning_text: str) -> gr.Blocks:
|
39 |
with gr.Blocks() as demo:
|
|
|
42 |
return demo
|
43 |
|
44 |
|
45 |
+
pipe = InferencePipeline(HF_TOKEN)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
trainer = Trainer()
|
47 |
|
48 |
with gr.Blocks(css='style.css') as demo:
|
|
|
52 |
show_warning(CUDA_NOT_AVAILABLE_WARNING)
|
53 |
|
54 |
gr.Markdown(TITLE)
|
|
|
|
|
55 |
with gr.Tabs():
|
56 |
with gr.TabItem('Train'):
|
57 |
create_training_demo(trainer, pipe)
|
58 |
with gr.TabItem('Test'):
|
59 |
+
create_inference_demo(pipe, HF_TOKEN)
|
60 |
with gr.TabItem('Upload'):
|
61 |
+
create_upload_demo(HF_TOKEN)
|
62 |
|
63 |
+
demo.queue(max_size=1).launch(share=False)
|
app_inference.py
ADDED
@@ -0,0 +1,150 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
|
3 |
+
from __future__ import annotations
|
4 |
+
|
5 |
+
import enum
|
6 |
+
|
7 |
+
import gradio as gr
|
8 |
+
from huggingface_hub import HfApi
|
9 |
+
|
10 |
+
from inference import InferencePipeline
|
11 |
+
from utils import find_exp_dirs
|
12 |
+
|
13 |
+
SAMPLE_MODEL_IDS = ['patrickvonplaten/lora_dreambooth_dog_example']
|
14 |
+
|
15 |
+
|
16 |
+
class ModelSource(enum.Enum):
|
17 |
+
SAMPLE = 'Sample'
|
18 |
+
HUB_LIB = 'Hub (lora-library)'
|
19 |
+
LOCAL = 'Local'
|
20 |
+
|
21 |
+
|
22 |
+
class InferenceUtil:
|
23 |
+
def __init__(self, hf_token: str | None):
|
24 |
+
self.hf_token = hf_token
|
25 |
+
|
26 |
+
@staticmethod
|
27 |
+
def load_sample_lora_model_list():
|
28 |
+
return gr.update(choices=SAMPLE_MODEL_IDS, value=SAMPLE_MODEL_IDS[0])
|
29 |
+
|
30 |
+
def load_hub_lora_model_list(self) -> dict:
|
31 |
+
api = HfApi(token=self.hf_token)
|
32 |
+
choices = [
|
33 |
+
info.modelId for info in api.list_models(author='lora-library')
|
34 |
+
]
|
35 |
+
return gr.update(choices=choices,
|
36 |
+
value=choices[0] if choices else None)
|
37 |
+
|
38 |
+
@staticmethod
|
39 |
+
def load_local_lora_model_list() -> dict:
|
40 |
+
choices = find_exp_dirs()
|
41 |
+
return gr.update(choices=choices,
|
42 |
+
value=choices[0] if choices else None)
|
43 |
+
|
44 |
+
def reload_lora_model_list(self, model_source: str) -> dict:
|
45 |
+
if model_source == ModelSource.SAMPLE.value:
|
46 |
+
return self.load_sample_lora_model_list()
|
47 |
+
elif model_source == ModelSource.HUB_LIB.value:
|
48 |
+
return self.load_hub_lora_model_list()
|
49 |
+
elif model_source == ModelSource.LOCAL.value:
|
50 |
+
return self.load_local_lora_model_list()
|
51 |
+
else:
|
52 |
+
raise ValueError
|
53 |
+
|
54 |
+
def load_model_info(self, lora_model_id: str) -> tuple[str, str]:
|
55 |
+
try:
|
56 |
+
card = InferencePipeline.get_model_card(lora_model_id,
|
57 |
+
self.hf_token)
|
58 |
+
except Exception:
|
59 |
+
return '', ''
|
60 |
+
base_model = getattr(card.data, 'base_model', '')
|
61 |
+
instance_prompt = getattr(card.data, 'instance_prompt', '')
|
62 |
+
return base_model, instance_prompt
|
63 |
+
|
64 |
+
|
65 |
+
def create_inference_demo(pipe: InferencePipeline,
|
66 |
+
hf_token: str | None = None) -> gr.Blocks:
|
67 |
+
app = InferenceUtil(hf_token)
|
68 |
+
|
69 |
+
with gr.Blocks() as demo:
|
70 |
+
with gr.Row():
|
71 |
+
with gr.Column():
|
72 |
+
with gr.Box():
|
73 |
+
model_source = gr.Radio(
|
74 |
+
label='Model Source',
|
75 |
+
choices=[_.value for _ in ModelSource],
|
76 |
+
value=ModelSource.SAMPLE.value)
|
77 |
+
reload_button = gr.Button('Reload Model List')
|
78 |
+
lora_model_id = gr.Dropdown(label='LoRA Model ID',
|
79 |
+
choices=SAMPLE_MODEL_IDS,
|
80 |
+
value=SAMPLE_MODEL_IDS[0])
|
81 |
+
with gr.Accordion(
|
82 |
+
label=
|
83 |
+
'Model info (Base model and instance prompt used for training)',
|
84 |
+
open=False):
|
85 |
+
with gr.Row():
|
86 |
+
base_model_used_for_training = gr.Text(
|
87 |
+
label='Base model', interactive=False)
|
88 |
+
instance_prompt_used_for_training = gr.Text(
|
89 |
+
label='Instance prompt', interactive=False)
|
90 |
+
prompt = gr.Textbox(
|
91 |
+
label='Prompt',
|
92 |
+
max_lines=1,
|
93 |
+
placeholder='Example: "A picture of a sks dog in a bucket"'
|
94 |
+
)
|
95 |
+
seed = gr.Slider(label='Seed',
|
96 |
+
minimum=0,
|
97 |
+
maximum=100000,
|
98 |
+
step=1,
|
99 |
+
value=0)
|
100 |
+
with gr.Accordion('Other Parameters', open=False):
|
101 |
+
num_steps = gr.Slider(label='Number of Steps',
|
102 |
+
minimum=0,
|
103 |
+
maximum=100,
|
104 |
+
step=1,
|
105 |
+
value=25)
|
106 |
+
guidance_scale = gr.Slider(label='CFG Scale',
|
107 |
+
minimum=0,
|
108 |
+
maximum=50,
|
109 |
+
step=0.1,
|
110 |
+
value=7.5)
|
111 |
+
|
112 |
+
run_button = gr.Button('Generate')
|
113 |
+
|
114 |
+
gr.Markdown('''
|
115 |
+
- After training, you can press "Reload Model List" button to load your trained model names.
|
116 |
+
''')
|
117 |
+
with gr.Column():
|
118 |
+
result = gr.Image(label='Result')
|
119 |
+
|
120 |
+
model_source.change(fn=app.reload_lora_model_list,
|
121 |
+
inputs=model_source,
|
122 |
+
outputs=lora_model_id)
|
123 |
+
reload_button.click(fn=app.reload_lora_model_list,
|
124 |
+
inputs=model_source,
|
125 |
+
outputs=lora_model_id)
|
126 |
+
lora_model_id.change(fn=app.load_model_info,
|
127 |
+
inputs=lora_model_id,
|
128 |
+
outputs=[
|
129 |
+
base_model_used_for_training,
|
130 |
+
instance_prompt_used_for_training,
|
131 |
+
])
|
132 |
+
inputs = [
|
133 |
+
lora_model_id,
|
134 |
+
prompt,
|
135 |
+
seed,
|
136 |
+
num_steps,
|
137 |
+
guidance_scale,
|
138 |
+
]
|
139 |
+
prompt.submit(fn=pipe.run, inputs=inputs, outputs=result)
|
140 |
+
run_button.click(fn=pipe.run, inputs=inputs, outputs=result)
|
141 |
+
return demo
|
142 |
+
|
143 |
+
|
144 |
+
if __name__ == '__main__':
|
145 |
+
import os
|
146 |
+
|
147 |
+
hf_token = os.getenv('HF_TOKEN')
|
148 |
+
pipe = InferencePipeline(hf_token)
|
149 |
+
demo = create_inference_demo(pipe, hf_token)
|
150 |
+
demo.queue(max_size=10).launch(share=False)
|
app_training.py
ADDED
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
|
3 |
+
from __future__ import annotations
|
4 |
+
|
5 |
+
import gradio as gr
|
6 |
+
|
7 |
+
from constants import UploadTarget
|
8 |
+
from inference import InferencePipeline
|
9 |
+
from trainer import Trainer
|
10 |
+
|
11 |
+
|
12 |
+
def create_training_demo(trainer: Trainer,
|
13 |
+
pipe: InferencePipeline | None = None) -> gr.Blocks:
|
14 |
+
with gr.Blocks() as demo:
|
15 |
+
with gr.Row():
|
16 |
+
with gr.Column():
|
17 |
+
with gr.Box():
|
18 |
+
gr.Markdown('Training Data')
|
19 |
+
instance_images = gr.Files(label='Instance images')
|
20 |
+
instance_prompt = gr.Textbox(label='Instance prompt',
|
21 |
+
max_lines=1)
|
22 |
+
gr.Markdown('''
|
23 |
+
- Upload images of the style you are planning on training on.
|
24 |
+
- For an instance prompt, use a unique, made up word to avoid collisions.
|
25 |
+
''')
|
26 |
+
with gr.Box():
|
27 |
+
gr.Markdown('Output Model')
|
28 |
+
output_model_name = gr.Text(label='Name of your model',
|
29 |
+
max_lines=1)
|
30 |
+
delete_existing_model = gr.Checkbox(
|
31 |
+
label='Delete existing model of the same name',
|
32 |
+
value=False)
|
33 |
+
validation_prompt = gr.Text(label='Validation Prompt')
|
34 |
+
with gr.Box():
|
35 |
+
gr.Markdown('Upload Settings')
|
36 |
+
with gr.Row():
|
37 |
+
upload_to_hub = gr.Checkbox(
|
38 |
+
label='Upload model to Hub', value=False)
|
39 |
+
use_private_repo = gr.Checkbox(label='Private',
|
40 |
+
value=False)
|
41 |
+
delete_existing_repo = gr.Checkbox(
|
42 |
+
label='Delete existing repo of the same name',
|
43 |
+
value=False)
|
44 |
+
upload_to = gr.Radio(
|
45 |
+
label='Upload to',
|
46 |
+
choices=[_.value for _ in UploadTarget],
|
47 |
+
value=UploadTarget.PERSONAL_PROFILE.value)
|
48 |
+
|
49 |
+
with gr.Box():
|
50 |
+
gr.Markdown('Training Parameters')
|
51 |
+
with gr.Row():
|
52 |
+
base_model = gr.Text(
|
53 |
+
label='Base Model',
|
54 |
+
value='stabilityai/stable-diffusion-2-1-base',
|
55 |
+
max_lines=1)
|
56 |
+
resolution = gr.Dropdown(choices=['512', '768'],
|
57 |
+
value='512',
|
58 |
+
label='Resolution')
|
59 |
+
num_training_steps = gr.Number(
|
60 |
+
label='Number of Training Steps', value=1000, precision=0)
|
61 |
+
learning_rate = gr.Number(label='Learning Rate', value=0.0001)
|
62 |
+
gradient_accumulation = gr.Number(
|
63 |
+
label='Number of Gradient Accumulation',
|
64 |
+
value=1,
|
65 |
+
precision=0)
|
66 |
+
seed = gr.Slider(label='Seed',
|
67 |
+
minimum=0,
|
68 |
+
maximum=100000,
|
69 |
+
step=1,
|
70 |
+
value=0)
|
71 |
+
fp16 = gr.Checkbox(label='FP16', value=True)
|
72 |
+
use_8bit_adam = gr.Checkbox(label='Use 8bit Adam', value=True)
|
73 |
+
checkpointing_steps = gr.Number(label='Checkpointing Steps',
|
74 |
+
value=100,
|
75 |
+
precision=0)
|
76 |
+
use_wandb = gr.Checkbox(label='Use W&B', value=False)
|
77 |
+
validation_epochs = gr.Number(label='Validation Epochs',
|
78 |
+
value=100,
|
79 |
+
precision=0)
|
80 |
+
gr.Markdown('''
|
81 |
+
- It will take about 8 minutes to train for 1000 steps with a T4 GPU.
|
82 |
+
- You may want to try a small number of steps first, like 1, to see if everything works fine in your environment.
|
83 |
+
- You need to set the environment variable `WANDB_API_KEY` if you'd like to use W&B. See [W&B documentation](https://docs.wandb.ai/guides/track/advanced/environment-variables).
|
84 |
+
- **Note:** Due to [this issue](https://github.com/huggingface/accelerate/issues/944), currently, training will not terminate properly if you use W&B.
|
85 |
+
''')
|
86 |
+
|
87 |
+
# TODO currently disabled
|
88 |
+
remove_gpu_after_training = gr.Checkbox(
|
89 |
+
label='Remove GPU after training', value=False, interactive=False)
|
90 |
+
run_button = gr.Button('Start Training')
|
91 |
+
|
92 |
+
with gr.Box():
|
93 |
+
gr.Markdown('Message')
|
94 |
+
message = gr.Markdown()
|
95 |
+
|
96 |
+
if pipe is not None:
|
97 |
+
run_button.click(fn=pipe.clear)
|
98 |
+
run_button.click(fn=trainer.run,
|
99 |
+
inputs=[
|
100 |
+
instance_images,
|
101 |
+
instance_prompt,
|
102 |
+
output_model_name,
|
103 |
+
delete_existing_model,
|
104 |
+
validation_prompt,
|
105 |
+
base_model,
|
106 |
+
resolution,
|
107 |
+
num_training_steps,
|
108 |
+
learning_rate,
|
109 |
+
gradient_accumulation,
|
110 |
+
seed,
|
111 |
+
fp16,
|
112 |
+
use_8bit_adam,
|
113 |
+
checkpointing_steps,
|
114 |
+
use_wandb,
|
115 |
+
validation_epochs,
|
116 |
+
upload_to_hub,
|
117 |
+
use_private_repo,
|
118 |
+
delete_existing_repo,
|
119 |
+
upload_to,
|
120 |
+
],
|
121 |
+
outputs=message)
|
122 |
+
return demo
|
123 |
+
|
124 |
+
|
125 |
+
if __name__ == '__main__':
|
126 |
+
trainer = Trainer()
|
127 |
+
demo = create_training_demo(trainer)
|
128 |
+
demo.queue(max_size=1).launch(share=False)
|
app_upload.py
ADDED
@@ -0,0 +1,95 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
|
3 |
+
from __future__ import annotations
|
4 |
+
|
5 |
+
import pathlib
|
6 |
+
|
7 |
+
import gradio as gr
|
8 |
+
import slugify
|
9 |
+
|
10 |
+
from constants import UploadTarget
|
11 |
+
from uploader import Uploader
|
12 |
+
from utils import find_exp_dirs
|
13 |
+
|
14 |
+
|
15 |
+
class LoRAModelUploader(Uploader):
|
16 |
+
def upload_lora_model(self, folder_path: str, repo_name: str,
|
17 |
+
upload_to: str, private: bool,
|
18 |
+
delete_existing_repo: bool) -> str:
|
19 |
+
if not repo_name:
|
20 |
+
repo_name = pathlib.Path(folder_path).name
|
21 |
+
repo_name = slugify.slugify(repo_name)
|
22 |
+
|
23 |
+
if upload_to == UploadTarget.PERSONAL_PROFILE.value:
|
24 |
+
organization = ''
|
25 |
+
elif upload_to == UploadTarget.LORA_LIBRARY.value:
|
26 |
+
organization = 'lora-library'
|
27 |
+
else:
|
28 |
+
raise ValueError
|
29 |
+
|
30 |
+
return self.upload(folder_path,
|
31 |
+
repo_name,
|
32 |
+
organization=organization,
|
33 |
+
private=private,
|
34 |
+
delete_existing_repo=delete_existing_repo)
|
35 |
+
|
36 |
+
|
37 |
+
def load_local_lora_model_list() -> dict:
|
38 |
+
choices = find_exp_dirs(ignore_repo=True)
|
39 |
+
return gr.update(choices=choices, value=choices[0] if choices else None)
|
40 |
+
|
41 |
+
|
42 |
+
def create_upload_demo(hf_token: str | None) -> gr.Blocks:
|
43 |
+
uploader = LoRAModelUploader(hf_token)
|
44 |
+
model_dirs = find_exp_dirs(ignore_repo=True)
|
45 |
+
|
46 |
+
with gr.Blocks() as demo:
|
47 |
+
with gr.Box():
|
48 |
+
gr.Markdown('Local Models')
|
49 |
+
reload_button = gr.Button('Reload Model List')
|
50 |
+
model_dir = gr.Dropdown(
|
51 |
+
label='LoRA Model ID',
|
52 |
+
choices=model_dirs,
|
53 |
+
value=model_dirs[0] if model_dirs else None)
|
54 |
+
gr.Markdown(
|
55 |
+
'- Models uploaded in training time will not be shown here.')
|
56 |
+
with gr.Box():
|
57 |
+
gr.Markdown('Upload Settings')
|
58 |
+
with gr.Row():
|
59 |
+
use_private_repo = gr.Checkbox(label='Private', value=False)
|
60 |
+
delete_existing_repo = gr.Checkbox(
|
61 |
+
label='Delete existing repo of the same name', value=False)
|
62 |
+
upload_to = gr.Radio(label='Upload to',
|
63 |
+
choices=[_.value for _ in UploadTarget],
|
64 |
+
value=UploadTarget.PERSONAL_PROFILE.value)
|
65 |
+
model_name = gr.Textbox(label='Model Name')
|
66 |
+
upload_button = gr.Button('Upload')
|
67 |
+
gr.Markdown('''
|
68 |
+
- You can upload your trained model to your personal profile (i.e. https://huggingface.co/{your_username}/{model_name}) or to the public [LoRA Concepts Library](https://huggingface.co/lora-library) (i.e. https://huggingface.co/lora-library/{model_name}).
|
69 |
+
''')
|
70 |
+
with gr.Box():
|
71 |
+
gr.Markdown('Message')
|
72 |
+
message = gr.Markdown()
|
73 |
+
|
74 |
+
reload_button.click(fn=load_local_lora_model_list,
|
75 |
+
inputs=None,
|
76 |
+
outputs=model_dir)
|
77 |
+
upload_button.click(fn=uploader.upload_lora_model,
|
78 |
+
inputs=[
|
79 |
+
model_dir,
|
80 |
+
model_name,
|
81 |
+
upload_to,
|
82 |
+
use_private_repo,
|
83 |
+
delete_existing_repo,
|
84 |
+
],
|
85 |
+
outputs=message)
|
86 |
+
|
87 |
+
return demo
|
88 |
+
|
89 |
+
|
90 |
+
if __name__ == '__main__':
|
91 |
+
import os
|
92 |
+
|
93 |
+
hf_token = os.getenv('HF_TOKEN')
|
94 |
+
demo = create_upload_demo(hf_token)
|
95 |
+
demo.queue(max_size=1).launch(share=False)
|
constants.py
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import enum
|
2 |
+
|
3 |
+
|
4 |
+
class UploadTarget(enum.Enum):
|
5 |
+
PERSONAL_PROFILE = 'Personal Profile'
|
6 |
+
LORA_LIBRARY = 'LoRA Library'
|
inference.py
CHANGED
@@ -2,78 +2,77 @@ from __future__ import annotations
|
|
2 |
|
3 |
import gc
|
4 |
import pathlib
|
5 |
-
import sys
|
6 |
|
7 |
import gradio as gr
|
8 |
import PIL.Image
|
9 |
import torch
|
10 |
-
from diffusers import
|
11 |
-
|
12 |
-
sys.path.insert(0, 'lora')
|
13 |
-
from lora_diffusion import monkeypatch_lora, tune_lora_scale
|
14 |
|
15 |
|
16 |
class InferencePipeline:
|
17 |
-
def __init__(self):
|
|
|
18 |
self.pipe = None
|
19 |
self.device = torch.device(
|
20 |
'cuda:0' if torch.cuda.is_available() else 'cpu')
|
21 |
-
self.
|
|
|
22 |
|
23 |
def clear(self) -> None:
|
24 |
-
self.
|
|
|
25 |
del self.pipe
|
26 |
self.pipe = None
|
27 |
torch.cuda.empty_cache()
|
28 |
gc.collect()
|
29 |
|
30 |
@staticmethod
|
31 |
-
def
|
32 |
-
|
33 |
-
return curr_dir / name
|
34 |
|
35 |
@staticmethod
|
36 |
-
def
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
path = parent_dir / text_encoder_filename
|
41 |
-
return path.as_posix() if path.exists() else ''
|
42 |
-
|
43 |
-
def load_pipe(self, model_id: str, lora_filename: str) -> None:
|
44 |
-
weight_path = self.get_lora_weight_path(lora_filename)
|
45 |
-
if weight_path == self.weight_path:
|
46 |
-
return
|
47 |
-
self.weight_path = weight_path
|
48 |
-
lora_weight = torch.load(self.weight_path, map_location=self.device)
|
49 |
-
|
50 |
-
if self.device.type == 'cpu':
|
51 |
-
pipe = StableDiffusionPipeline.from_pretrained(model_id)
|
52 |
else:
|
53 |
-
|
54 |
-
|
55 |
-
pipe = pipe.to(self.device)
|
56 |
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
lora_text_encoder_weight = torch.load(
|
63 |
-
lora_text_encoder_weight_path, map_location=self.device)
|
64 |
-
monkeypatch_lora(pipe.text_encoder,
|
65 |
-
lora_text_encoder_weight,
|
66 |
-
target_replace_module=['CLIPAttention'])
|
67 |
|
68 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
|
70 |
def run(
|
71 |
self,
|
72 |
-
|
73 |
-
lora_weight_name: str,
|
74 |
prompt: str,
|
75 |
-
alpha: float,
|
76 |
-
alpha_for_text: float,
|
77 |
seed: int,
|
78 |
n_steps: int,
|
79 |
guidance_scale: float,
|
@@ -81,11 +80,9 @@ class InferencePipeline:
|
|
81 |
if not torch.cuda.is_available():
|
82 |
raise gr.Error('CUDA is not available.')
|
83 |
|
84 |
-
self.load_pipe(
|
85 |
|
86 |
generator = torch.Generator(device=self.device).manual_seed(seed)
|
87 |
-
tune_lora_scale(self.pipe.unet, alpha) # type: ignore
|
88 |
-
tune_lora_scale(self.pipe.text_encoder, alpha_for_text) # type: ignore
|
89 |
out = self.pipe(prompt,
|
90 |
num_inference_steps=n_steps,
|
91 |
guidance_scale=guidance_scale,
|
|
|
2 |
|
3 |
import gc
|
4 |
import pathlib
|
|
|
5 |
|
6 |
import gradio as gr
|
7 |
import PIL.Image
|
8 |
import torch
|
9 |
+
from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
|
10 |
+
from huggingface_hub import ModelCard
|
|
|
|
|
11 |
|
12 |
|
13 |
class InferencePipeline:
|
14 |
+
def __init__(self, hf_token: str | None = None):
|
15 |
+
self.hf_token = hf_token
|
16 |
self.pipe = None
|
17 |
self.device = torch.device(
|
18 |
'cuda:0' if torch.cuda.is_available() else 'cpu')
|
19 |
+
self.lora_model_id = None
|
20 |
+
self.base_model_id = None
|
21 |
|
22 |
def clear(self) -> None:
|
23 |
+
self.lora_model_id = None
|
24 |
+
self.base_model_id = None
|
25 |
del self.pipe
|
26 |
self.pipe = None
|
27 |
torch.cuda.empty_cache()
|
28 |
gc.collect()
|
29 |
|
30 |
@staticmethod
|
31 |
+
def check_if_model_is_local(lora_model_id: str) -> bool:
|
32 |
+
return pathlib.Path(lora_model_id).exists()
|
|
|
33 |
|
34 |
@staticmethod
|
35 |
+
def get_model_card(model_id: str,
|
36 |
+
hf_token: str | None = None) -> ModelCard:
|
37 |
+
if InferencePipeline.check_if_model_is_local(model_id):
|
38 |
+
card_path = (pathlib.Path(model_id) / 'README.md').as_posix()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
else:
|
40 |
+
card_path = model_id
|
41 |
+
return ModelCard.load(card_path, token=hf_token)
|
|
|
42 |
|
43 |
+
@staticmethod
|
44 |
+
def get_base_model_info(lora_model_id: str,
|
45 |
+
hf_token: str | None = None) -> str:
|
46 |
+
card = InferencePipeline.get_model_card(lora_model_id, hf_token)
|
47 |
+
return card.data.base_model
|
|
|
|
|
|
|
|
|
|
|
48 |
|
49 |
+
def load_pipe(self, lora_model_id: str) -> None:
|
50 |
+
if lora_model_id == self.lora_model_id:
|
51 |
+
return
|
52 |
+
base_model_id = self.get_base_model_info(lora_model_id, self.hf_token)
|
53 |
+
if base_model_id != self.base_model_id:
|
54 |
+
if self.device.type == 'cpu':
|
55 |
+
pipe = DiffusionPipeline.from_pretrained(
|
56 |
+
base_model_id, use_auth_token=self.hf_token)
|
57 |
+
else:
|
58 |
+
pipe = DiffusionPipeline.from_pretrained(
|
59 |
+
base_model_id,
|
60 |
+
torch_dtype=torch.float16,
|
61 |
+
use_auth_token=self.hf_token)
|
62 |
+
pipe = pipe.to(self.device)
|
63 |
+
pipe.scheduler = DPMSolverMultistepScheduler.from_config(
|
64 |
+
pipe.scheduler.config)
|
65 |
+
self.pipe = pipe
|
66 |
+
self.pipe.unet.load_attn_procs( # type: ignore
|
67 |
+
lora_model_id, use_auth_token=self.hf_token)
|
68 |
+
|
69 |
+
self.lora_model_id = lora_model_id # type: ignore
|
70 |
+
self.base_model_id = base_model_id # type: ignore
|
71 |
|
72 |
def run(
|
73 |
self,
|
74 |
+
lora_model_id: str,
|
|
|
75 |
prompt: str,
|
|
|
|
|
76 |
seed: int,
|
77 |
n_steps: int,
|
78 |
guidance_scale: float,
|
|
|
80 |
if not torch.cuda.is_available():
|
81 |
raise gr.Error('CUDA is not available.')
|
82 |
|
83 |
+
self.load_pipe(lora_model_id)
|
84 |
|
85 |
generator = torch.Generator(device=self.device).manual_seed(seed)
|
|
|
|
|
86 |
out = self.pipe(prompt,
|
87 |
num_inference_steps=n_steps,
|
88 |
guidance_scale=guidance_scale,
|
lora
DELETED
@@ -1 +0,0 @@
|
|
1 |
-
Subproject commit 26787a09bff4ebcb08f0ad4e848b67bce4389a7a
|
|
|
|
requirements.txt
CHANGED
@@ -1,10 +1,12 @@
|
|
1 |
accelerate==0.15.0
|
2 |
-
bitsandbytes==0.
|
3 |
-
|
|
|
4 |
ftfy==6.1.1
|
5 |
-
|
6 |
-
|
7 |
-
|
|
|
|
|
8 |
transformers==4.25.1
|
9 |
-
|
10 |
-
xformers==0.0.13
|
|
|
1 |
accelerate==0.15.0
|
2 |
+
bitsandbytes==0.36.0.post2
|
3 |
+
datasets==2.8.0
|
4 |
+
git+https://github.com/huggingface/diffusers@a66f2baeb782e091dde4e1e6394e46f169e5ba58#egg=diffusers
|
5 |
ftfy==6.1.1
|
6 |
+
gradio==3.14.0
|
7 |
+
Pillow==9.4.0
|
8 |
+
python-slugify==7.0.0
|
9 |
+
torch==1.13.1
|
10 |
+
torchvision==0.14.1
|
11 |
transformers==4.25.1
|
12 |
+
wandb==0.13.9
|
|
train_dreambooth_lora.py
ADDED
@@ -0,0 +1,1018 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# This file is adapted from https://github.com/huggingface/diffusers/blob/a66f2baeb782e091dde4e1e6394e46f169e5ba58/examples/dreambooth/train_dreambooth_lora.py
|
3 |
+
# The original license is as below.
|
4 |
+
#
|
5 |
+
# coding=utf-8
|
6 |
+
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
|
7 |
+
#
|
8 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
9 |
+
# you may not use this file except in compliance with the License.
|
10 |
+
# You may obtain a copy of the License at
|
11 |
+
#
|
12 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
13 |
+
#
|
14 |
+
# Unless required by applicable law or agreed to in writing, software
|
15 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
16 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
17 |
+
# See the License for the specific language governing permissions and
|
18 |
+
|
19 |
+
import argparse
|
20 |
+
import hashlib
|
21 |
+
import logging
|
22 |
+
import math
|
23 |
+
import os
|
24 |
+
import warnings
|
25 |
+
from pathlib import Path
|
26 |
+
from typing import Optional
|
27 |
+
|
28 |
+
import torch
|
29 |
+
import torch.nn.functional as F
|
30 |
+
import torch.utils.checkpoint
|
31 |
+
from torch.utils.data import Dataset
|
32 |
+
|
33 |
+
import datasets
|
34 |
+
import diffusers
|
35 |
+
import transformers
|
36 |
+
from accelerate import Accelerator
|
37 |
+
from accelerate.logging import get_logger
|
38 |
+
from accelerate.utils import set_seed
|
39 |
+
from diffusers import (
|
40 |
+
AutoencoderKL,
|
41 |
+
DDPMScheduler,
|
42 |
+
DiffusionPipeline,
|
43 |
+
DPMSolverMultistepScheduler,
|
44 |
+
UNet2DConditionModel,
|
45 |
+
)
|
46 |
+
from diffusers.loaders import AttnProcsLayers
|
47 |
+
from diffusers.models.cross_attention import LoRACrossAttnProcessor
|
48 |
+
from diffusers.optimization import get_scheduler
|
49 |
+
from diffusers.utils import check_min_version, is_wandb_available
|
50 |
+
from diffusers.utils.import_utils import is_xformers_available
|
51 |
+
from huggingface_hub import HfFolder, Repository, create_repo, delete_repo, whoami
|
52 |
+
from PIL import Image
|
53 |
+
from torchvision import transforms
|
54 |
+
from tqdm.auto import tqdm
|
55 |
+
from transformers import AutoTokenizer, PretrainedConfig
|
56 |
+
|
57 |
+
|
58 |
+
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
59 |
+
check_min_version("0.12.0.dev0")
|
60 |
+
|
61 |
+
logger = get_logger(__name__)
|
62 |
+
|
63 |
+
|
64 |
+
def save_model_card(repo_name, base_model, instance_prompt, test_prompt="", images=None, repo_folder=""):
|
65 |
+
img_str = f"Test prompt: {test_prompt}\n" if test_prompt else ""
|
66 |
+
for i, image in enumerate(images or []):
|
67 |
+
image.save(os.path.join(repo_folder, f"image_{i}.png"))
|
68 |
+
img_str += f"![img_{i}](./image_{i}.png)\n"
|
69 |
+
|
70 |
+
yaml = f"""
|
71 |
+
---
|
72 |
+
license: creativeml-openrail-m
|
73 |
+
base_model: {base_model}
|
74 |
+
instance_prompt: {instance_prompt}
|
75 |
+
tags:
|
76 |
+
- stable-diffusion
|
77 |
+
- stable-diffusion-diffusers
|
78 |
+
- text-to-image
|
79 |
+
- diffusers
|
80 |
+
inference: true
|
81 |
+
---
|
82 |
+
"""
|
83 |
+
model_card = f"""
|
84 |
+
# LoRA DreamBooth - {repo_name}
|
85 |
+
|
86 |
+
These are LoRA adaption weights for [{base_model}](https://huggingface.co/{base_model}). The weights were trained on the instance prompt "{instance_prompt}" using [DreamBooth](https://dreambooth.github.io/). You can find some example images in the following.\n
|
87 |
+
{img_str}
|
88 |
+
"""
|
89 |
+
with open(os.path.join(repo_folder, "README.md"), "w") as f:
|
90 |
+
f.write(yaml + model_card)
|
91 |
+
|
92 |
+
|
93 |
+
def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: str, revision: str):
|
94 |
+
text_encoder_config = PretrainedConfig.from_pretrained(
|
95 |
+
pretrained_model_name_or_path,
|
96 |
+
subfolder="text_encoder",
|
97 |
+
revision=revision,
|
98 |
+
)
|
99 |
+
model_class = text_encoder_config.architectures[0]
|
100 |
+
|
101 |
+
if model_class == "CLIPTextModel":
|
102 |
+
from transformers import CLIPTextModel
|
103 |
+
|
104 |
+
return CLIPTextModel
|
105 |
+
elif model_class == "RobertaSeriesModelWithTransformation":
|
106 |
+
from diffusers.pipelines.alt_diffusion.modeling_roberta_series import RobertaSeriesModelWithTransformation
|
107 |
+
|
108 |
+
return RobertaSeriesModelWithTransformation
|
109 |
+
else:
|
110 |
+
raise ValueError(f"{model_class} is not supported.")
|
111 |
+
|
112 |
+
|
113 |
+
def parse_args(input_args=None):
|
114 |
+
parser = argparse.ArgumentParser(description="Simple example of a training script.")
|
115 |
+
parser.add_argument(
|
116 |
+
"--pretrained_model_name_or_path",
|
117 |
+
type=str,
|
118 |
+
default=None,
|
119 |
+
required=True,
|
120 |
+
help="Path to pretrained model or model identifier from huggingface.co/models.",
|
121 |
+
)
|
122 |
+
parser.add_argument(
|
123 |
+
"--revision",
|
124 |
+
type=str,
|
125 |
+
default=None,
|
126 |
+
required=False,
|
127 |
+
help="Revision of pretrained model identifier from huggingface.co/models.",
|
128 |
+
)
|
129 |
+
parser.add_argument(
|
130 |
+
"--tokenizer_name",
|
131 |
+
type=str,
|
132 |
+
default=None,
|
133 |
+
help="Pretrained tokenizer name or path if not the same as model_name",
|
134 |
+
)
|
135 |
+
parser.add_argument(
|
136 |
+
"--instance_data_dir",
|
137 |
+
type=str,
|
138 |
+
default=None,
|
139 |
+
required=True,
|
140 |
+
help="A folder containing the training data of instance images.",
|
141 |
+
)
|
142 |
+
parser.add_argument(
|
143 |
+
"--class_data_dir",
|
144 |
+
type=str,
|
145 |
+
default=None,
|
146 |
+
required=False,
|
147 |
+
help="A folder containing the training data of class images.",
|
148 |
+
)
|
149 |
+
parser.add_argument(
|
150 |
+
"--instance_prompt",
|
151 |
+
type=str,
|
152 |
+
default=None,
|
153 |
+
required=True,
|
154 |
+
help="The prompt with identifier specifying the instance",
|
155 |
+
)
|
156 |
+
parser.add_argument(
|
157 |
+
"--class_prompt",
|
158 |
+
type=str,
|
159 |
+
default=None,
|
160 |
+
help="The prompt to specify images in the same class as provided instance images.",
|
161 |
+
)
|
162 |
+
parser.add_argument(
|
163 |
+
"--validation_prompt",
|
164 |
+
type=str,
|
165 |
+
default=None,
|
166 |
+
help="A prompt that is used during validation to verify that the model is learning.",
|
167 |
+
)
|
168 |
+
parser.add_argument(
|
169 |
+
"--num_validation_images",
|
170 |
+
type=int,
|
171 |
+
default=4,
|
172 |
+
help="Number of images that should be generated during validation with `validation_prompt`.",
|
173 |
+
)
|
174 |
+
parser.add_argument(
|
175 |
+
"--validation_epochs",
|
176 |
+
type=int,
|
177 |
+
default=50,
|
178 |
+
help=(
|
179 |
+
"Run dreambooth validation every X epochs. Dreambooth validation consists of running the prompt"
|
180 |
+
" `args.validation_prompt` multiple times: `args.num_validation_images`."
|
181 |
+
),
|
182 |
+
)
|
183 |
+
parser.add_argument(
|
184 |
+
"--with_prior_preservation",
|
185 |
+
default=False,
|
186 |
+
action="store_true",
|
187 |
+
help="Flag to add prior preservation loss.",
|
188 |
+
)
|
189 |
+
parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.")
|
190 |
+
parser.add_argument(
|
191 |
+
"--num_class_images",
|
192 |
+
type=int,
|
193 |
+
default=100,
|
194 |
+
help=(
|
195 |
+
"Minimal class images for prior preservation loss. If there are not enough images already present in"
|
196 |
+
" class_data_dir, additional images will be sampled with class_prompt."
|
197 |
+
),
|
198 |
+
)
|
199 |
+
parser.add_argument(
|
200 |
+
"--output_dir",
|
201 |
+
type=str,
|
202 |
+
default="lora-dreambooth-model",
|
203 |
+
help="The output directory where the model predictions and checkpoints will be written.",
|
204 |
+
)
|
205 |
+
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
|
206 |
+
parser.add_argument(
|
207 |
+
"--resolution",
|
208 |
+
type=int,
|
209 |
+
default=512,
|
210 |
+
help=(
|
211 |
+
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
|
212 |
+
" resolution"
|
213 |
+
),
|
214 |
+
)
|
215 |
+
parser.add_argument(
|
216 |
+
"--center_crop", action="store_true", help="Whether to center crop images before resizing to resolution"
|
217 |
+
)
|
218 |
+
parser.add_argument(
|
219 |
+
"--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
|
220 |
+
)
|
221 |
+
parser.add_argument(
|
222 |
+
"--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images."
|
223 |
+
)
|
224 |
+
parser.add_argument("--num_train_epochs", type=int, default=1)
|
225 |
+
parser.add_argument(
|
226 |
+
"--max_train_steps",
|
227 |
+
type=int,
|
228 |
+
default=None,
|
229 |
+
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
|
230 |
+
)
|
231 |
+
parser.add_argument(
|
232 |
+
"--checkpointing_steps",
|
233 |
+
type=int,
|
234 |
+
default=500,
|
235 |
+
help=(
|
236 |
+
"Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
|
237 |
+
" checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
|
238 |
+
" training using `--resume_from_checkpoint`."
|
239 |
+
),
|
240 |
+
)
|
241 |
+
parser.add_argument(
|
242 |
+
"--resume_from_checkpoint",
|
243 |
+
type=str,
|
244 |
+
default=None,
|
245 |
+
help=(
|
246 |
+
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
|
247 |
+
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
|
248 |
+
),
|
249 |
+
)
|
250 |
+
parser.add_argument(
|
251 |
+
"--gradient_accumulation_steps",
|
252 |
+
type=int,
|
253 |
+
default=1,
|
254 |
+
help="Number of updates steps to accumulate before performing a backward/update pass.",
|
255 |
+
)
|
256 |
+
parser.add_argument(
|
257 |
+
"--gradient_checkpointing",
|
258 |
+
action="store_true",
|
259 |
+
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
|
260 |
+
)
|
261 |
+
parser.add_argument(
|
262 |
+
"--learning_rate",
|
263 |
+
type=float,
|
264 |
+
default=5e-4,
|
265 |
+
help="Initial learning rate (after the potential warmup period) to use.",
|
266 |
+
)
|
267 |
+
parser.add_argument(
|
268 |
+
"--scale_lr",
|
269 |
+
action="store_true",
|
270 |
+
default=False,
|
271 |
+
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
|
272 |
+
)
|
273 |
+
parser.add_argument(
|
274 |
+
"--lr_scheduler",
|
275 |
+
type=str,
|
276 |
+
default="constant",
|
277 |
+
help=(
|
278 |
+
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
|
279 |
+
' "constant", "constant_with_warmup"]'
|
280 |
+
),
|
281 |
+
)
|
282 |
+
parser.add_argument(
|
283 |
+
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
|
284 |
+
)
|
285 |
+
parser.add_argument(
|
286 |
+
"--lr_num_cycles",
|
287 |
+
type=int,
|
288 |
+
default=1,
|
289 |
+
help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
|
290 |
+
)
|
291 |
+
parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
|
292 |
+
parser.add_argument(
|
293 |
+
"--dataloader_num_workers",
|
294 |
+
type=int,
|
295 |
+
default=0,
|
296 |
+
help=(
|
297 |
+
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
|
298 |
+
),
|
299 |
+
)
|
300 |
+
parser.add_argument(
|
301 |
+
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
|
302 |
+
)
|
303 |
+
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
|
304 |
+
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
|
305 |
+
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
|
306 |
+
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
|
307 |
+
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
|
308 |
+
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
|
309 |
+
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
|
310 |
+
parser.add_argument(
|
311 |
+
"--hub_model_id",
|
312 |
+
type=str,
|
313 |
+
default=None,
|
314 |
+
help="The name of the repository to keep in sync with the local `output_dir`.",
|
315 |
+
)
|
316 |
+
parser.add_argument(
|
317 |
+
"--logging_dir",
|
318 |
+
type=str,
|
319 |
+
default="logs",
|
320 |
+
help=(
|
321 |
+
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
|
322 |
+
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
|
323 |
+
),
|
324 |
+
)
|
325 |
+
parser.add_argument(
|
326 |
+
"--allow_tf32",
|
327 |
+
action="store_true",
|
328 |
+
help=(
|
329 |
+
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
|
330 |
+
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
|
331 |
+
),
|
332 |
+
)
|
333 |
+
parser.add_argument(
|
334 |
+
"--report_to",
|
335 |
+
type=str,
|
336 |
+
default="tensorboard",
|
337 |
+
help=(
|
338 |
+
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
|
339 |
+
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
|
340 |
+
),
|
341 |
+
)
|
342 |
+
parser.add_argument(
|
343 |
+
"--mixed_precision",
|
344 |
+
type=str,
|
345 |
+
default=None,
|
346 |
+
choices=["no", "fp16", "bf16"],
|
347 |
+
help=(
|
348 |
+
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
|
349 |
+
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
|
350 |
+
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
|
351 |
+
),
|
352 |
+
)
|
353 |
+
parser.add_argument(
|
354 |
+
"--prior_generation_precision",
|
355 |
+
type=str,
|
356 |
+
default=None,
|
357 |
+
choices=["no", "fp32", "fp16", "bf16"],
|
358 |
+
help=(
|
359 |
+
"Choose prior generation precision between fp32, fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
|
360 |
+
" 1.10.and an Nvidia Ampere GPU. Default to fp16 if a GPU is available else fp32."
|
361 |
+
),
|
362 |
+
)
|
363 |
+
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
|
364 |
+
parser.add_argument(
|
365 |
+
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
|
366 |
+
)
|
367 |
+
parser.add_argument("--private_repo", action="store_true")
|
368 |
+
parser.add_argument("--delete_existing_repo", action="store_true")
|
369 |
+
parser.add_argument("--upload_to_lora_library", action="store_true")
|
370 |
+
|
371 |
+
if input_args is not None:
|
372 |
+
args = parser.parse_args(input_args)
|
373 |
+
else:
|
374 |
+
args = parser.parse_args()
|
375 |
+
|
376 |
+
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
|
377 |
+
if env_local_rank != -1 and env_local_rank != args.local_rank:
|
378 |
+
args.local_rank = env_local_rank
|
379 |
+
|
380 |
+
if args.with_prior_preservation:
|
381 |
+
if args.class_data_dir is None:
|
382 |
+
raise ValueError("You must specify a data directory for class images.")
|
383 |
+
if args.class_prompt is None:
|
384 |
+
raise ValueError("You must specify prompt for class images.")
|
385 |
+
else:
|
386 |
+
# logger is not available yet
|
387 |
+
if args.class_data_dir is not None:
|
388 |
+
warnings.warn("You need not use --class_data_dir without --with_prior_preservation.")
|
389 |
+
if args.class_prompt is not None:
|
390 |
+
warnings.warn("You need not use --class_prompt without --with_prior_preservation.")
|
391 |
+
|
392 |
+
return args
|
393 |
+
|
394 |
+
|
395 |
+
class DreamBoothDataset(Dataset):
|
396 |
+
"""
|
397 |
+
A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
|
398 |
+
It pre-processes the images and the tokenizes prompts.
|
399 |
+
"""
|
400 |
+
|
401 |
+
def __init__(
|
402 |
+
self,
|
403 |
+
instance_data_root,
|
404 |
+
instance_prompt,
|
405 |
+
tokenizer,
|
406 |
+
class_data_root=None,
|
407 |
+
class_prompt=None,
|
408 |
+
size=512,
|
409 |
+
center_crop=False,
|
410 |
+
):
|
411 |
+
self.size = size
|
412 |
+
self.center_crop = center_crop
|
413 |
+
self.tokenizer = tokenizer
|
414 |
+
|
415 |
+
self.instance_data_root = Path(instance_data_root)
|
416 |
+
if not self.instance_data_root.exists():
|
417 |
+
raise ValueError("Instance images root doesn't exists.")
|
418 |
+
|
419 |
+
self.instance_images_path = list(Path(instance_data_root).iterdir())
|
420 |
+
self.num_instance_images = len(self.instance_images_path)
|
421 |
+
self.instance_prompt = instance_prompt
|
422 |
+
self._length = self.num_instance_images
|
423 |
+
|
424 |
+
if class_data_root is not None:
|
425 |
+
self.class_data_root = Path(class_data_root)
|
426 |
+
self.class_data_root.mkdir(parents=True, exist_ok=True)
|
427 |
+
self.class_images_path = list(self.class_data_root.iterdir())
|
428 |
+
self.num_class_images = len(self.class_images_path)
|
429 |
+
self._length = max(self.num_class_images, self.num_instance_images)
|
430 |
+
self.class_prompt = class_prompt
|
431 |
+
else:
|
432 |
+
self.class_data_root = None
|
433 |
+
|
434 |
+
self.image_transforms = transforms.Compose(
|
435 |
+
[
|
436 |
+
transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
|
437 |
+
transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
|
438 |
+
transforms.ToTensor(),
|
439 |
+
transforms.Normalize([0.5], [0.5]),
|
440 |
+
]
|
441 |
+
)
|
442 |
+
|
443 |
+
def __len__(self):
|
444 |
+
return self._length
|
445 |
+
|
446 |
+
def __getitem__(self, index):
|
447 |
+
example = {}
|
448 |
+
instance_image = Image.open(self.instance_images_path[index % self.num_instance_images])
|
449 |
+
if not instance_image.mode == "RGB":
|
450 |
+
instance_image = instance_image.convert("RGB")
|
451 |
+
example["instance_images"] = self.image_transforms(instance_image)
|
452 |
+
example["instance_prompt_ids"] = self.tokenizer(
|
453 |
+
self.instance_prompt,
|
454 |
+
truncation=True,
|
455 |
+
padding="max_length",
|
456 |
+
max_length=self.tokenizer.model_max_length,
|
457 |
+
return_tensors="pt",
|
458 |
+
).input_ids
|
459 |
+
|
460 |
+
if self.class_data_root:
|
461 |
+
class_image = Image.open(self.class_images_path[index % self.num_class_images])
|
462 |
+
if not class_image.mode == "RGB":
|
463 |
+
class_image = class_image.convert("RGB")
|
464 |
+
example["class_images"] = self.image_transforms(class_image)
|
465 |
+
example["class_prompt_ids"] = self.tokenizer(
|
466 |
+
self.class_prompt,
|
467 |
+
truncation=True,
|
468 |
+
padding="max_length",
|
469 |
+
max_length=self.tokenizer.model_max_length,
|
470 |
+
return_tensors="pt",
|
471 |
+
).input_ids
|
472 |
+
|
473 |
+
return example
|
474 |
+
|
475 |
+
|
476 |
+
def collate_fn(examples, with_prior_preservation=False):
|
477 |
+
input_ids = [example["instance_prompt_ids"] for example in examples]
|
478 |
+
pixel_values = [example["instance_images"] for example in examples]
|
479 |
+
|
480 |
+
# Concat class and instance examples for prior preservation.
|
481 |
+
# We do this to avoid doing two forward passes.
|
482 |
+
if with_prior_preservation:
|
483 |
+
input_ids += [example["class_prompt_ids"] for example in examples]
|
484 |
+
pixel_values += [example["class_images"] for example in examples]
|
485 |
+
|
486 |
+
pixel_values = torch.stack(pixel_values)
|
487 |
+
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
|
488 |
+
|
489 |
+
input_ids = torch.cat(input_ids, dim=0)
|
490 |
+
|
491 |
+
batch = {
|
492 |
+
"input_ids": input_ids,
|
493 |
+
"pixel_values": pixel_values,
|
494 |
+
}
|
495 |
+
return batch
|
496 |
+
|
497 |
+
|
498 |
+
class PromptDataset(Dataset):
|
499 |
+
"A simple dataset to prepare the prompts to generate class images on multiple GPUs."
|
500 |
+
|
501 |
+
def __init__(self, prompt, num_samples):
|
502 |
+
self.prompt = prompt
|
503 |
+
self.num_samples = num_samples
|
504 |
+
|
505 |
+
def __len__(self):
|
506 |
+
return self.num_samples
|
507 |
+
|
508 |
+
def __getitem__(self, index):
|
509 |
+
example = {}
|
510 |
+
example["prompt"] = self.prompt
|
511 |
+
example["index"] = index
|
512 |
+
return example
|
513 |
+
|
514 |
+
|
515 |
+
def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
|
516 |
+
if token is None:
|
517 |
+
token = HfFolder.get_token()
|
518 |
+
if organization is None:
|
519 |
+
username = whoami(token)["name"]
|
520 |
+
return f"{username}/{model_id}"
|
521 |
+
else:
|
522 |
+
return f"{organization}/{model_id}"
|
523 |
+
|
524 |
+
|
525 |
+
def main(args):
|
526 |
+
logging_dir = Path(args.output_dir, args.logging_dir)
|
527 |
+
|
528 |
+
accelerator = Accelerator(
|
529 |
+
gradient_accumulation_steps=args.gradient_accumulation_steps,
|
530 |
+
mixed_precision=args.mixed_precision,
|
531 |
+
log_with=args.report_to,
|
532 |
+
logging_dir=logging_dir,
|
533 |
+
)
|
534 |
+
|
535 |
+
if args.report_to == "wandb":
|
536 |
+
if not is_wandb_available():
|
537 |
+
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
|
538 |
+
import wandb
|
539 |
+
|
540 |
+
# Currently, it's not possible to do gradient accumulation when training two models with accelerate.accumulate
|
541 |
+
# This will be enabled soon in accelerate. For now, we don't allow gradient accumulation when training two models.
|
542 |
+
# TODO (patil-suraj): Remove this check when gradient accumulation with two models is enabled in accelerate.
|
543 |
+
# Make one log on every process with the configuration for debugging.
|
544 |
+
logging.basicConfig(
|
545 |
+
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
546 |
+
datefmt="%m/%d/%Y %H:%M:%S",
|
547 |
+
level=logging.INFO,
|
548 |
+
)
|
549 |
+
logger.info(accelerator.state, main_process_only=False)
|
550 |
+
if accelerator.is_local_main_process:
|
551 |
+
datasets.utils.logging.set_verbosity_warning()
|
552 |
+
transformers.utils.logging.set_verbosity_warning()
|
553 |
+
diffusers.utils.logging.set_verbosity_info()
|
554 |
+
else:
|
555 |
+
datasets.utils.logging.set_verbosity_error()
|
556 |
+
transformers.utils.logging.set_verbosity_error()
|
557 |
+
diffusers.utils.logging.set_verbosity_error()
|
558 |
+
|
559 |
+
# If passed along, set the training seed now.
|
560 |
+
if args.seed is not None:
|
561 |
+
set_seed(args.seed)
|
562 |
+
|
563 |
+
# Generate class images if prior preservation is enabled.
|
564 |
+
if args.with_prior_preservation:
|
565 |
+
class_images_dir = Path(args.class_data_dir)
|
566 |
+
if not class_images_dir.exists():
|
567 |
+
class_images_dir.mkdir(parents=True)
|
568 |
+
cur_class_images = len(list(class_images_dir.iterdir()))
|
569 |
+
|
570 |
+
if cur_class_images < args.num_class_images:
|
571 |
+
torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
|
572 |
+
if args.prior_generation_precision == "fp32":
|
573 |
+
torch_dtype = torch.float32
|
574 |
+
elif args.prior_generation_precision == "fp16":
|
575 |
+
torch_dtype = torch.float16
|
576 |
+
elif args.prior_generation_precision == "bf16":
|
577 |
+
torch_dtype = torch.bfloat16
|
578 |
+
pipeline = DiffusionPipeline.from_pretrained(
|
579 |
+
args.pretrained_model_name_or_path,
|
580 |
+
torch_dtype=torch_dtype,
|
581 |
+
safety_checker=None,
|
582 |
+
revision=args.revision,
|
583 |
+
)
|
584 |
+
pipeline.set_progress_bar_config(disable=True)
|
585 |
+
|
586 |
+
num_new_images = args.num_class_images - cur_class_images
|
587 |
+
logger.info(f"Number of class images to sample: {num_new_images}.")
|
588 |
+
|
589 |
+
sample_dataset = PromptDataset(args.class_prompt, num_new_images)
|
590 |
+
sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=args.sample_batch_size)
|
591 |
+
|
592 |
+
sample_dataloader = accelerator.prepare(sample_dataloader)
|
593 |
+
pipeline.to(accelerator.device)
|
594 |
+
|
595 |
+
for example in tqdm(
|
596 |
+
sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process
|
597 |
+
):
|
598 |
+
images = pipeline(example["prompt"]).images
|
599 |
+
|
600 |
+
for i, image in enumerate(images):
|
601 |
+
hash_image = hashlib.sha1(image.tobytes()).hexdigest()
|
602 |
+
image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
|
603 |
+
image.save(image_filename)
|
604 |
+
|
605 |
+
del pipeline
|
606 |
+
if torch.cuda.is_available():
|
607 |
+
torch.cuda.empty_cache()
|
608 |
+
|
609 |
+
# Handle the repository creation
|
610 |
+
if accelerator.is_main_process:
|
611 |
+
if args.push_to_hub:
|
612 |
+
if args.hub_model_id is None:
|
613 |
+
organization = 'lora-library' if args.upload_to_lora_library else None
|
614 |
+
repo_name = get_full_repo_name(Path(args.output_dir).name, organization=organization, token=args.hub_token)
|
615 |
+
else:
|
616 |
+
repo_name = args.hub_model_id
|
617 |
+
|
618 |
+
if args.delete_existing_repo:
|
619 |
+
try:
|
620 |
+
delete_repo(repo_name, token=args.hub_token)
|
621 |
+
except Exception:
|
622 |
+
pass
|
623 |
+
create_repo(repo_name, token=args.hub_token, private=args.private_repo)
|
624 |
+
repo = Repository(args.output_dir, clone_from=repo_name, token=args.hub_token)
|
625 |
+
|
626 |
+
with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
|
627 |
+
if "step_*" not in gitignore:
|
628 |
+
gitignore.write("step_*\n")
|
629 |
+
if "epoch_*" not in gitignore:
|
630 |
+
gitignore.write("epoch_*\n")
|
631 |
+
elif args.output_dir is not None:
|
632 |
+
os.makedirs(args.output_dir, exist_ok=True)
|
633 |
+
|
634 |
+
# Load the tokenizer
|
635 |
+
if args.tokenizer_name:
|
636 |
+
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
|
637 |
+
elif args.pretrained_model_name_or_path:
|
638 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
639 |
+
args.pretrained_model_name_or_path,
|
640 |
+
subfolder="tokenizer",
|
641 |
+
revision=args.revision,
|
642 |
+
use_fast=False,
|
643 |
+
)
|
644 |
+
|
645 |
+
# import correct text encoder class
|
646 |
+
text_encoder_cls = import_model_class_from_model_name_or_path(args.pretrained_model_name_or_path, args.revision)
|
647 |
+
|
648 |
+
# Load scheduler and models
|
649 |
+
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
|
650 |
+
text_encoder = text_encoder_cls.from_pretrained(
|
651 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
|
652 |
+
)
|
653 |
+
vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision)
|
654 |
+
unet = UNet2DConditionModel.from_pretrained(
|
655 |
+
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision
|
656 |
+
)
|
657 |
+
|
658 |
+
# We only train the additional adapter LoRA layers
|
659 |
+
vae.requires_grad_(False)
|
660 |
+
text_encoder.requires_grad_(False)
|
661 |
+
unet.requires_grad_(False)
|
662 |
+
|
663 |
+
# For mixed precision training we cast the text_encoder and vae weights to half-precision
|
664 |
+
# as these models are only used for inference, keeping weights in full precision is not required.
|
665 |
+
weight_dtype = torch.float32
|
666 |
+
if accelerator.mixed_precision == "fp16":
|
667 |
+
weight_dtype = torch.float16
|
668 |
+
elif accelerator.mixed_precision == "bf16":
|
669 |
+
weight_dtype = torch.bfloat16
|
670 |
+
|
671 |
+
# Move unet, vae and text_encoder to device and cast to weight_dtype
|
672 |
+
unet.to(accelerator.device, dtype=weight_dtype)
|
673 |
+
vae.to(accelerator.device, dtype=weight_dtype)
|
674 |
+
text_encoder.to(accelerator.device, dtype=weight_dtype)
|
675 |
+
|
676 |
+
if args.enable_xformers_memory_efficient_attention:
|
677 |
+
if is_xformers_available():
|
678 |
+
unet.enable_xformers_memory_efficient_attention()
|
679 |
+
else:
|
680 |
+
raise ValueError("xformers is not available. Make sure it is installed correctly")
|
681 |
+
|
682 |
+
# now we will add new LoRA weights to the attention layers
|
683 |
+
# It's important to realize here how many attention weights will be added and of which sizes
|
684 |
+
# The sizes of the attention layers consist only of two different variables:
|
685 |
+
# 1) - the "hidden_size", which is increased according to `unet.config.block_out_channels`.
|
686 |
+
# 2) - the "cross attention size", which is set to `unet.config.cross_attention_dim`.
|
687 |
+
|
688 |
+
# Let's first see how many attention processors we will have to set.
|
689 |
+
# For Stable Diffusion, it should be equal to:
|
690 |
+
# - down blocks (2x attention layers) * (2x transformer layers) * (3x down blocks) = 12
|
691 |
+
# - mid blocks (2x attention layers) * (1x transformer layers) * (1x mid blocks) = 2
|
692 |
+
# - up blocks (2x attention layers) * (3x transformer layers) * (3x down blocks) = 18
|
693 |
+
# => 32 layers
|
694 |
+
|
695 |
+
# Set correct lora layers
|
696 |
+
lora_attn_procs = {}
|
697 |
+
for name in unet.attn_processors.keys():
|
698 |
+
cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
|
699 |
+
if name.startswith("mid_block"):
|
700 |
+
hidden_size = unet.config.block_out_channels[-1]
|
701 |
+
elif name.startswith("up_blocks"):
|
702 |
+
block_id = int(name[len("up_blocks.")])
|
703 |
+
hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
|
704 |
+
elif name.startswith("down_blocks"):
|
705 |
+
block_id = int(name[len("down_blocks.")])
|
706 |
+
hidden_size = unet.config.block_out_channels[block_id]
|
707 |
+
|
708 |
+
lora_attn_procs[name] = LoRACrossAttnProcessor(
|
709 |
+
hidden_size=hidden_size, cross_attention_dim=cross_attention_dim
|
710 |
+
)
|
711 |
+
|
712 |
+
unet.set_attn_processor(lora_attn_procs)
|
713 |
+
lora_layers = AttnProcsLayers(unet.attn_processors)
|
714 |
+
|
715 |
+
accelerator.register_for_checkpointing(lora_layers)
|
716 |
+
|
717 |
+
if args.scale_lr:
|
718 |
+
args.learning_rate = (
|
719 |
+
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
|
720 |
+
)
|
721 |
+
|
722 |
+
# Enable TF32 for faster training on Ampere GPUs,
|
723 |
+
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
|
724 |
+
if args.allow_tf32:
|
725 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
726 |
+
|
727 |
+
if args.scale_lr:
|
728 |
+
args.learning_rate = (
|
729 |
+
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
|
730 |
+
)
|
731 |
+
|
732 |
+
# Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
|
733 |
+
if args.use_8bit_adam:
|
734 |
+
try:
|
735 |
+
import bitsandbytes as bnb
|
736 |
+
except ImportError:
|
737 |
+
raise ImportError(
|
738 |
+
"To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
|
739 |
+
)
|
740 |
+
|
741 |
+
optimizer_class = bnb.optim.AdamW8bit
|
742 |
+
else:
|
743 |
+
optimizer_class = torch.optim.AdamW
|
744 |
+
|
745 |
+
# Optimizer creation
|
746 |
+
optimizer = optimizer_class(
|
747 |
+
lora_layers.parameters(),
|
748 |
+
lr=args.learning_rate,
|
749 |
+
betas=(args.adam_beta1, args.adam_beta2),
|
750 |
+
weight_decay=args.adam_weight_decay,
|
751 |
+
eps=args.adam_epsilon,
|
752 |
+
)
|
753 |
+
|
754 |
+
# Dataset and DataLoaders creation:
|
755 |
+
train_dataset = DreamBoothDataset(
|
756 |
+
instance_data_root=args.instance_data_dir,
|
757 |
+
instance_prompt=args.instance_prompt,
|
758 |
+
class_data_root=args.class_data_dir if args.with_prior_preservation else None,
|
759 |
+
class_prompt=args.class_prompt,
|
760 |
+
tokenizer=tokenizer,
|
761 |
+
size=args.resolution,
|
762 |
+
center_crop=args.center_crop,
|
763 |
+
)
|
764 |
+
|
765 |
+
train_dataloader = torch.utils.data.DataLoader(
|
766 |
+
train_dataset,
|
767 |
+
batch_size=args.train_batch_size,
|
768 |
+
shuffle=True,
|
769 |
+
collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
|
770 |
+
num_workers=args.dataloader_num_workers,
|
771 |
+
)
|
772 |
+
|
773 |
+
# Scheduler and math around the number of training steps.
|
774 |
+
overrode_max_train_steps = False
|
775 |
+
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
776 |
+
if args.max_train_steps is None:
|
777 |
+
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
778 |
+
overrode_max_train_steps = True
|
779 |
+
|
780 |
+
lr_scheduler = get_scheduler(
|
781 |
+
args.lr_scheduler,
|
782 |
+
optimizer=optimizer,
|
783 |
+
num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
|
784 |
+
num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
|
785 |
+
num_cycles=args.lr_num_cycles,
|
786 |
+
power=args.lr_power,
|
787 |
+
)
|
788 |
+
|
789 |
+
# Prepare everything with our `accelerator`.
|
790 |
+
lora_layers, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
791 |
+
lora_layers, optimizer, train_dataloader, lr_scheduler
|
792 |
+
)
|
793 |
+
|
794 |
+
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
|
795 |
+
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
796 |
+
if overrode_max_train_steps:
|
797 |
+
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
798 |
+
# Afterwards we recalculate our number of training epochs
|
799 |
+
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
|
800 |
+
|
801 |
+
# We need to initialize the trackers we use, and also store our configuration.
|
802 |
+
# The trackers initializes automatically on the main process.
|
803 |
+
if accelerator.is_main_process:
|
804 |
+
accelerator.init_trackers("dreambooth-lora", config=vars(args))
|
805 |
+
|
806 |
+
# Train!
|
807 |
+
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
|
808 |
+
|
809 |
+
logger.info("***** Running training *****")
|
810 |
+
logger.info(f" Num examples = {len(train_dataset)}")
|
811 |
+
logger.info(f" Num batches each epoch = {len(train_dataloader)}")
|
812 |
+
logger.info(f" Num Epochs = {args.num_train_epochs}")
|
813 |
+
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
|
814 |
+
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
|
815 |
+
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
|
816 |
+
logger.info(f" Total optimization steps = {args.max_train_steps}")
|
817 |
+
global_step = 0
|
818 |
+
first_epoch = 0
|
819 |
+
|
820 |
+
# Potentially load in the weights and states from a previous save
|
821 |
+
if args.resume_from_checkpoint:
|
822 |
+
if args.resume_from_checkpoint != "latest":
|
823 |
+
path = os.path.basename(args.resume_from_checkpoint)
|
824 |
+
else:
|
825 |
+
# Get the mos recent checkpoint
|
826 |
+
dirs = os.listdir(args.output_dir)
|
827 |
+
dirs = [d for d in dirs if d.startswith("checkpoint")]
|
828 |
+
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
|
829 |
+
path = dirs[-1]
|
830 |
+
accelerator.print(f"Resuming from checkpoint {path}")
|
831 |
+
accelerator.load_state(os.path.join(args.output_dir, path))
|
832 |
+
global_step = int(path.split("-")[1])
|
833 |
+
|
834 |
+
resume_global_step = global_step * args.gradient_accumulation_steps
|
835 |
+
first_epoch = resume_global_step // num_update_steps_per_epoch
|
836 |
+
resume_step = resume_global_step % num_update_steps_per_epoch
|
837 |
+
|
838 |
+
# Only show the progress bar once on each machine.
|
839 |
+
progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
|
840 |
+
progress_bar.set_description("Steps")
|
841 |
+
|
842 |
+
for epoch in range(first_epoch, args.num_train_epochs):
|
843 |
+
unet.train()
|
844 |
+
for step, batch in enumerate(train_dataloader):
|
845 |
+
# Skip steps until we reach the resumed step
|
846 |
+
if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
|
847 |
+
if step % args.gradient_accumulation_steps == 0:
|
848 |
+
progress_bar.update(1)
|
849 |
+
continue
|
850 |
+
|
851 |
+
with accelerator.accumulate(unet):
|
852 |
+
# Convert images to latent space
|
853 |
+
latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
|
854 |
+
latents = latents * 0.18215
|
855 |
+
|
856 |
+
# Sample noise that we'll add to the latents
|
857 |
+
noise = torch.randn_like(latents)
|
858 |
+
bsz = latents.shape[0]
|
859 |
+
# Sample a random timestep for each image
|
860 |
+
timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
|
861 |
+
timesteps = timesteps.long()
|
862 |
+
|
863 |
+
# Add noise to the latents according to the noise magnitude at each timestep
|
864 |
+
# (this is the forward diffusion process)
|
865 |
+
noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
|
866 |
+
|
867 |
+
# Get the text embedding for conditioning
|
868 |
+
encoder_hidden_states = text_encoder(batch["input_ids"])[0]
|
869 |
+
|
870 |
+
# Predict the noise residual
|
871 |
+
model_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
|
872 |
+
|
873 |
+
# Get the target for loss depending on the prediction type
|
874 |
+
if noise_scheduler.config.prediction_type == "epsilon":
|
875 |
+
target = noise
|
876 |
+
elif noise_scheduler.config.prediction_type == "v_prediction":
|
877 |
+
target = noise_scheduler.get_velocity(latents, noise, timesteps)
|
878 |
+
else:
|
879 |
+
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
|
880 |
+
|
881 |
+
if args.with_prior_preservation:
|
882 |
+
# Chunk the noise and model_pred into two parts and compute the loss on each part separately.
|
883 |
+
model_pred, model_pred_prior = torch.chunk(model_pred, 2, dim=0)
|
884 |
+
target, target_prior = torch.chunk(target, 2, dim=0)
|
885 |
+
|
886 |
+
# Compute instance loss
|
887 |
+
loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
|
888 |
+
|
889 |
+
# Compute prior loss
|
890 |
+
prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean")
|
891 |
+
|
892 |
+
# Add the prior loss to the instance loss.
|
893 |
+
loss = loss + args.prior_loss_weight * prior_loss
|
894 |
+
else:
|
895 |
+
loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
|
896 |
+
|
897 |
+
accelerator.backward(loss)
|
898 |
+
if accelerator.sync_gradients:
|
899 |
+
params_to_clip = lora_layers.parameters()
|
900 |
+
accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
|
901 |
+
optimizer.step()
|
902 |
+
lr_scheduler.step()
|
903 |
+
optimizer.zero_grad()
|
904 |
+
|
905 |
+
# Checks if the accelerator has performed an optimization step behind the scenes
|
906 |
+
if accelerator.sync_gradients:
|
907 |
+
progress_bar.update(1)
|
908 |
+
global_step += 1
|
909 |
+
|
910 |
+
if global_step % args.checkpointing_steps == 0:
|
911 |
+
if accelerator.is_main_process:
|
912 |
+
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
|
913 |
+
accelerator.save_state(save_path)
|
914 |
+
logger.info(f"Saved state to {save_path}")
|
915 |
+
|
916 |
+
logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
|
917 |
+
progress_bar.set_postfix(**logs)
|
918 |
+
accelerator.log(logs, step=global_step)
|
919 |
+
|
920 |
+
if global_step >= args.max_train_steps:
|
921 |
+
break
|
922 |
+
|
923 |
+
if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
|
924 |
+
logger.info(
|
925 |
+
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
|
926 |
+
f" {args.validation_prompt}."
|
927 |
+
)
|
928 |
+
# create pipeline
|
929 |
+
pipeline = DiffusionPipeline.from_pretrained(
|
930 |
+
args.pretrained_model_name_or_path,
|
931 |
+
unet=accelerator.unwrap_model(unet),
|
932 |
+
text_encoder=accelerator.unwrap_model(text_encoder),
|
933 |
+
revision=args.revision,
|
934 |
+
torch_dtype=weight_dtype,
|
935 |
+
)
|
936 |
+
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
|
937 |
+
pipeline = pipeline.to(accelerator.device)
|
938 |
+
pipeline.set_progress_bar_config(disable=True)
|
939 |
+
|
940 |
+
# run inference
|
941 |
+
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)
|
942 |
+
prompt = args.num_validation_images * [args.validation_prompt]
|
943 |
+
images = pipeline(prompt, num_inference_steps=25, generator=generator).images
|
944 |
+
|
945 |
+
for tracker in accelerator.trackers:
|
946 |
+
if tracker.name == "wandb":
|
947 |
+
tracker.log(
|
948 |
+
{
|
949 |
+
"validation": [
|
950 |
+
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
|
951 |
+
for i, image in enumerate(images)
|
952 |
+
]
|
953 |
+
}
|
954 |
+
)
|
955 |
+
|
956 |
+
del pipeline
|
957 |
+
torch.cuda.empty_cache()
|
958 |
+
|
959 |
+
# Save the lora layers
|
960 |
+
accelerator.wait_for_everyone()
|
961 |
+
if accelerator.is_main_process:
|
962 |
+
unet = unet.to(torch.float32)
|
963 |
+
unet.save_attn_procs(args.output_dir)
|
964 |
+
|
965 |
+
# Final inference
|
966 |
+
# Load previous pipeline
|
967 |
+
pipeline = DiffusionPipeline.from_pretrained(
|
968 |
+
args.pretrained_model_name_or_path, revision=args.revision, torch_dtype=weight_dtype
|
969 |
+
)
|
970 |
+
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
|
971 |
+
pipeline = pipeline.to(accelerator.device)
|
972 |
+
|
973 |
+
# load attention processors
|
974 |
+
pipeline.unet.load_attn_procs(args.output_dir)
|
975 |
+
|
976 |
+
# run inference
|
977 |
+
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)
|
978 |
+
prompt = args.num_validation_images * [args.validation_prompt]
|
979 |
+
images = pipeline(prompt, num_inference_steps=25, generator=generator).images
|
980 |
+
|
981 |
+
for tracker in accelerator.trackers:
|
982 |
+
if tracker.name == "wandb":
|
983 |
+
tracker.log(
|
984 |
+
{
|
985 |
+
"test": [
|
986 |
+
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
|
987 |
+
for i, image in enumerate(images)
|
988 |
+
]
|
989 |
+
}
|
990 |
+
)
|
991 |
+
|
992 |
+
if args.push_to_hub:
|
993 |
+
save_model_card(
|
994 |
+
repo_name,
|
995 |
+
base_model=args.pretrained_model_name_or_path,
|
996 |
+
instance_prompt=args.instance_prompt,
|
997 |
+
test_prompt=args.validation_prompt,
|
998 |
+
images=images,
|
999 |
+
repo_folder=args.output_dir,
|
1000 |
+
)
|
1001 |
+
repo.push_to_hub(commit_message="End of training", blocking=False, auto_lfs_prune=True)
|
1002 |
+
else:
|
1003 |
+
repo_name = Path(args.output_dir).name
|
1004 |
+
save_model_card(
|
1005 |
+
repo_name,
|
1006 |
+
base_model=args.pretrained_model_name_or_path,
|
1007 |
+
instance_prompt=args.instance_prompt,
|
1008 |
+
test_prompt=args.validation_prompt,
|
1009 |
+
images=images,
|
1010 |
+
repo_folder=args.output_dir,
|
1011 |
+
)
|
1012 |
+
|
1013 |
+
accelerator.end_training()
|
1014 |
+
|
1015 |
+
|
1016 |
+
if __name__ == "__main__":
|
1017 |
+
args = parse_args()
|
1018 |
+
main(args)
|
trainer.py
CHANGED
@@ -1,5 +1,6 @@
|
|
1 |
from __future__ import annotations
|
2 |
|
|
|
3 |
import os
|
4 |
import pathlib
|
5 |
import shlex
|
@@ -8,9 +9,10 @@ import subprocess
|
|
8 |
|
9 |
import gradio as gr
|
10 |
import PIL.Image
|
|
|
11 |
import torch
|
12 |
|
13 |
-
|
14 |
|
15 |
|
16 |
def pad_image(image: PIL.Image.Image) -> PIL.Image.Image:
|
@@ -28,94 +30,105 @@ def pad_image(image: PIL.Image.Image) -> PIL.Image.Image:
|
|
28 |
|
29 |
|
30 |
class Trainer:
|
31 |
-
def
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
self.instance_data_dir = self.output_dir / 'training_data'
|
37 |
-
|
38 |
-
def check_if_running(self) -> dict:
|
39 |
-
if self.is_running:
|
40 |
-
return gr.update(value=self.is_running_message)
|
41 |
-
else:
|
42 |
-
return gr.update(value='No training is running.')
|
43 |
-
|
44 |
-
def cleanup_dirs(self) -> None:
|
45 |
-
shutil.rmtree(self.output_dir, ignore_errors=True)
|
46 |
-
|
47 |
-
def prepare_dataset(self, concept_images: list, resolution: int) -> None:
|
48 |
-
self.instance_data_dir.mkdir(parents=True)
|
49 |
-
for i, temp_path in enumerate(concept_images):
|
50 |
image = PIL.Image.open(temp_path.name)
|
51 |
image = pad_image(image)
|
52 |
image = image.resize((resolution, resolution))
|
53 |
image = image.convert('RGB')
|
54 |
-
out_path =
|
55 |
image.save(out_path, format='JPEG', quality=100)
|
56 |
|
57 |
def run(
|
58 |
self,
|
|
|
|
|
|
|
|
|
|
|
59 |
base_model: str,
|
60 |
resolution_s: str,
|
61 |
-
concept_images: list | None,
|
62 |
-
concept_prompt: str,
|
63 |
n_steps: int,
|
64 |
learning_rate: float,
|
65 |
-
train_text_encoder: bool,
|
66 |
-
learning_rate_text: float,
|
67 |
gradient_accumulation: int,
|
|
|
68 |
fp16: bool,
|
69 |
use_8bit_adam: bool,
|
70 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
if not torch.cuda.is_available():
|
72 |
raise gr.Error('CUDA is not available.')
|
73 |
-
|
74 |
-
if self.is_running:
|
75 |
-
return gr.update(value=self.is_running_message), []
|
76 |
-
|
77 |
-
if concept_images is None:
|
78 |
raise gr.Error('You need to upload images.')
|
79 |
-
if not
|
80 |
-
raise gr.Error('The
|
|
|
|
|
81 |
|
82 |
resolution = int(resolution_s)
|
83 |
|
84 |
-
|
85 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
|
87 |
command = f'''
|
88 |
-
accelerate launch
|
89 |
--pretrained_model_name_or_path={base_model} \
|
90 |
-
--instance_data_dir={
|
91 |
-
--output_dir={
|
92 |
-
--instance_prompt="{
|
93 |
--resolution={resolution} \
|
94 |
--train_batch_size=1 \
|
95 |
--gradient_accumulation_steps={gradient_accumulation} \
|
96 |
--learning_rate={learning_rate} \
|
97 |
--lr_scheduler=constant \
|
98 |
--lr_warmup_steps=0 \
|
99 |
-
--max_train_steps={n_steps}
|
|
|
|
|
|
|
|
|
100 |
'''
|
101 |
if fp16:
|
102 |
command += ' --mixed_precision fp16'
|
103 |
if use_8bit_adam:
|
104 |
command += ' --use_8bit_adam'
|
105 |
-
if
|
106 |
-
command +=
|
107 |
-
|
108 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
109 |
command_s = ' '.join(command.split())
|
110 |
f.write(command_s)
|
111 |
|
112 |
-
|
113 |
-
res = subprocess.run(shlex.split(command))
|
114 |
-
self.is_running = False
|
115 |
-
|
116 |
-
if res.returncode == 0:
|
117 |
-
result_message = 'Training Completed!'
|
118 |
-
else:
|
119 |
-
result_message = 'Training Failed!'
|
120 |
-
weight_paths = sorted(self.output_dir.glob('*.pt'))
|
121 |
-
return gr.update(value=result_message), weight_paths
|
|
|
1 |
from __future__ import annotations
|
2 |
|
3 |
+
import datetime
|
4 |
import os
|
5 |
import pathlib
|
6 |
import shlex
|
|
|
9 |
|
10 |
import gradio as gr
|
11 |
import PIL.Image
|
12 |
+
import slugify
|
13 |
import torch
|
14 |
|
15 |
+
from constants import UploadTarget
|
16 |
|
17 |
|
18 |
def pad_image(image: PIL.Image.Image) -> PIL.Image.Image:
|
|
|
30 |
|
31 |
|
32 |
class Trainer:
|
33 |
+
def prepare_dataset(self, instance_images: list, resolution: int,
|
34 |
+
instance_data_dir: pathlib.Path) -> None:
|
35 |
+
shutil.rmtree(instance_data_dir, ignore_errors=True)
|
36 |
+
instance_data_dir.mkdir(parents=True)
|
37 |
+
for i, temp_path in enumerate(instance_images):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
image = PIL.Image.open(temp_path.name)
|
39 |
image = pad_image(image)
|
40 |
image = image.resize((resolution, resolution))
|
41 |
image = image.convert('RGB')
|
42 |
+
out_path = instance_data_dir / f'{i:03d}.jpg'
|
43 |
image.save(out_path, format='JPEG', quality=100)
|
44 |
|
45 |
def run(
|
46 |
self,
|
47 |
+
instance_images: list | None,
|
48 |
+
instance_prompt: str,
|
49 |
+
output_model_name: str,
|
50 |
+
overwrite_existing_model: bool,
|
51 |
+
validation_prompt: str,
|
52 |
base_model: str,
|
53 |
resolution_s: str,
|
|
|
|
|
54 |
n_steps: int,
|
55 |
learning_rate: float,
|
|
|
|
|
56 |
gradient_accumulation: int,
|
57 |
+
seed: int,
|
58 |
fp16: bool,
|
59 |
use_8bit_adam: bool,
|
60 |
+
checkpointing_steps: int,
|
61 |
+
use_wandb: bool,
|
62 |
+
validation_epochs: int,
|
63 |
+
upload_to_hub: bool,
|
64 |
+
use_private_repo: bool,
|
65 |
+
delete_existing_repo: bool,
|
66 |
+
upload_to: str,
|
67 |
+
) -> str:
|
68 |
if not torch.cuda.is_available():
|
69 |
raise gr.Error('CUDA is not available.')
|
70 |
+
if instance_images is None:
|
|
|
|
|
|
|
|
|
71 |
raise gr.Error('You need to upload images.')
|
72 |
+
if not instance_prompt:
|
73 |
+
raise gr.Error('The instance prompt is missing.')
|
74 |
+
if not validation_prompt:
|
75 |
+
raise gr.Error('The validation prompt is missing.')
|
76 |
|
77 |
resolution = int(resolution_s)
|
78 |
|
79 |
+
if not output_model_name:
|
80 |
+
output_model_name = datetime.datetime.now().strftime(
|
81 |
+
'%Y-%m-%d-%H-%M-%S')
|
82 |
+
output_model_name = slugify.slugify(output_model_name)
|
83 |
+
|
84 |
+
repo_dir = pathlib.Path(__file__).parent
|
85 |
+
output_dir = repo_dir / 'experiments' / output_model_name
|
86 |
+
if overwrite_existing_model or upload_to_hub:
|
87 |
+
shutil.rmtree(output_dir, ignore_errors=True)
|
88 |
+
if not upload_to_hub:
|
89 |
+
output_dir.mkdir(parents=True)
|
90 |
+
|
91 |
+
instance_data_dir = repo_dir / 'training_data' / output_model_name
|
92 |
+
self.prepare_dataset(instance_images, resolution, instance_data_dir)
|
93 |
|
94 |
command = f'''
|
95 |
+
accelerate launch train_dreambooth_lora.py \
|
96 |
--pretrained_model_name_or_path={base_model} \
|
97 |
+
--instance_data_dir={instance_data_dir} \
|
98 |
+
--output_dir={output_dir} \
|
99 |
+
--instance_prompt="{instance_prompt}" \
|
100 |
--resolution={resolution} \
|
101 |
--train_batch_size=1 \
|
102 |
--gradient_accumulation_steps={gradient_accumulation} \
|
103 |
--learning_rate={learning_rate} \
|
104 |
--lr_scheduler=constant \
|
105 |
--lr_warmup_steps=0 \
|
106 |
+
--max_train_steps={n_steps} \
|
107 |
+
--checkpointing_steps={checkpointing_steps} \
|
108 |
+
--validation_prompt="{validation_prompt}" \
|
109 |
+
--validation_epochs={validation_epochs} \
|
110 |
+
--seed={seed}
|
111 |
'''
|
112 |
if fp16:
|
113 |
command += ' --mixed_precision fp16'
|
114 |
if use_8bit_adam:
|
115 |
command += ' --use_8bit_adam'
|
116 |
+
if use_wandb:
|
117 |
+
command += ' --report_to wandb'
|
118 |
+
if upload_to_hub:
|
119 |
+
hf_token = os.getenv('HF_TOKEN')
|
120 |
+
command += f' --push_to_hub --hub_token {hf_token}'
|
121 |
+
if use_private_repo:
|
122 |
+
command += ' --private_repo'
|
123 |
+
if delete_existing_repo:
|
124 |
+
command += ' --delete_existing_repo'
|
125 |
+
if upload_to == UploadTarget.LORA_LIBRARY.value:
|
126 |
+
command += ' --upload_to_lora_library'
|
127 |
+
|
128 |
+
subprocess.run(shlex.split(command))
|
129 |
+
|
130 |
+
with open(output_dir / 'train.sh', 'w') as f:
|
131 |
command_s = ' '.join(command.split())
|
132 |
f.write(command_s)
|
133 |
|
134 |
+
return 'Training completed!'
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
uploader.py
CHANGED
@@ -1,20 +1,39 @@
|
|
1 |
-
|
|
|
2 |
from huggingface_hub import HfApi
|
3 |
|
4 |
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
api.
|
11 |
-
api.upload_folder(repo_id=model_id,
|
12 |
-
folder_path='results',
|
13 |
-
path_in_repo='results',
|
14 |
-
repo_type='model')
|
15 |
-
url = f'https://huggingface.co/{model_id}'
|
16 |
-
message = f'Your model was successfully uploaded to [{url}]({url}).'
|
17 |
-
except Exception as e:
|
18 |
-
message = str(e)
|
19 |
|
20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import annotations
|
2 |
+
|
3 |
from huggingface_hub import HfApi
|
4 |
|
5 |
|
6 |
+
class Uploader:
|
7 |
+
def __init__(self, hf_token: str | None):
|
8 |
+
self.api = HfApi(token=hf_token)
|
9 |
+
|
10 |
+
def get_username(self) -> str:
|
11 |
+
return self.api.whoami()['name']
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
+
def upload(self,
|
14 |
+
folder_path: str,
|
15 |
+
repo_name: str,
|
16 |
+
organization: str = '',
|
17 |
+
repo_type: str = 'model',
|
18 |
+
private: bool = True,
|
19 |
+
delete_existing_repo: bool = False) -> str:
|
20 |
+
if not organization:
|
21 |
+
organization = self.get_username()
|
22 |
+
repo_id = f'{organization}/{repo_name}'
|
23 |
+
if delete_existing_repo:
|
24 |
+
try:
|
25 |
+
self.api.delete_repo(repo_id, repo_type=repo_type)
|
26 |
+
except Exception:
|
27 |
+
pass
|
28 |
+
try:
|
29 |
+
self.api.create_repo(repo_id, repo_type=repo_type, private=private)
|
30 |
+
self.api.upload_folder(repo_id=repo_id,
|
31 |
+
folder_path=folder_path,
|
32 |
+
path_in_repo='.',
|
33 |
+
repo_type=repo_type)
|
34 |
+
url = f'https://huggingface.co/{repo_id}'
|
35 |
+
message = f'Your model was successfully uploaded to <a href="{url}" target="_blank">{url}</a>.'
|
36 |
+
except Exception as e:
|
37 |
+
message = str(e)
|
38 |
+
print(message)
|
39 |
+
return message
|
utils.py
ADDED
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pathlib
|
2 |
+
|
3 |
+
|
4 |
+
def find_exp_dirs(ignore_repo: bool = False) -> list[str]:
|
5 |
+
repo_dir = pathlib.Path(__file__).parent
|
6 |
+
exp_root_dir = repo_dir / 'experiments'
|
7 |
+
if not exp_root_dir.exists():
|
8 |
+
return []
|
9 |
+
exp_dirs = sorted(exp_root_dir.glob('*'))
|
10 |
+
exp_dirs = [
|
11 |
+
exp_dir for exp_dir in exp_dirs
|
12 |
+
if (exp_dir / 'pytorch_lora_weights.bin').exists()
|
13 |
+
]
|
14 |
+
if ignore_repo:
|
15 |
+
exp_dirs = [
|
16 |
+
exp_dir for exp_dir in exp_dirs if not (exp_dir / '.git').exists()
|
17 |
+
]
|
18 |
+
return [path.relative_to(repo_dir).as_posix() for path in exp_dirs]
|