Spaces:
Build error
Build error
Update app.py
Browse files
app.py
CHANGED
@@ -1,51 +1,282 @@
|
|
1 |
-
import os
|
2 |
-
import sys
|
3 |
-
from fastapi import FastAPI, Request
|
4 |
-
from fastapi.middleware.cors import CORSMiddleware
|
5 |
-
from fastapi.staticfiles import StaticFiles
|
6 |
-
from fastapi.responses import FileResponse, HTMLResponse
|
7 |
-
import uvicorn
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
app
|
29 |
-
|
30 |
-
#
|
31 |
-
app.
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
from fastapi import FastAPI, Request, UploadFile, File, HTTPException
|
4 |
+
from fastapi.middleware.cors import CORSMiddleware
|
5 |
+
from fastapi.staticfiles import StaticFiles
|
6 |
+
from fastapi.responses import JSONResponse, FileResponse, HTMLResponse
|
7 |
+
import uvicorn
|
8 |
+
import time
|
9 |
+
import shutil
|
10 |
+
import glob
|
11 |
+
import datetime
|
12 |
+
from random import choice
|
13 |
+
import torch
|
14 |
+
import torchvision
|
15 |
+
from torchvision import transforms
|
16 |
+
from torch import nn
|
17 |
+
import numpy as np
|
18 |
+
import cv2
|
19 |
+
import face_recognition
|
20 |
+
from PIL import Image as pImage
|
21 |
+
import matplotlib.pyplot as plt
|
22 |
+
import matplotlib
|
23 |
+
matplotlib.use('Agg') # Use non-GUI backend for matplotlib
|
24 |
+
from typing import List
|
25 |
+
import base64
|
26 |
+
import io
|
27 |
+
|
28 |
+
app = FastAPI()
|
29 |
+
|
30 |
+
# Configure CORS
|
31 |
+
app.add_middleware(
|
32 |
+
CORSMiddleware,
|
33 |
+
allow_origins=["*"],
|
34 |
+
allow_credentials=True,
|
35 |
+
allow_methods=["*"],
|
36 |
+
allow_headers=["*"],
|
37 |
+
)
|
38 |
+
|
39 |
+
# Create directories if they don't exist
|
40 |
+
os.makedirs("uploaded_images", exist_ok=True)
|
41 |
+
os.makedirs("static", exist_ok=True)
|
42 |
+
os.makedirs("uploaded_videos", exist_ok=True)
|
43 |
+
os.makedirs("models", exist_ok=True)
|
44 |
+
|
45 |
+
# Mount static files
|
46 |
+
app.mount("/uploaded_images", StaticFiles(directory="uploaded_images"), name="uploaded_images")
|
47 |
+
app.mount("/static", StaticFiles(directory="static"), name="static")
|
48 |
+
app.mount("/assets", StaticFiles(directory="frontend/dist/assets"), name="assets")
|
49 |
+
|
50 |
+
# Configuration
|
51 |
+
im_size = 112
|
52 |
+
mean = [0.485, 0.456, 0.406]
|
53 |
+
std = [0.229, 0.224, 0.225]
|
54 |
+
sm = nn.Softmax(dim=1)
|
55 |
+
inv_normalize = transforms.Normalize(
|
56 |
+
mean=-1*np.divide(mean, std), std=np.divide([1, 1, 1], std))
|
57 |
+
|
58 |
+
train_transforms = transforms.Compose([
|
59 |
+
transforms.ToPILImage(),
|
60 |
+
transforms.Resize((im_size, im_size)),
|
61 |
+
transforms.ToTensor(),
|
62 |
+
transforms.Normalize(mean, std)])
|
63 |
+
|
64 |
+
ALLOWED_VIDEO_EXTENSIONS = {'mp4', 'gif', 'webm', 'avi', '3gp', 'wmv', 'flv', 'mkv'}
|
65 |
+
|
66 |
+
# Detects GPU in device
|
67 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
68 |
+
|
69 |
+
class Model(nn.Module):
|
70 |
+
def __init__(self, num_classes, latent_dim=2048, lstm_layers=1, hidden_dim=2048, bidirectional=False):
|
71 |
+
super(Model, self).__init__()
|
72 |
+
model = torchvision.models.resnext50_32x4d(weights=torchvision.models.ResNeXt50_32X4D_Weights.DEFAULT)
|
73 |
+
self.model = nn.Sequential(*list(model.children())[:-2])
|
74 |
+
self.lstm = nn.LSTM(latent_dim, hidden_dim, lstm_layers, bidirectional)
|
75 |
+
self.relu = nn.LeakyReLU()
|
76 |
+
self.dp = nn.Dropout(0.4)
|
77 |
+
self.linear1 = nn.Linear(2048, num_classes)
|
78 |
+
self.avgpool = nn.AdaptiveAvgPool2d(1)
|
79 |
+
|
80 |
+
def forward(self, x):
|
81 |
+
batch_size, seq_length, c, h, w = x.shape
|
82 |
+
x = x.view(batch_size * seq_length, c, h, w)
|
83 |
+
fmap = self.model(x)
|
84 |
+
x = self.avgpool(fmap)
|
85 |
+
x = x.view(batch_size, seq_length, 2048)
|
86 |
+
x_lstm, _ = self.lstm(x, None)
|
87 |
+
return fmap, self.dp(self.linear1(x_lstm[:, -1, :]))
|
88 |
+
|
89 |
+
class ValidationDataset(torch.utils.data.Dataset):
|
90 |
+
def __init__(self, video_names, sequence_length=60, transform=None):
|
91 |
+
self.video_names = video_names
|
92 |
+
self.transform = transform
|
93 |
+
self.count = sequence_length
|
94 |
+
|
95 |
+
def __len__(self):
|
96 |
+
return len(self.video_names)
|
97 |
+
|
98 |
+
def __getitem__(self, idx):
|
99 |
+
video_path = self.video_names[idx]
|
100 |
+
frames = []
|
101 |
+
a = int(100/self.count)
|
102 |
+
first_frame = np.random.randint(0, a)
|
103 |
+
for i, frame in enumerate(self.frame_extract(video_path)):
|
104 |
+
faces = face_recognition.face_locations(frame)
|
105 |
+
try:
|
106 |
+
top, right, bottom, left = faces[0]
|
107 |
+
frame = frame[top:bottom, left:right, :]
|
108 |
+
except:
|
109 |
+
pass
|
110 |
+
frames.append(self.transform(frame))
|
111 |
+
if (len(frames) == self.count):
|
112 |
+
break
|
113 |
+
frames = torch.stack(frames)
|
114 |
+
frames = frames[:self.count]
|
115 |
+
return frames.unsqueeze(0) # Shape: (1, seq_len, C, H, W)
|
116 |
+
|
117 |
+
def frame_extract(self, path):
|
118 |
+
vidObj = cv2.VideoCapture(path)
|
119 |
+
success = 1
|
120 |
+
while success:
|
121 |
+
success, image = vidObj.read()
|
122 |
+
if success:
|
123 |
+
yield image
|
124 |
+
|
125 |
+
def allowed_video_file(filename):
|
126 |
+
return filename.split('.')[-1].lower() in ALLOWED_VIDEO_EXTENSIONS
|
127 |
+
|
128 |
+
def load_model(sequence_length=20):
|
129 |
+
"""Load the model from Hugging Face Hub if not available locally."""
|
130 |
+
model_path = os.path.join("models", "model.pt")
|
131 |
+
|
132 |
+
if not os.path.exists(model_path):
|
133 |
+
try:
|
134 |
+
from huggingface_hub import hf_hub_download
|
135 |
+
model_path = hf_hub_download(repo_id="tayyabimam/Deepfake",
|
136 |
+
filename="model.pt",
|
137 |
+
local_dir="models")
|
138 |
+
except Exception as e:
|
139 |
+
raise Exception(f"Failed to download model: {str(e)}")
|
140 |
+
|
141 |
+
# Load model
|
142 |
+
model = Model(2).to(device)
|
143 |
+
model.load_state_dict(torch.load(model_path, map_location=device))
|
144 |
+
model.eval()
|
145 |
+
return model
|
146 |
+
|
147 |
+
def im_convert(tensor, video_file_name=""):
|
148 |
+
"""Convert tensor to image for visualization."""
|
149 |
+
image = tensor.to("cpu").clone().detach()
|
150 |
+
image = image.squeeze()
|
151 |
+
image = inv_normalize(image)
|
152 |
+
image = image.numpy()
|
153 |
+
image = image.transpose(1, 2, 0)
|
154 |
+
image = image.clip(0, 1)
|
155 |
+
return image
|
156 |
+
|
157 |
+
def generate_gradcam_heatmap(model, img, video_file_name=""):
|
158 |
+
"""Generate GradCAM heatmap showing areas of focus for deepfake detection."""
|
159 |
+
# Forward pass
|
160 |
+
fmap, logits = model(img)
|
161 |
+
|
162 |
+
# Softmax on logits
|
163 |
+
logits_softmax = sm(logits)
|
164 |
+
confidence, prediction = torch.max(logits_softmax, 1)
|
165 |
+
confidence_val = confidence.item() * 100
|
166 |
+
pred_idx = prediction.item()
|
167 |
+
|
168 |
+
# Get weights and feature maps
|
169 |
+
weight_softmax = model.linear1.weight.detach().cpu().numpy()
|
170 |
+
fmap_last = fmap[-1].detach().cpu().numpy()
|
171 |
+
nc, h, w = fmap_last.shape
|
172 |
+
fmap_reshaped = fmap_last.reshape(nc, h*w)
|
173 |
+
|
174 |
+
# Compute GradCAM heatmap
|
175 |
+
heatmap_raw = np.dot(fmap_reshaped.T, weight_softmax[pred_idx, :].T)
|
176 |
+
heatmap_raw -= heatmap_raw.min()
|
177 |
+
heatmap_raw /= heatmap_raw.max()
|
178 |
+
heatmap_img = np.uint8(255 * heatmap_raw.reshape(h, w))
|
179 |
+
|
180 |
+
# Resize heatmap to model input size
|
181 |
+
heatmap_resized = cv2.resize(heatmap_img, (im_size, im_size))
|
182 |
+
heatmap_colored = cv2.applyColorMap(heatmap_resized, cv2.COLORMAP_JET)
|
183 |
+
|
184 |
+
# Convert original image tensor to numpy
|
185 |
+
original_img = im_convert(img[:, -1, :, :, :])
|
186 |
+
original_img_uint8 = (original_img * 255).astype(np.uint8)
|
187 |
+
|
188 |
+
# Overlay heatmap on original image
|
189 |
+
overlay = cv2.addWeighted(original_img_uint8, 0.6, heatmap_colored, 0.4, 0)
|
190 |
+
|
191 |
+
# Save overlayed image
|
192 |
+
timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
|
193 |
+
result_filename = f"result_{timestamp}.jpg"
|
194 |
+
save_path = os.path.join("static", result_filename)
|
195 |
+
plt.figure(figsize=(10, 5))
|
196 |
+
|
197 |
+
# Plot original and heatmap
|
198 |
+
plt.subplot(1, 2, 1)
|
199 |
+
plt.imshow(original_img)
|
200 |
+
plt.title("Original")
|
201 |
+
plt.axis('off')
|
202 |
+
|
203 |
+
plt.subplot(1, 2, 2)
|
204 |
+
plt.imshow(cv2.cvtColor(overlay, cv2.COLOR_BGR2RGB))
|
205 |
+
plt.title(f"{'FAKE' if pred_idx == 1 else 'REAL'} ({confidence_val:.2f}%)")
|
206 |
+
plt.axis('off')
|
207 |
+
|
208 |
+
plt.tight_layout()
|
209 |
+
plt.savefig(save_path)
|
210 |
+
plt.close()
|
211 |
+
|
212 |
+
return {
|
213 |
+
"prediction": "FAKE" if pred_idx == 1 else "REAL",
|
214 |
+
"confidence": confidence_val,
|
215 |
+
"heatmap_url": f"/static/{result_filename}",
|
216 |
+
"original_filename": video_file_name
|
217 |
+
}
|
218 |
+
|
219 |
+
def predict_with_gradcam(model, img, video_file_name=""):
|
220 |
+
"""Predict with GradCAM visualization."""
|
221 |
+
return generate_gradcam_heatmap(model, img, video_file_name)
|
222 |
+
|
223 |
+
@app.post("/api/upload-video")
|
224 |
+
async def api_upload_video(file: UploadFile = File(...), sequence_length: int = 20):
|
225 |
+
"""API endpoint for video upload and analysis."""
|
226 |
+
if not allowed_video_file(file.filename):
|
227 |
+
raise HTTPException(status_code=400, detail="Invalid file format. Supported formats: mp4, gif, webm, avi, 3gp, wmv, flv, mkv")
|
228 |
+
|
229 |
+
# Save uploaded file
|
230 |
+
temp_file = f"uploaded_videos/{file.filename}"
|
231 |
+
with open(temp_file, "wb") as buffer:
|
232 |
+
shutil.copyfileobj(file.file, buffer)
|
233 |
+
|
234 |
+
try:
|
235 |
+
# Process the video
|
236 |
+
result = process_video(temp_file, sequence_length)
|
237 |
+
return result
|
238 |
+
except Exception as e:
|
239 |
+
raise HTTPException(status_code=500, detail=str(e))
|
240 |
+
|
241 |
+
def process_video(video_file, sequence_length):
|
242 |
+
"""Process video for deepfake detection."""
|
243 |
+
# Load model
|
244 |
+
model = load_model(sequence_length)
|
245 |
+
|
246 |
+
# Prepare dataset
|
247 |
+
test_dataset = ValidationDataset(video_names=[video_file],
|
248 |
+
sequence_length=sequence_length,
|
249 |
+
transform=train_transforms)
|
250 |
+
|
251 |
+
# Get frames
|
252 |
+
frames = test_dataset[0]
|
253 |
+
frames = frames.to(device)
|
254 |
+
|
255 |
+
# Make prediction with GradCAM
|
256 |
+
result = predict_with_gradcam(model, frames, os.path.basename(video_file))
|
257 |
+
|
258 |
+
return result
|
259 |
+
|
260 |
+
@app.get("/{path:path}")
|
261 |
+
async def serve_frontend(path: str):
|
262 |
+
# First check if the path exists in the frontend dist
|
263 |
+
if os.path.exists(f"frontend/dist/{path}"):
|
264 |
+
return FileResponse(f"frontend/dist/{path}")
|
265 |
+
|
266 |
+
# Otherwise return the index.html
|
267 |
+
return FileResponse("frontend/dist/index.html")
|
268 |
+
|
269 |
+
@app.get("/", response_class=HTMLResponse)
|
270 |
+
async def root():
|
271 |
+
return FileResponse("frontend/dist/index.html")
|
272 |
+
|
273 |
+
@app.get("/api")
|
274 |
+
async def api_root():
|
275 |
+
"""Root endpoint with API documentation."""
|
276 |
+
return {
|
277 |
+
"message": "Welcome to DeepSight DeepFake Detection API",
|
278 |
+
"usage": "POST /api/upload-video with a video file to detect deepfakes"
|
279 |
+
}
|
280 |
+
|
281 |
+
if __name__ == "__main__":
|
282 |
+
uvicorn.run(app, host="0.0.0.0", port=7860)
|