Spaces:
Running
on
Zero
Running
on
Zero
simplify UI
Browse files- .gitattributes +2 -0
- README.md +47 -1
- app.py +44 -95
- samples/n02086240_2799.JPEG +3 -0
- samples/n03417042_5234.JPEG +3 -0
- samples/unsafe.jpeg +3 -0
- tools/synth.py +0 -1
.gitattributes
CHANGED
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
*.JPEG filter=lfs diff=lfs merge=lfs -text
|
37 |
+
*.jpeg filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
@@ -9,4 +9,50 @@ app_file: app.py
|
|
9 |
pinned: false
|
10 |
---
|
11 |
|
12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
pinned: false
|
10 |
---
|
11 |
|
12 |
+
# Generative Data Augmentation Demo
|
13 |
+
|
14 |
+
Main GitHub Repo: [Generative Data Augmentation](https://github.com/zhulinchng/generative-data-augmentation) | Image Classification Demo: [Generative Augmented Classifiers](https://huggingface.co/spaces/czl/generative-augmented-classifiers).
|
15 |
+
|
16 |
+
This demo is created as part of the 'Investigating the Effectiveness of Generative Diffusion Models in Synthesizing Images for Data Augmentation in Image Classification' dissertation.
|
17 |
+
|
18 |
+
The user can augment an image by interpolating between two prompts, and specify the number of interpolation steps and the specific step to generate the image.
|
19 |
+
|
20 |
+
## Demo Usage Instructions
|
21 |
+
|
22 |
+
1. Upload an image.
|
23 |
+
2. Enter the two prompts to interpolate between, the first prompt should contain the desired class of the augmented image, the second prompt should contain the undesired class (i.e., confusing class).
|
24 |
+
|
25 |
+
## Configuration
|
26 |
+
|
27 |
+
- Total Interpolation Steps: The number of steps to interpolate between the two prompts.
|
28 |
+
- Interpolation Step: The specific step to generate the image.
|
29 |
+
- Example for 10 steps:
|
30 |
+
|
31 |
+
```python
|
32 |
+
Total: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
|
33 |
+
Sampled: 4
|
34 |
+
```
|
35 |
+
|
36 |
+
- Seed: Seed value for reproducibility.
|
37 |
+
- Negative Prompt: Prompt to guide the model away from generating the image.
|
38 |
+
- Width, Height: The dimensions of the generated image.
|
39 |
+
- Guidance Scale: The scale of the guide the model on how closely to follow the prompts.
|
40 |
+
|
41 |
+
## Metadata
|
42 |
+
|
43 |
+
[SSIM Score](https://lightning.ai/docs/torchmetrics/stable/image/structural_similarity.html): Structural Similarity Index (SSIM) score between the original and generated image, ranges from 0 to 1.
|
44 |
+
[CLIP Score](https://lightning.ai/docs/torchmetrics/stable/multimodal/clip_score.html): CLIP similarity score between the original and generated image, ranges from 0 to 100.
|
45 |
+
|
46 |
+
## Local Setup
|
47 |
+
|
48 |
+
```bash
|
49 |
+
git clone https://huggingface.co/spaces/czl/generative-data-augmentation-demo
|
50 |
+
cd generative-data-augmentation-demo
|
51 |
+
# Setup the data directory structure as shown above
|
52 |
+
conda create --name $env_name python=3.11.* # Replace $env_name with your environment name
|
53 |
+
conda activate $env_name
|
54 |
+
# Visit PyTorch website https://pytorch.org/get-started/previous-versions/#v212 for PyTorch installation instructions.
|
55 |
+
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url # Obtain the correct URL from the PyTorch website
|
56 |
+
pip install -r requirements.txt
|
57 |
+
python app.py
|
58 |
+
```
|
app.py
CHANGED
@@ -4,6 +4,7 @@ import gradio as gr
|
|
4 |
import numpy as np
|
5 |
import torch
|
6 |
import torchvision.transforms as transforms
|
|
|
7 |
from torchmetrics.functional.image import structural_similarity_index_measure as ssim
|
8 |
from transformers import CLIPModel, CLIPProcessor
|
9 |
|
@@ -45,8 +46,6 @@ def infer(
|
|
45 |
interpolation_step,
|
46 |
num_inference_steps,
|
47 |
num_interpolation_steps,
|
48 |
-
sample_mid_interpolation,
|
49 |
-
remove_n_middle,
|
50 |
):
|
51 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
52 |
|
@@ -55,38 +54,15 @@ def infer(
|
|
55 |
assert num_interpolation_steps % 2 == 0
|
56 |
except AssertionError:
|
57 |
raise ValueError("num_interpolation_steps must be an even number")
|
58 |
-
try:
|
59 |
-
assert sample_mid_interpolation % 2 == 0
|
60 |
-
except AssertionError:
|
61 |
-
raise ValueError("sample_mid_interpolation must be an even number")
|
62 |
-
try:
|
63 |
-
assert remove_n_middle % 2 == 0
|
64 |
-
except AssertionError:
|
65 |
-
raise ValueError("remove_n_middle must be an even number")
|
66 |
-
try:
|
67 |
-
assert num_interpolation_steps >= sample_mid_interpolation
|
68 |
-
except AssertionError:
|
69 |
-
raise ValueError(
|
70 |
-
"num_interpolation_steps must be greater than or equal to sample_mid_interpolation"
|
71 |
-
)
|
72 |
-
try:
|
73 |
-
assert num_interpolation_steps >= 2 and sample_mid_interpolation >= 2
|
74 |
-
except AssertionError:
|
75 |
-
raise ValueError(
|
76 |
-
"num_interpolation_steps and sample_mid_interpolation must be greater than or equal to 2"
|
77 |
-
)
|
78 |
-
try:
|
79 |
-
assert sample_mid_interpolation - remove_n_middle >= 2
|
80 |
-
except AssertionError:
|
81 |
-
raise ValueError(
|
82 |
-
"sample_mid_interpolation must be greater than or equal to remove_n_middle + 2"
|
83 |
-
)
|
84 |
|
85 |
if randomize_seed:
|
86 |
seed = random.randint(0, MAX_SEED)
|
87 |
prompts = [prompt1, prompt2]
|
88 |
generator = torch.Generator().manual_seed(seed)
|
89 |
|
|
|
|
|
|
|
90 |
interpolated_prompt_embeds, prompt_metadata = synth.interpolatePrompts(
|
91 |
prompts,
|
92 |
pipe,
|
@@ -116,7 +92,6 @@ def infer(
|
|
116 |
).to(device)
|
117 |
embed_pairs = zip(interpolated_prompt_embeds, interpolated_negative_prompts_embeds)
|
118 |
embed_pairs_list = list(embed_pairs)
|
119 |
-
print(len(embed_pairs_list))
|
120 |
# offset step by -1
|
121 |
prompt_embeds, negative_prompt_embeds = embed_pairs_list[interpolation_step - 1]
|
122 |
preprocess_input = transforms.Compose(
|
@@ -127,7 +102,7 @@ def infer(
|
|
127 |
npe = negative_prompt_embeds[None, ...]
|
128 |
else:
|
129 |
npe = None
|
130 |
-
|
131 |
height=height,
|
132 |
width=width,
|
133 |
num_images_per_prompt=1,
|
@@ -138,7 +113,13 @@ def infer(
|
|
138 |
generator=generator,
|
139 |
latents=latents,
|
140 |
image=input_img_tensor,
|
141 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
142 |
pred_image = transforms.ToTensor()(image).unsqueeze(0)
|
143 |
ssim_score = ssim(pred_image, input_img_tensor).item()
|
144 |
real_inputs = clip_processor(
|
@@ -163,25 +144,17 @@ def infer(
|
|
163 |
|
164 |
|
165 |
examples1 = [
|
166 |
-
"A photo of a
|
167 |
"A photo of a Shih-Tzu, a type of dog",
|
168 |
]
|
169 |
examples2 = [
|
170 |
-
"A photo of a
|
171 |
"A photo of a beagle, a type of dog",
|
172 |
]
|
173 |
|
174 |
|
175 |
def update_steps(total_steps, interpolation_step):
|
176 |
-
|
177 |
-
return gr.update(maximum=total_steps // 2, value=total_steps)
|
178 |
-
return gr.update(maximum=total_steps // 2)
|
179 |
-
|
180 |
-
|
181 |
-
def update_sampling_steps(total_steps, sample_steps):
|
182 |
-
# if sample_steps > total_steps:
|
183 |
-
# return gr.update(value=total_steps)
|
184 |
-
return gr.update(value=total_steps)
|
185 |
|
186 |
|
187 |
def update_format(image_format):
|
@@ -211,7 +184,7 @@ with gr.Blocks(title="Generative Date Augmentation Demo") as demo:
|
|
211 |
label="Prompt for the image to synthesize. (Actual class)",
|
212 |
show_label=True,
|
213 |
max_lines=1,
|
214 |
-
placeholder="Enter
|
215 |
container=False,
|
216 |
)
|
217 |
with gr.Row():
|
@@ -219,32 +192,44 @@ with gr.Blocks(title="Generative Date Augmentation Demo") as demo:
|
|
219 |
label="Prompt to augment against. (Confusing class)",
|
220 |
show_label=True,
|
221 |
max_lines=1,
|
222 |
-
placeholder="Enter
|
223 |
container=False,
|
224 |
)
|
225 |
with gr.Row():
|
226 |
gr.Examples(
|
227 |
-
examples=
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
228 |
)
|
229 |
gr.Examples(
|
230 |
-
examples=examples2,
|
|
|
|
|
231 |
)
|
232 |
|
233 |
with gr.Row():
|
234 |
-
interpolation_step = gr.Slider(
|
235 |
-
label="Specific Interpolation Step",
|
236 |
-
minimum=1,
|
237 |
-
maximum=8,
|
238 |
-
step=1,
|
239 |
-
value=8,
|
240 |
-
)
|
241 |
num_interpolation_steps = gr.Slider(
|
242 |
-
label="Total
|
243 |
minimum=2,
|
244 |
-
maximum=
|
245 |
step=2,
|
246 |
value=16,
|
247 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
248 |
num_interpolation_steps.change(
|
249 |
fn=update_steps,
|
250 |
inputs=[num_interpolation_steps, interpolation_step],
|
@@ -305,27 +290,6 @@ with gr.Blocks(title="Generative Date Augmentation Demo") as demo:
|
|
305 |
step=1,
|
306 |
value=25,
|
307 |
)
|
308 |
-
with gr.Row():
|
309 |
-
sample_mid_interpolation = gr.Slider(
|
310 |
-
label="Number of sampling steps in the middle of interpolation",
|
311 |
-
minimum=2,
|
312 |
-
maximum=80,
|
313 |
-
step=2,
|
314 |
-
value=16,
|
315 |
-
)
|
316 |
-
num_interpolation_steps.change(
|
317 |
-
fn=update_sampling_steps,
|
318 |
-
inputs=[num_interpolation_steps, sample_mid_interpolation],
|
319 |
-
outputs=[sample_mid_interpolation],
|
320 |
-
)
|
321 |
-
with gr.Row():
|
322 |
-
remove_n_middle = gr.Slider(
|
323 |
-
label="Number of middle steps to remove from interpolation",
|
324 |
-
minimum=0,
|
325 |
-
maximum=80,
|
326 |
-
step=2,
|
327 |
-
value=0,
|
328 |
-
)
|
329 |
with gr.Row():
|
330 |
image_type = gr.Radio(
|
331 |
choices=[
|
@@ -372,6 +336,10 @@ Note: Running on CPU will take longer (approx. 6 minutes with default settings).
|
|
372 |
This demo is created as part of the 'Investigating the Effectiveness of Generative Diffusion Models in Synthesizing Images for Data Augmentation in Image Classification' dissertation.
|
373 |
|
374 |
The user can augment an image by interpolating between two prompts, and specify the number of interpolation steps and the specific step to generate the image.
|
|
|
|
|
|
|
|
|
375 |
"""
|
376 |
)
|
377 |
run_button.click(
|
@@ -389,27 +357,8 @@ The user can augment an image by interpolating between two prompts, and specify
|
|
389 |
interpolation_step,
|
390 |
num_inference_steps,
|
391 |
num_interpolation_steps,
|
392 |
-
sample_mid_interpolation,
|
393 |
-
remove_n_middle,
|
394 |
],
|
395 |
outputs=[result, show_seed, ssim_score, cos_sim],
|
396 |
)
|
397 |
|
398 |
demo.queue().launch(show_error=True)
|
399 |
-
|
400 |
-
"""
|
401 |
-
input_image,
|
402 |
-
prompt1,
|
403 |
-
prompt2,
|
404 |
-
negative_prompt,
|
405 |
-
seed,
|
406 |
-
randomize_seed,
|
407 |
-
width,
|
408 |
-
height,
|
409 |
-
guidance_scale,
|
410 |
-
interpolation_step,
|
411 |
-
num_inference_steps,
|
412 |
-
num_interpolation_steps,
|
413 |
-
sample_mid_interpolation,
|
414 |
-
remove_n_middle,
|
415 |
-
"""
|
|
|
4 |
import numpy as np
|
5 |
import torch
|
6 |
import torchvision.transforms as transforms
|
7 |
+
from PIL import Image
|
8 |
from torchmetrics.functional.image import structural_similarity_index_measure as ssim
|
9 |
from transformers import CLIPModel, CLIPProcessor
|
10 |
|
|
|
46 |
interpolation_step,
|
47 |
num_inference_steps,
|
48 |
num_interpolation_steps,
|
|
|
|
|
49 |
):
|
50 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
51 |
|
|
|
54 |
assert num_interpolation_steps % 2 == 0
|
55 |
except AssertionError:
|
56 |
raise ValueError("num_interpolation_steps must be an even number")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
|
58 |
if randomize_seed:
|
59 |
seed = random.randint(0, MAX_SEED)
|
60 |
prompts = [prompt1, prompt2]
|
61 |
generator = torch.Generator().manual_seed(seed)
|
62 |
|
63 |
+
sample_mid_interpolation = num_interpolation_steps
|
64 |
+
remove_n_middle = 0
|
65 |
+
|
66 |
interpolated_prompt_embeds, prompt_metadata = synth.interpolatePrompts(
|
67 |
prompts,
|
68 |
pipe,
|
|
|
92 |
).to(device)
|
93 |
embed_pairs = zip(interpolated_prompt_embeds, interpolated_negative_prompts_embeds)
|
94 |
embed_pairs_list = list(embed_pairs)
|
|
|
95 |
# offset step by -1
|
96 |
prompt_embeds, negative_prompt_embeds = embed_pairs_list[interpolation_step - 1]
|
97 |
preprocess_input = transforms.Compose(
|
|
|
102 |
npe = negative_prompt_embeds[None, ...]
|
103 |
else:
|
104 |
npe = None
|
105 |
+
images_list = pipe(
|
106 |
height=height,
|
107 |
width=width,
|
108 |
num_images_per_prompt=1,
|
|
|
113 |
generator=generator,
|
114 |
latents=latents,
|
115 |
image=input_img_tensor,
|
116 |
+
)
|
117 |
+
if images_list["nsfw_content_detected"][0]:
|
118 |
+
image = Image.open("samples/unsafe.jpeg")
|
119 |
+
return image, seed, "Unsafe content detected", "Unsafe content detected"
|
120 |
+
else:
|
121 |
+
image = images_list.images[0]
|
122 |
+
|
123 |
pred_image = transforms.ToTensor()(image).unsqueeze(0)
|
124 |
ssim_score = ssim(pred_image, input_img_tensor).item()
|
125 |
real_inputs = clip_processor(
|
|
|
144 |
|
145 |
|
146 |
examples1 = [
|
147 |
+
"A photo of a garbage truck, dustcart",
|
148 |
"A photo of a Shih-Tzu, a type of dog",
|
149 |
]
|
150 |
examples2 = [
|
151 |
+
"A photo of a cassette player",
|
152 |
"A photo of a beagle, a type of dog",
|
153 |
]
|
154 |
|
155 |
|
156 |
def update_steps(total_steps, interpolation_step):
|
157 |
+
return gr.update(maximum=total_steps)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
158 |
|
159 |
|
160 |
def update_format(image_format):
|
|
|
184 |
label="Prompt for the image to synthesize. (Actual class)",
|
185 |
show_label=True,
|
186 |
max_lines=1,
|
187 |
+
placeholder="Enter Prompt for the image to synthesize. (Actual class)",
|
188 |
container=False,
|
189 |
)
|
190 |
with gr.Row():
|
|
|
192 |
label="Prompt to augment against. (Confusing class)",
|
193 |
show_label=True,
|
194 |
max_lines=1,
|
195 |
+
placeholder="Enter Prompt to augment against. (Confusing class)",
|
196 |
container=False,
|
197 |
)
|
198 |
with gr.Row():
|
199 |
gr.Examples(
|
200 |
+
examples=[
|
201 |
+
"samples/n03417042_5234.JPEG",
|
202 |
+
"samples/n02086240_2799.JPEG",
|
203 |
+
],
|
204 |
+
inputs=[input_image],
|
205 |
+
label="Example Images",
|
206 |
+
)
|
207 |
+
gr.Examples(
|
208 |
+
examples=examples1,
|
209 |
+
inputs=[prompt1],
|
210 |
+
label="Example for Prompt 1 (Actual class)",
|
211 |
)
|
212 |
gr.Examples(
|
213 |
+
examples=examples2,
|
214 |
+
inputs=[prompt2],
|
215 |
+
label="Example for Prompt 2 (Confusing class)",
|
216 |
)
|
217 |
|
218 |
with gr.Row():
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
219 |
num_interpolation_steps = gr.Slider(
|
220 |
+
label="Total Interpolation Steps",
|
221 |
minimum=2,
|
222 |
+
maximum=128,
|
223 |
step=2,
|
224 |
value=16,
|
225 |
)
|
226 |
+
interpolation_step = gr.Slider(
|
227 |
+
label="Sample Interpolation Step",
|
228 |
+
minimum=1,
|
229 |
+
maximum=16,
|
230 |
+
step=1,
|
231 |
+
value=8,
|
232 |
+
)
|
233 |
num_interpolation_steps.change(
|
234 |
fn=update_steps,
|
235 |
inputs=[num_interpolation_steps, interpolation_step],
|
|
|
290 |
step=1,
|
291 |
value=25,
|
292 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
293 |
with gr.Row():
|
294 |
image_type = gr.Radio(
|
295 |
choices=[
|
|
|
336 |
This demo is created as part of the 'Investigating the Effectiveness of Generative Diffusion Models in Synthesizing Images for Data Augmentation in Image Classification' dissertation.
|
337 |
|
338 |
The user can augment an image by interpolating between two prompts, and specify the number of interpolation steps and the specific step to generate the image.
|
339 |
+
|
340 |
+
View the files used in this demo [here](https://huggingface.co/spaces/czl/generative-data-augmentation-demo/tree/main).
|
341 |
+
|
342 |
+
Note: Safety checker is enabled to prevent unsafe content from being displayed in this public demo.
|
343 |
"""
|
344 |
)
|
345 |
run_button.click(
|
|
|
357 |
interpolation_step,
|
358 |
num_inference_steps,
|
359 |
num_interpolation_steps,
|
|
|
|
|
360 |
],
|
361 |
outputs=[result, show_seed, ssim_score, cos_sim],
|
362 |
)
|
363 |
|
364 |
demo.queue().launch(show_error=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
samples/n02086240_2799.JPEG
ADDED
Git LFS Details
|
samples/n03417042_5234.JPEG
ADDED
Git LFS Details
|
samples/unsafe.jpeg
ADDED
Git LFS Details
|
tools/synth.py
CHANGED
@@ -157,7 +157,6 @@ def pipe_img(
|
|
157 |
scheduler=scheduler,
|
158 |
torch_dtype=torch.float32,
|
159 |
use_safetensors=use_safetensors,
|
160 |
-
safety_checker=None,
|
161 |
).to(device)
|
162 |
if cpu_offload:
|
163 |
pipe.enable_model_cpu_offload()
|
|
|
157 |
scheduler=scheduler,
|
158 |
torch_dtype=torch.float32,
|
159 |
use_safetensors=use_safetensors,
|
|
|
160 |
).to(device)
|
161 |
if cpu_offload:
|
162 |
pipe.enable_model_cpu_offload()
|