Raman Dutt
commited on
Commit
·
ee6eca1
1
Parent(s):
740ee27
app.py added
Browse files
app.py
ADDED
@@ -0,0 +1,329 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import PIL.Image
|
3 |
+
from pathlib import Path
|
4 |
+
import pandas as pd
|
5 |
+
from diffusers.pipelines import StableDiffusionPipeline
|
6 |
+
import torch
|
7 |
+
import argparse
|
8 |
+
import os
|
9 |
+
import warnings
|
10 |
+
from safetensors.torch import load_file
|
11 |
+
import yaml
|
12 |
+
|
13 |
+
warnings.filterwarnings("ignore")
|
14 |
+
|
15 |
+
OUTPUT_DIR = "OUTPUT"
|
16 |
+
cuda_device = 1
|
17 |
+
device = f"cuda:{cuda_device}" if torch.cuda.is_available() else "cpu"
|
18 |
+
|
19 |
+
TITLE = "Demo for Generating Chest X-rays using Diferent Parameter-Efficient Fine-Tuned Stable Diffusion Pipelines"
|
20 |
+
INFO_ABOUT_TEXT_PROMPT = "INFO_ABOUT_TEXT_PROMPT"
|
21 |
+
INFO_ABOUT_GUIDANCE_SCALE = "INFO_ABOUT_GUIDANCE_SCALE"
|
22 |
+
INFO_ABOUT_INFERENCE_STEPS = "INFO_ABOUT_INFERENCE_STEPS"
|
23 |
+
EXAMPLE_TEXT_PROMPTS = [
|
24 |
+
"No acute cardiopulmonary abnormality.",
|
25 |
+
"Normal chest radiograph.",
|
26 |
+
"No acute intrathoracic process.",
|
27 |
+
"Mild pulmonary edema.",
|
28 |
+
"No focal consolidation concerning for pneumonia",
|
29 |
+
"No radiographic evidence for acute cardiopulmonary process",
|
30 |
+
]
|
31 |
+
|
32 |
+
|
33 |
+
def load_adapted_unet(unet_pretraining_type, exp_path, pipe):
|
34 |
+
|
35 |
+
"""
|
36 |
+
Loads the adapted U-Net for the selected PEFT Type
|
37 |
+
|
38 |
+
Parameters:
|
39 |
+
unet_pretraining_type (str): The type of PEFT to use for generating the X-ray
|
40 |
+
exp_path (str): The path to the best trained model for the selected PEFT Type
|
41 |
+
pipe (StableDiffusionPipeline): The Stable Diffusion Pipeline to use for generating the X-ray
|
42 |
+
|
43 |
+
Returns:
|
44 |
+
None
|
45 |
+
"""
|
46 |
+
|
47 |
+
sd_folder_path = "runwayml/stable-diffusion-v1-5"
|
48 |
+
|
49 |
+
if unet_pretraining_type == "freeze":
|
50 |
+
pass
|
51 |
+
|
52 |
+
elif unet_pretraining_type == "svdiff":
|
53 |
+
print("SV-DIFF UNET")
|
54 |
+
|
55 |
+
pipe.unet = load_unet_for_svdiff(
|
56 |
+
sd_folder_path,
|
57 |
+
spectral_shifts_ckpt=os.path.join(
|
58 |
+
os.path.join(exp_path, "unet"), "spectral_shifts.safetensors"
|
59 |
+
),
|
60 |
+
subfolder="unet",
|
61 |
+
)
|
62 |
+
for module in pipe.unet.modules():
|
63 |
+
if hasattr(module, "perform_svd"):
|
64 |
+
module.perform_svd()
|
65 |
+
|
66 |
+
elif unet_pretraining_type == "lorav2":
|
67 |
+
exp_path = os.path.join(exp_path, "pytorch_lora_weights.safetensors")
|
68 |
+
pipe.unet.load_attn_procs(exp_path)
|
69 |
+
else:
|
70 |
+
exp_path = unet_pretraining_type + "_" + "diffusion_pytorch_model.safetensors"
|
71 |
+
state_dict = load_file(exp_path)
|
72 |
+
print(pipe.unet.load_state_dict(state_dict, strict=False))
|
73 |
+
|
74 |
+
|
75 |
+
def loadSDModel(unet_pretraining_type, exp_path, cuda_device):
|
76 |
+
|
77 |
+
"""
|
78 |
+
Loads the Stable Diffusion Model for the selected PEFT Type
|
79 |
+
|
80 |
+
Parameters:
|
81 |
+
unet_pretraining_type (str): The type of PEFT to use for generating the X-ray
|
82 |
+
exp_path (str): The path to the best trained model for the selected PEFT Type
|
83 |
+
cuda_device (str): The CUDA device to use for generating the X-ray
|
84 |
+
|
85 |
+
Returns:
|
86 |
+
pipe (StableDiffusionPipeline): The Stable Diffusion Pipeline to use for generating the X-ray
|
87 |
+
"""
|
88 |
+
|
89 |
+
sd_folder_path = "runwayml/stable-diffusion-v1-5"
|
90 |
+
|
91 |
+
pipe = StableDiffusionPipeline.from_pretrained(sd_folder_path, revision="fp16")
|
92 |
+
|
93 |
+
load_adapted_unet(unet_pretraining_type, exp_path, pipe)
|
94 |
+
pipe.safety_checker = None
|
95 |
+
|
96 |
+
return pipe
|
97 |
+
|
98 |
+
|
99 |
+
def load_all_pipelines():
|
100 |
+
|
101 |
+
"""
|
102 |
+
Loads all the Stable Diffusion Pipelines for each PEFT Type for efficient caching (Design Choice 2)
|
103 |
+
|
104 |
+
Parameters:
|
105 |
+
None
|
106 |
+
|
107 |
+
Returns:
|
108 |
+
sd_pipeline_full (StableDiffusionPipeline): The Stable Diffusion Pipeline for Full Fine-Tuning
|
109 |
+
sd_pipeline_norm (StableDiffusionPipeline): The Stable Diffusion Pipeline for Norm Fine-Tuning
|
110 |
+
sd_pipeline_bias (StableDiffusionPipeline): The Stable Diffusion Pipeline for Bias Fine-Tuning
|
111 |
+
sd_pipeline_attention (StableDiffusionPipeline): The Stable Diffusion Pipeline for Attention Fine-Tuning
|
112 |
+
sd_pipeline_NBA (StableDiffusionPipeline): The Stable Diffusion Pipeline for NBA Fine-Tuning
|
113 |
+
sd_pipeline_difffit (StableDiffusionPipeline): The Stable Diffusion Pipeline for Difffit Fine-Tuning
|
114 |
+
"""
|
115 |
+
|
116 |
+
# Dictionary containing the path to the best trained models for each PEFT type
|
117 |
+
MODEL_PATH_DICT = {
|
118 |
+
"full": "full_diffusion_pytorch_model.safetensors",
|
119 |
+
"norm": "norm_diffusion_pytorch_model.safetensors",
|
120 |
+
"bias": "bias_diffusion_pytorch_model.safetensors",
|
121 |
+
"attention": "attention_diffusion_pytorch_model.safetensors",
|
122 |
+
"norm_bias_attention": "norm_bias_attention_diffusion_pytorch_model.safetensors",
|
123 |
+
"difffit": "difffit_diffusion_pytorch_model.safetensors",
|
124 |
+
}
|
125 |
+
|
126 |
+
device = "0"
|
127 |
+
cuda_device = f"cuda:{device}" if torch.cuda.is_available() else "cpu"
|
128 |
+
|
129 |
+
# Full FT
|
130 |
+
unet_pretraining_type = "full"
|
131 |
+
print("Loading Pipeline for Full Fine-Tuning")
|
132 |
+
sd_pipeline_full = loadSDModel(
|
133 |
+
unet_pretraining_type=unet_pretraining_type,
|
134 |
+
exp_path=MODEL_PATH_DICT[unet_pretraining_type],
|
135 |
+
cuda_device=cuda_device,
|
136 |
+
)
|
137 |
+
|
138 |
+
# Norm
|
139 |
+
unet_pretraining_type = "norm"
|
140 |
+
print("Loading Pipeline for Norm Fine-Tuning")
|
141 |
+
sd_pipeline_norm = loadSDModel(
|
142 |
+
unet_pretraining_type=unet_pretraining_type,
|
143 |
+
exp_path=MODEL_PATH_DICT[unet_pretraining_type],
|
144 |
+
cuda_device=cuda_device,
|
145 |
+
)
|
146 |
+
|
147 |
+
# bias
|
148 |
+
unet_pretraining_type = "bias"
|
149 |
+
print("Loading Pipeline for Bias Fine-Tuning")
|
150 |
+
sd_pipeline_bias = loadSDModel(
|
151 |
+
unet_pretraining_type=unet_pretraining_type,
|
152 |
+
exp_path=MODEL_PATH_DICT[unet_pretraining_type],
|
153 |
+
cuda_device=cuda_device,
|
154 |
+
)
|
155 |
+
|
156 |
+
# attention
|
157 |
+
unet_pretraining_type = "attention"
|
158 |
+
print("Loading Pipeline for Attention Fine-Tuning")
|
159 |
+
sd_pipeline_attention = loadSDModel(
|
160 |
+
unet_pretraining_type=unet_pretraining_type,
|
161 |
+
exp_path=MODEL_PATH_DICT[unet_pretraining_type],
|
162 |
+
cuda_device=cuda_device,
|
163 |
+
)
|
164 |
+
|
165 |
+
# NBA
|
166 |
+
unet_pretraining_type = "norm_bias_attention"
|
167 |
+
print("Loading Pipeline for NBA Fine-Tuning")
|
168 |
+
sd_pipeline_NBA = loadSDModel(
|
169 |
+
unet_pretraining_type=unet_pretraining_type,
|
170 |
+
exp_path=MODEL_PATH_DICT[unet_pretraining_type],
|
171 |
+
cuda_device=cuda_device,
|
172 |
+
)
|
173 |
+
|
174 |
+
# difffit
|
175 |
+
unet_pretraining_type = "difffit"
|
176 |
+
print("Loading Pipeline for Difffit Fine-Tuning")
|
177 |
+
sd_pipeline_difffit = loadSDModel(
|
178 |
+
unet_pretraining_type=unet_pretraining_type,
|
179 |
+
exp_path=MODEL_PATH_DICT[unet_pretraining_type],
|
180 |
+
cuda_device=cuda_device,
|
181 |
+
)
|
182 |
+
|
183 |
+
return (
|
184 |
+
sd_pipeline_full,
|
185 |
+
sd_pipeline_norm,
|
186 |
+
sd_pipeline_bias,
|
187 |
+
sd_pipeline_attention,
|
188 |
+
sd_pipeline_NBA,
|
189 |
+
sd_pipeline_difffit,
|
190 |
+
)
|
191 |
+
|
192 |
+
|
193 |
+
# LOAD ALL PIPELINES FIRST AND CACHE THEM
|
194 |
+
# (
|
195 |
+
# sd_pipeline_full,
|
196 |
+
# sd_pipeline_norm,
|
197 |
+
# sd_pipeline_bias,
|
198 |
+
# sd_pipeline_attention,
|
199 |
+
# sd_pipeline_NBA,
|
200 |
+
# sd_pipeline_difffit,
|
201 |
+
# ) = load_all_pipelines()
|
202 |
+
|
203 |
+
# PIPELINE_DICT = {
|
204 |
+
# "full": sd_pipeline_full,
|
205 |
+
# "norm": sd_pipeline_norm,
|
206 |
+
# "bias": sd_pipeline_bias,
|
207 |
+
# "attention": sd_pipeline_attention,
|
208 |
+
# "norm_bias_attention": sd_pipeline_NBA,
|
209 |
+
# "difffit": sd_pipeline_difffit,
|
210 |
+
# }
|
211 |
+
|
212 |
+
|
213 |
+
def predict(
|
214 |
+
unet_pretraining_type,
|
215 |
+
input_text,
|
216 |
+
guidance_scale=4,
|
217 |
+
num_inference_steps=75,
|
218 |
+
device="0",
|
219 |
+
OUTPUT_DIR="OUTPUT",
|
220 |
+
PIPELINE_DICT=PIPELINE_DICT,
|
221 |
+
):
|
222 |
+
|
223 |
+
NUM_TUNABLE_PARAMS = {
|
224 |
+
"full": 86,
|
225 |
+
"attention": 26.7,
|
226 |
+
"bias": 0.343,
|
227 |
+
"norm": 0.2,
|
228 |
+
"norm_bias_attention": 26.7,
|
229 |
+
"lorav2": 0.8,
|
230 |
+
"svdiff": 0.222,
|
231 |
+
"difffit": 0.581,
|
232 |
+
}
|
233 |
+
|
234 |
+
cuda_device = f"cuda:{device}" if torch.cuda.is_available() else "cpu"
|
235 |
+
|
236 |
+
|
237 |
+
#sd_pipeline = PIPELINE_DICT[unet_pretraining_type]
|
238 |
+
print("Loading Pipeline for {} Fine-Tuning".format(unet_pretraining_type))
|
239 |
+
sd_pipeline_norm = loadSDModel(
|
240 |
+
unet_pretraining_type=unet_pretraining_type,
|
241 |
+
exp_path=MODEL_PATH_DICT[unet_pretraining_type],
|
242 |
+
cuda_device=cuda_device,
|
243 |
+
)
|
244 |
+
|
245 |
+
sd_pipeline.to(cuda_device)
|
246 |
+
|
247 |
+
result_image = sd_pipeline(
|
248 |
+
prompt=input_text,
|
249 |
+
height=224,
|
250 |
+
width=224,
|
251 |
+
guidance_scale=guidance_scale,
|
252 |
+
num_inference_steps=num_inference_steps,
|
253 |
+
)
|
254 |
+
|
255 |
+
result_pil_image = result_image["images"][0]
|
256 |
+
|
257 |
+
# Create a Bar Plot displaying the number of tunable parameters for the selected PEFT Type
|
258 |
+
# Create a Pandas DataFrame
|
259 |
+
|
260 |
+
df = pd.DataFrame(
|
261 |
+
{
|
262 |
+
"PEFT Type": list(NUM_TUNABLE_PARAMS.keys()),
|
263 |
+
"Number of Tunable Parameters": list(NUM_TUNABLE_PARAMS.values()),
|
264 |
+
}
|
265 |
+
)
|
266 |
+
|
267 |
+
df = df[df["PEFT Type"].isin(["full", unet_pretraining_type])].reset_index(
|
268 |
+
drop=True
|
269 |
+
)
|
270 |
+
|
271 |
+
bar_plot = gr.BarPlot(
|
272 |
+
value=df,
|
273 |
+
x="PEFT Type",
|
274 |
+
y="Number of Tunable Parameters",
|
275 |
+
label="PEFT Type",
|
276 |
+
title="Number of Tunable Parameters",
|
277 |
+
vertical=False,
|
278 |
+
)
|
279 |
+
|
280 |
+
return result_pil_image, bar_plot
|
281 |
+
|
282 |
+
|
283 |
+
# Create a Gradio interface
|
284 |
+
"""
|
285 |
+
Input Parameters:
|
286 |
+
1. PEFT Type: (Dropdown) The type of PEFT to use for generating the X-ray
|
287 |
+
2. Input Text: (Textbox) The text prompt to use for generating the X-ray
|
288 |
+
3. Guidance Scale: (Slider) The guidance scale to use for generating the X-ray
|
289 |
+
4. Num Inference Steps: (Slider) The number of inference steps to use for generating the X-ray
|
290 |
+
|
291 |
+
Output Parameters:
|
292 |
+
1. Generated X-ray Image: (Image) The generated X-ray image
|
293 |
+
2. Number of Tunable Parameters: (Bar Plot) The number of tunable parameters for the selected PEFT Type
|
294 |
+
"""
|
295 |
+
iface = gr.Interface(
|
296 |
+
fn=predict,
|
297 |
+
inputs=[
|
298 |
+
gr.Dropdown(
|
299 |
+
["full", "difffit", "svdiff", "norm", "bias", "attention"],
|
300 |
+
label="PEFT Type",
|
301 |
+
),
|
302 |
+
gr.Dropdown(
|
303 |
+
EXAMPLE_TEXT_PROMPTS, info=INFO_ABOUT_TEXT_PROMPT, label="Input Text"
|
304 |
+
),
|
305 |
+
gr.Slider(
|
306 |
+
minimum=1,
|
307 |
+
maximum=10,
|
308 |
+
value=4,
|
309 |
+
step=1,
|
310 |
+
info=INFO_ABOUT_GUIDANCE_SCALE,
|
311 |
+
label="Guidance Scale",
|
312 |
+
),
|
313 |
+
gr.Slider(
|
314 |
+
minimum=1,
|
315 |
+
maximum=100,
|
316 |
+
value=75,
|
317 |
+
step=1,
|
318 |
+
info=INFO_ABOUT_INFERENCE_STEPS,
|
319 |
+
label="Num Inference Steps",
|
320 |
+
),
|
321 |
+
],
|
322 |
+
outputs=[gr.Image(type="pil"), gr.BarPlot()],
|
323 |
+
live=True,
|
324 |
+
analytics_enabled=False,
|
325 |
+
title=TITLE,
|
326 |
+
)
|
327 |
+
|
328 |
+
# Launch the Gradio interface
|
329 |
+
iface.launch(share=True)
|