Spaces:
Sleeping
Sleeping
lmao lets goooo
Browse files- app.py +40 -5
- requirements.txt +5 -1
app.py
CHANGED
@@ -1,18 +1,53 @@
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
2 |
|
3 |
def process_image(input_image, learning_rate, iterations):
|
4 |
-
|
5 |
-
|
6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
iface = gr.Interface(
|
9 |
fn=process_image,
|
10 |
inputs=[
|
11 |
-
gr.Image(),
|
12 |
gr.Number(value=0.01, label="Learning Rate"),
|
13 |
gr.Number(value=1, label="Iterations")
|
14 |
],
|
15 |
-
outputs=gr.Image()
|
16 |
)
|
17 |
|
18 |
iface.launch()
|
|
|
1 |
import gradio as gr
|
2 |
+
import torch
|
3 |
+
import numpy as np
|
4 |
+
from transformers import ViTImageProcessor, ViTForImageClassification
|
5 |
+
from PIL import Image
|
6 |
|
7 |
def process_image(input_image, learning_rate, iterations):
|
8 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
9 |
+
|
10 |
+
feature_extractor = ViTImageProcessor.from_pretrained('google/vit-large-patch32-384')
|
11 |
+
model = ViTForImageClassification.from_pretrained('google/vit-large-patch32-384')
|
12 |
+
model.to(device)
|
13 |
+
model.eval()
|
14 |
+
|
15 |
+
def get_encoder_activations(x):
|
16 |
+
encoder_output = model.vit(x)
|
17 |
+
final_activations = encoder_output.last_hidden_state
|
18 |
+
return final_activations
|
19 |
+
|
20 |
+
image = input_image.convert('RGB')
|
21 |
+
pixel_values = feature_extractor(images=image, return_tensors="pt").pixel_values
|
22 |
+
pixel_values.to(device)
|
23 |
+
pixel_values.requires_grad_(True)
|
24 |
+
|
25 |
+
for iteration in range(iterations.value):
|
26 |
+
model.zero_grad()
|
27 |
+
if pixel_values.grad is not None:
|
28 |
+
pixel_values.grad.data.zero_()
|
29 |
+
|
30 |
+
final_activations = get_encoder_activations(pixel_values.to('cuda'))
|
31 |
+
target_sum = final_activations.sum()
|
32 |
+
target_sum.backward()
|
33 |
+
|
34 |
+
with torch.no_grad():
|
35 |
+
pixel_values.data += learning_rate.value * pixel_values.grad.data
|
36 |
+
pixel_values.data = torch.clamp(pixel_values.data, -1, 1)
|
37 |
+
|
38 |
+
updated_pixel_values_np = 127.5 + pixel_values.squeeze().permute(1, 2, 0).detach().cpu() * 127.5
|
39 |
+
updated_pixel_values_np = updated_pixel_values_np.numpy().astype(np.uint8)
|
40 |
+
|
41 |
+
return updated_pixel_values_np
|
42 |
|
43 |
iface = gr.Interface(
|
44 |
fn=process_image,
|
45 |
inputs=[
|
46 |
+
gr.Image(type="pil"),
|
47 |
gr.Number(value=0.01, label="Learning Rate"),
|
48 |
gr.Number(value=1, label="Iterations")
|
49 |
],
|
50 |
+
outputs=gr.Image(type="numpy", label="Processed Image")
|
51 |
)
|
52 |
|
53 |
iface.launch()
|
requirements.txt
CHANGED
@@ -1 +1,5 @@
|
|
1 |
-
gradio
|
|
|
|
|
|
|
|
|
|
1 |
+
gradio
|
2 |
+
torch
|
3 |
+
numpy
|
4 |
+
transformers
|
5 |
+
PIL
|