lam107 commited on
Commit
2bb45c2
ยท
verified ยท
1 Parent(s): 0599ae6

Upload app.py

Browse files
Files changed (1) hide show
  1. app.py +111 -0
app.py ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import sys
3
+ import zipfile
4
+ import requests
5
+ import pickle
6
+ import torch
7
+ from PIL import Image
8
+ import numpy as np
9
+ import torchvision.transforms as transforms
10
+ import cv2
11
+ from timm import create_model
12
+ import gradio as gr
13
+
14
+ # ุชุญู…ูŠู„ ูƒูˆุฏ StyleGAN2
15
+ if not os.path.exists("stylegan2-ada-pytorch"):
16
+ os.system("git clone https://github.com/NVlabs/stylegan2-ada-pytorch.git")
17
+ sys.path.append("stylegan2-ada-pytorch")
18
+
19
+ # ุชุญู…ูŠู„ ู…ู„ู zip ู…ู† HF model repo
20
+ ZIP_URL = "https://huggingface.co/lam107/stylegan2-models/resolve/main/stylegan2-ffhq-512x512.zip"
21
+ ZIP_PATH = "stylegan2-ffhq-512x512.zip"
22
+ PKL_PATH = "stylegan2-ffhq-512x512.pkl"
23
+
24
+ if not os.path.exists(PKL_PATH):
25
+ print("Downloading StyleGAN2 model...")
26
+ r = requests.get(ZIP_URL)
27
+ with open(ZIP_PATH, "wb") as f:
28
+ f.write(r.content)
29
+ print("Extracting...")
30
+ with zipfile.ZipFile(ZIP_PATH, 'r') as zip_ref:
31
+ zip_ref.extractall(".")
32
+
33
+ # ุชุฌู‡ูŠุฒ ุงู„ุฌู‡ุงุฒ
34
+ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
35
+
36
+ # ุชุญู…ูŠู„ Discriminator
37
+ with open(PKL_PATH, "rb") as f:
38
+ data = pickle.load(f)
39
+ D = data['D'].to(device)
40
+ D.eval()
41
+
42
+ # ุชุญู…ูŠู„ Swin model
43
+ class_names = ["real", "fake"]
44
+ model = create_model('swin_large_patch4_window12_384', pretrained=False, num_classes=len(class_names))
45
+ model.load_state_dict(torch.load("best_model_final.pth", map_location=device))
46
+ model = model.to(device)
47
+ model.eval()
48
+
49
+ def compute_gradients(model, pil_image):
50
+ transform = transforms.Compose([
51
+ transforms.Resize((512, 512)),
52
+ transforms.ToTensor(),
53
+ transforms.Normalize([0.5], [0.5])
54
+ ])
55
+ image = transform(pil_image.convert("RGB")).unsqueeze(0).to(device).requires_grad_(True)
56
+
57
+ c = torch.zeros([image.shape[0], model.c_dim], device=device) if hasattr(model, "c_dim") and model.c_dim > 0 else None
58
+ output = model(image, c)
59
+
60
+ loss = -output.mean()
61
+ loss.backward()
62
+
63
+ gradients = image.grad.detach().cpu().numpy()
64
+ gradients = np.abs(gradients).mean(axis=1)[0]
65
+ gradients = (gradients - gradients.min()) / (gradients.max() - gradients.min() + 1e-8)
66
+ gradients = (gradients * 255).astype(np.uint8)
67
+
68
+ grad_color = cv2.applyColorMap(gradients, cv2.COLORMAP_JET)
69
+ grad_rgb = cv2.cvtColor(grad_color, cv2.COLOR_BGR2RGB)
70
+
71
+ return Image.fromarray(grad_rgb)
72
+
73
+ def swin_classifier(pil_image):
74
+ transform = transforms.Compose([
75
+ transforms.Resize((384, 384)),
76
+ transforms.ToTensor(),
77
+ transforms.Normalize([0.485, 0.456, 0.406],
78
+ [0.229, 0.224, 0.225])
79
+ ])
80
+ image = transform(pil_image).unsqueeze(0).to(device)
81
+
82
+ with torch.no_grad():
83
+ output = model(image)
84
+ probabilities = torch.softmax(output, dim=1).squeeze().cpu().numpy()
85
+ return {"Real": float(probabilities[0]), "Fake": float(probabilities[1])}
86
+
87
+ def process_image(pil_img):
88
+ grad_pil = compute_gradients(D, pil_img).resize((384, 384))
89
+ prediction = swin_classifier(grad_pil)
90
+ return grad_pil, prediction
91
+
92
+ # ูˆุงุฌู‡ุฉ Gradio
93
+ def main():
94
+ with gr.Blocks() as demo:
95
+ gr.Markdown("**๐Ÿ’œ Mukhbir: AI Generated Image Detection**")
96
+ with gr.Row():
97
+ with gr.Column():
98
+ image_input = gr.Image(type="pil", label="Upload Image")
99
+ clear_button = gr.Button("Clear")
100
+ submit_button = gr.Button("Submit", variant="primary")
101
+ with gr.Column():
102
+ output_image = gr.Image(label="Gradient Image", type="pil")
103
+ output_label = gr.Label(label="Prediction")
104
+
105
+ submit_button.click(fn=process_image, inputs=image_input, outputs=[output_image, output_label])
106
+ clear_button.click(fn=lambda: (None, None, None), inputs=None, outputs=[image_input, output_image, output_label])
107
+
108
+ demo.launch(server_name="0.0.0.0", server_port=7860)
109
+
110
+ if __name__ == "__main__":
111
+ main()