Update app.py
Browse files
app.py
CHANGED
@@ -3,8 +3,7 @@ import torch
|
|
3 |
import torch.nn as nn
|
4 |
from torch.utils.data import DataLoader
|
5 |
from torchvision import transforms
|
6 |
-
from transformers import
|
7 |
-
from transformers.models.clip import CLIPModel
|
8 |
from PIL import Image
|
9 |
import numpy as np
|
10 |
import io
|
@@ -16,6 +15,7 @@ from unsloth import FastVisionModel
|
|
16 |
import os
|
17 |
import tempfile
|
18 |
import warnings
|
|
|
19 |
warnings.filterwarnings("ignore", category=UserWarning)
|
20 |
|
21 |
# App title and description
|
@@ -42,14 +42,13 @@ def check_gpu():
|
|
42 |
# Sidebar components
|
43 |
st.sidebar.title("About")
|
44 |
st.sidebar.markdown("""
|
45 |
-
This tool detects deepfakes using
|
46 |
-
- **
|
47 |
-
- **GradCAM**: Highlights suspicious regions
|
48 |
- **BLIP**: Describes image content
|
49 |
- **Llama 3.2**: Explains potential manipulations
|
50 |
|
51 |
### Quick Start
|
52 |
-
1. **Load Models** - Start with
|
53 |
2. **Upload Image** - View classification and heat map
|
54 |
3. **Analyze** - Get explanations and ask questions
|
55 |
|
@@ -72,8 +71,7 @@ if use_custom_instructions:
|
|
72 |
else:
|
73 |
custom_instruction = ""
|
74 |
|
75 |
-
# ----- GradCAM Implementation -----
|
76 |
-
|
77 |
class ImageDataset(torch.utils.data.Dataset):
|
78 |
def __init__(self, image, transform=None, face_only=True, dataset_name=None):
|
79 |
self.image = image
|
@@ -149,262 +147,45 @@ class ImageDataset(torch.utils.data.Dataset):
|
|
149 |
|
150 |
return image_tensor, label, "uploaded_image", original_image, None, self.dataset_name
|
151 |
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
if isinstance(output, tuple):
|
163 |
-
self.activations = output[0]
|
164 |
-
else:
|
165 |
-
self.activations = output
|
166 |
-
|
167 |
-
def backward_hook(module, grad_in, grad_out):
|
168 |
-
if isinstance(grad_out, tuple):
|
169 |
-
self.gradients = grad_out[0]
|
170 |
-
else:
|
171 |
-
self.gradients = grad_out
|
172 |
-
|
173 |
-
layer = dict([*self.model.named_modules()])[self.target_layer]
|
174 |
-
layer.register_forward_hook(forward_hook)
|
175 |
-
layer.register_backward_hook(backward_hook)
|
176 |
-
|
177 |
-
def generate(self, input_tensor, class_idx):
|
178 |
-
self.model.zero_grad()
|
179 |
-
|
180 |
-
try:
|
181 |
-
# Use only the vision part of the model for gradient calculation
|
182 |
-
vision_outputs = self.model.vision_model(pixel_values=input_tensor)
|
183 |
-
|
184 |
-
# Get the pooler output
|
185 |
-
features = vision_outputs.pooler_output
|
186 |
-
|
187 |
-
# Create a dummy gradient for the feature based on the class idx
|
188 |
-
one_hot = torch.zeros_like(features)
|
189 |
-
one_hot[0, class_idx] = 1
|
190 |
-
|
191 |
-
# Manually backpropagate
|
192 |
-
features.backward(gradient=one_hot)
|
193 |
-
|
194 |
-
# Check for None values
|
195 |
-
if self.gradients is None or self.activations is None:
|
196 |
-
st.warning("Warning: Gradients or activations are None. Using fallback CAM.")
|
197 |
-
return np.ones((14, 14), dtype=np.float32) * 0.5
|
198 |
-
|
199 |
-
# Process gradients and activations for transformer-based model
|
200 |
-
gradients = self.gradients.cpu().detach().numpy()
|
201 |
-
activations = self.activations.cpu().detach().numpy()
|
202 |
-
|
203 |
-
if len(activations.shape) == 3: # [batch, sequence_length, hidden_dim]
|
204 |
-
seq_len = activations.shape[1]
|
205 |
-
|
206 |
-
# CLIP ViT typically has 196 patch tokens (14Γ14) + 1 class token = 197
|
207 |
-
if seq_len >= 197:
|
208 |
-
# Skip the class token (first token) and reshape the patch tokens into a square
|
209 |
-
patch_tokens = activations[0, 1:197, :] # Remove the class token
|
210 |
-
# Take the mean across the hidden dimension
|
211 |
-
token_importance = np.mean(np.abs(patch_tokens), axis=1)
|
212 |
-
# Reshape to the expected grid size (14Γ14 for CLIP ViT)
|
213 |
-
cam = token_importance.reshape(14, 14)
|
214 |
-
else:
|
215 |
-
# Try to find factors close to a square
|
216 |
-
side_len = int(np.sqrt(seq_len))
|
217 |
-
# Use the mean across features as importance
|
218 |
-
token_importance = np.mean(np.abs(activations[0]), axis=1)
|
219 |
-
# Create as square-like shape as possible
|
220 |
-
cam = np.zeros((side_len, side_len))
|
221 |
-
# Fill the cam with available values
|
222 |
-
flat_cam = cam.flatten()
|
223 |
-
flat_cam[:min(len(token_importance), len(flat_cam))] = token_importance[:min(len(token_importance), len(flat_cam))]
|
224 |
-
cam = flat_cam.reshape(side_len, side_len)
|
225 |
-
else:
|
226 |
-
# Fallback
|
227 |
-
st.info("Using fallback CAM shape (14x14)")
|
228 |
-
cam = np.ones((14, 14), dtype=np.float32) * 0.5 # Default fallback
|
229 |
-
|
230 |
-
# Ensure we have valid values
|
231 |
-
cam = np.maximum(cam, 0)
|
232 |
-
if np.max(cam) > 0:
|
233 |
-
cam = cam / np.max(cam)
|
234 |
-
|
235 |
-
return cam
|
236 |
-
|
237 |
-
except Exception as e:
|
238 |
-
st.error(f"Error in GradCAM.generate: {str(e)}")
|
239 |
-
return np.ones((14, 14), dtype=np.float32) * 0.5
|
240 |
-
|
241 |
-
def overlay_cam_on_image(image, cam, face_box=None, alpha=0.5):
|
242 |
-
"""Overlay the CAM on the image"""
|
243 |
-
if face_box is not None:
|
244 |
-
x, y, w, h = face_box
|
245 |
-
# Create a mask for the entire image (all zeros initially)
|
246 |
-
img_np = np.array(image)
|
247 |
-
full_h, full_w = img_np.shape[:2]
|
248 |
-
full_cam = np.zeros((full_h, full_w), dtype=np.float32)
|
249 |
-
|
250 |
-
# Resize CAM to match face region
|
251 |
-
face_cam = cv2.resize(cam, (w, h))
|
252 |
-
|
253 |
-
# Copy the face CAM into the full image CAM at the face position
|
254 |
-
full_cam[y:y+h, x:x+w] = face_cam
|
255 |
-
|
256 |
-
# Convert full CAM to image
|
257 |
-
cam_resized = Image.fromarray((full_cam * 255).astype(np.uint8))
|
258 |
-
cam_colormap = plt.cm.jet(np.array(cam_resized) / 255.0)[:, :, :3] # Apply colormap
|
259 |
-
cam_colormap = (cam_colormap * 255).astype(np.uint8)
|
260 |
-
else:
|
261 |
-
# Resize CAM to match image dimensions
|
262 |
-
img_np = np.array(image)
|
263 |
-
h, w = img_np.shape[:2]
|
264 |
-
cam_resized = cv2.resize(cam, (w, h))
|
265 |
-
|
266 |
-
# Apply colormap
|
267 |
-
cam_colormap = plt.cm.jet(cam_resized)[:, :, :3] # Apply colormap
|
268 |
-
cam_colormap = (cam_colormap * 255).astype(np.uint8)
|
269 |
-
|
270 |
-
# Blend the original image with the colormap
|
271 |
-
img_np_float = img_np.astype(float) / 255.0
|
272 |
-
cam_colormap_float = cam_colormap.astype(float) / 255.0
|
273 |
-
|
274 |
-
blended = img_np_float * (1 - alpha) + cam_colormap_float * alpha
|
275 |
-
blended = (blended * 255).astype(np.uint8)
|
276 |
|
277 |
-
|
278 |
-
|
279 |
-
def save_comparison(image, cam, overlay, face_box=None):
|
280 |
-
"""Create a side-by-side comparison of the original, CAM, and overlay"""
|
281 |
-
fig, axes = plt.subplots(1, 3, figsize=(15, 5))
|
282 |
-
|
283 |
-
# Original Image
|
284 |
-
axes[0].imshow(image)
|
285 |
-
axes[0].set_title("Original")
|
286 |
-
if face_box is not None:
|
287 |
-
x, y, w, h = face_box
|
288 |
-
rect = plt.Rectangle((x, y), w, h, edgecolor='lime', linewidth=2, fill=False)
|
289 |
-
axes[0].add_patch(rect)
|
290 |
-
axes[0].axis("off")
|
291 |
-
|
292 |
-
# CAM
|
293 |
-
if face_box is not None:
|
294 |
-
# Create a full image CAM that highlights only the face
|
295 |
-
img_np = np.array(image)
|
296 |
-
h, w = img_np.shape[:2]
|
297 |
-
full_cam = np.zeros((h, w))
|
298 |
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
else:
|
306 |
-
|
307 |
-
|
308 |
-
axes[1].set_title("CAM")
|
309 |
-
axes[1].axis("off")
|
310 |
-
|
311 |
-
# Overlay
|
312 |
-
axes[2].imshow(overlay)
|
313 |
-
axes[2].set_title("Overlay")
|
314 |
-
axes[2].axis("off")
|
315 |
|
316 |
-
|
317 |
-
|
318 |
-
# Convert plot to PIL Image for Streamlit display
|
319 |
-
buf = io.BytesIO()
|
320 |
-
plt.savefig(buf, format="png", bbox_inches="tight")
|
321 |
-
plt.close()
|
322 |
-
buf.seek(0)
|
323 |
-
return Image.open(buf)
|
324 |
-
|
325 |
-
# Function to load GradCAM CLIP model
|
326 |
@st.cache_resource
|
327 |
-
def
|
328 |
-
|
|
|
329 |
try:
|
330 |
-
model =
|
331 |
-
|
332 |
-
|
333 |
-
model.
|
334 |
-
model.classification_head.weight.data.normal_(mean=0.0, std=0.02)
|
335 |
-
model.classification_head.bias.data.zero_()
|
336 |
-
|
337 |
model.eval()
|
338 |
-
return model
|
339 |
-
except Exception as e:
|
340 |
-
st.error(f"Error loading CLIP model: {str(e)}")
|
341 |
-
return None
|
342 |
-
|
343 |
-
def get_target_layer_clip(model):
|
344 |
-
"""Get the target layer for GradCAM"""
|
345 |
-
return "vision_model.encoder.layers.23"
|
346 |
-
|
347 |
-
def process_image_with_gradcam(image, model, device, pred_class):
|
348 |
-
"""Process an image with GradCAM"""
|
349 |
-
# Set up transformations
|
350 |
-
transform = transforms.Compose([
|
351 |
-
transforms.Resize((224, 224)),
|
352 |
-
transforms.ToTensor(),
|
353 |
-
transforms.Normalize(mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711]),
|
354 |
-
])
|
355 |
-
|
356 |
-
# Create dataset for the single image
|
357 |
-
dataset = ImageDataset(image, transform=transform, face_only=True)
|
358 |
-
|
359 |
-
# Custom collate function
|
360 |
-
def custom_collate(batch):
|
361 |
-
tensors = [item[0] for item in batch]
|
362 |
-
labels = [item[1] for item in batch]
|
363 |
-
paths = [item[2] for item in batch]
|
364 |
-
images = [item[3] for item in batch]
|
365 |
-
face_boxes = [item[4] for item in batch]
|
366 |
-
dataset_names = [item[5] for item in batch]
|
367 |
-
|
368 |
-
tensors = torch.stack(tensors)
|
369 |
-
labels = torch.tensor(labels)
|
370 |
-
|
371 |
-
return tensors, labels, paths, images, face_boxes, dataset_names
|
372 |
-
|
373 |
-
# Create dataloader
|
374 |
-
dataloader = DataLoader(dataset, batch_size=1, shuffle=False, collate_fn=custom_collate)
|
375 |
-
|
376 |
-
# Extract the batch
|
377 |
-
for batch in dataloader:
|
378 |
-
input_tensor, label, img_paths, original_images, face_boxes, dataset_names = batch
|
379 |
-
original_image = original_images[0]
|
380 |
-
face_box = face_boxes[0]
|
381 |
-
|
382 |
-
# Move tensors and model to device
|
383 |
-
input_tensor = input_tensor.to(device)
|
384 |
-
model = model.to(device)
|
385 |
-
|
386 |
-
try:
|
387 |
-
# Create GradCAM extractor
|
388 |
-
target_layer = get_target_layer_clip(model)
|
389 |
-
cam_extractor = GradCAM(model, target_layer)
|
390 |
-
|
391 |
-
# Generate CAM
|
392 |
-
cam = cam_extractor.generate(input_tensor, pred_class)
|
393 |
-
|
394 |
-
# Create visualizations
|
395 |
-
overlay = overlay_cam_on_image(original_image, cam, face_box)
|
396 |
-
comparison = save_comparison(original_image, cam, overlay, face_box)
|
397 |
-
|
398 |
-
# Return results
|
399 |
-
return cam, overlay, comparison, face_box
|
400 |
-
|
401 |
except Exception as e:
|
402 |
-
st.error(f"Error
|
403 |
-
|
404 |
-
default_cam = np.ones((14, 14), dtype=np.float32) * 0.5
|
405 |
-
overlay = overlay_cam_on_image(original_image, default_cam, face_box)
|
406 |
-
comparison = save_comparison(original_image, default_cam, overlay, face_box)
|
407 |
-
return default_cam, overlay, comparison, face_box
|
408 |
|
409 |
# ----- BLIP Image Captioning -----
|
410 |
|
@@ -624,12 +405,47 @@ def analyze_image_with_llm(image, gradcam_overlay, face_box, pred_label, confide
|
|
624 |
st.error(f"Error during LLM analysis: {str(e)}")
|
625 |
return f"Error analyzing image: {str(e)}"
|
626 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
627 |
# Main app
|
628 |
def main():
|
629 |
# Initialize session state variables
|
630 |
-
if '
|
631 |
-
st.session_state.
|
632 |
-
st.session_state.
|
633 |
|
634 |
if 'llm_model_loaded' not in st.session_state:
|
635 |
st.session_state.llm_model_loaded = False
|
@@ -652,21 +468,22 @@ def main():
|
|
652 |
st.write("Please load the models using the buttons below:")
|
653 |
|
654 |
# Button for loading models
|
655 |
-
|
656 |
|
657 |
-
with
|
658 |
-
if not st.session_state.
|
659 |
-
if st.button("π₯ Load
|
660 |
-
# Load
|
661 |
-
model =
|
662 |
if model is not None:
|
663 |
-
st.session_state.
|
664 |
-
st.session_state.
|
665 |
-
st.
|
|
|
666 |
else:
|
667 |
-
st.error("β Failed to load
|
668 |
else:
|
669 |
-
st.success("β
|
670 |
|
671 |
with blip_col:
|
672 |
if not st.session_state.blip_model_loaded:
|
@@ -724,56 +541,56 @@ def main():
|
|
724 |
st.session_state.original_model
|
725 |
)
|
726 |
st.session_state.image_caption = caption
|
727 |
-
|
728 |
-
# Store caption but don't display it yet
|
729 |
|
730 |
-
# Detect with
|
731 |
-
if st.session_state.
|
732 |
-
with st.spinner("Analyzing image with
|
733 |
-
# Preprocess image for
|
734 |
-
|
735 |
-
transforms.Resize((224, 224)),
|
736 |
-
transforms.ToTensor(),
|
737 |
-
transforms.Normalize(mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711]),
|
738 |
-
])
|
739 |
|
740 |
-
#
|
741 |
-
|
742 |
-
|
743 |
-
tensor = tensor.unsqueeze(0)
|
744 |
|
745 |
-
#
|
746 |
-
|
747 |
-
|
748 |
-
# Move model and tensor to device
|
749 |
-
model = st.session_state.clip_model.to(device)
|
750 |
-
tensor = tensor.to(device)
|
751 |
|
752 |
# Forward pass
|
753 |
with torch.no_grad():
|
754 |
-
|
755 |
-
|
756 |
-
|
757 |
-
|
758 |
-
|
759 |
-
pred_label = "Fake" if pred_class == 1 else "Real"
|
760 |
|
761 |
# Display results
|
762 |
with col2:
|
763 |
st.markdown("### Detection Result")
|
764 |
st.markdown(f"**Classification:** {pred_label} (Confidence: {confidence:.2%})")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
765 |
|
766 |
# GradCAM visualization
|
767 |
st.subheader("GradCAM Visualization")
|
768 |
-
cam, overlay, comparison, detected_face_box =
|
769 |
image, model, device, pred_class
|
770 |
)
|
771 |
|
772 |
-
|
773 |
-
|
|
|
|
|
|
|
|
|
774 |
|
775 |
# Generate caption for GradCAM overlay image if BLIP model is loaded
|
776 |
-
if st.session_state.blip_model_loaded:
|
777 |
with st.spinner("Analyzing GradCAM visualization..."):
|
778 |
gradcam_caption = generate_gradcam_caption(
|
779 |
overlay,
|
@@ -781,8 +598,6 @@ def main():
|
|
781 |
st.session_state.finetuned_model
|
782 |
)
|
783 |
st.session_state.gradcam_caption = gradcam_caption
|
784 |
-
|
785 |
-
# Store caption but don't display it yet
|
786 |
|
787 |
# Save results in session state for LLM analysis
|
788 |
st.session_state.current_image = image
|
@@ -793,7 +608,7 @@ def main():
|
|
793 |
|
794 |
st.success("β
Initial detection and GradCAM visualization complete!")
|
795 |
else:
|
796 |
-
st.warning("β οΈ Please load the
|
797 |
except Exception as e:
|
798 |
st.error(f"Error processing image: {str(e)}")
|
799 |
import traceback
|
@@ -930,7 +745,7 @@ def main():
|
|
930 |
st.markdown("---")
|
931 |
|
932 |
# Add model version indicator in sidebar
|
933 |
-
st.sidebar.info("Using deepfake-explainer-2
|
934 |
|
935 |
if __name__ == "__main__":
|
936 |
main()
|
|
|
3 |
import torch.nn as nn
|
4 |
from torch.utils.data import DataLoader
|
5 |
from torchvision import transforms
|
6 |
+
from transformers import BlipProcessor, BlipForConditionalGeneration
|
|
|
7 |
from PIL import Image
|
8 |
import numpy as np
|
9 |
import io
|
|
|
15 |
import os
|
16 |
import tempfile
|
17 |
import warnings
|
18 |
+
from gradcam_xception import load_xception_model, generate_smoothgrad_visualizations_xception, get_xception_transform
|
19 |
warnings.filterwarnings("ignore", category=UserWarning)
|
20 |
|
21 |
# App title and description
|
|
|
42 |
# Sidebar components
|
43 |
st.sidebar.title("About")
|
44 |
st.sidebar.markdown("""
|
45 |
+
This tool detects deepfakes using three AI models:
|
46 |
+
- **Xception**: Initial Real/Fake classification
|
|
|
47 |
- **BLIP**: Describes image content
|
48 |
- **Llama 3.2**: Explains potential manipulations
|
49 |
|
50 |
### Quick Start
|
51 |
+
1. **Load Models** - Start with Xception, add others as needed
|
52 |
2. **Upload Image** - View classification and heat map
|
53 |
3. **Analyze** - Get explanations and ask questions
|
54 |
|
|
|
71 |
else:
|
72 |
custom_instruction = ""
|
73 |
|
74 |
+
# ----- GradCAM Implementation for Xception -----
|
|
|
75 |
class ImageDataset(torch.utils.data.Dataset):
|
76 |
def __init__(self, image, transform=None, face_only=True, dataset_name=None):
|
77 |
self.image = image
|
|
|
147 |
|
148 |
return image_tensor, label, "uploaded_image", original_image, None, self.dataset_name
|
149 |
|
150 |
+
# Function to process image with Xception GradCAM
|
151 |
+
def process_image_with_xception_gradcam(image, model, device, pred_class):
|
152 |
+
"""Process an image with Xception GradCAM"""
|
153 |
+
cam_results = generate_smoothgrad_visualizations_xception(
|
154 |
+
model=model,
|
155 |
+
image=image,
|
156 |
+
target_class=pred_class,
|
157 |
+
face_only=True,
|
158 |
+
num_samples=5 # Can be adjusted
|
159 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
160 |
|
161 |
+
if cam_results and len(cam_results) == 4:
|
162 |
+
raw_cam, cam_img, overlay, comparison = cam_results
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
163 |
|
164 |
+
# Extract the face box from the dataset if needed
|
165 |
+
transform = get_xception_transform()
|
166 |
+
dataset = ImageDataset(image, transform=transform, face_only=True)
|
167 |
+
_, _, _, _, face_box, _ = dataset[0]
|
168 |
+
|
169 |
+
return raw_cam, overlay, comparison, face_box
|
170 |
else:
|
171 |
+
st.error("Failed to generate GradCAM visualization")
|
172 |
+
return None, None, None, None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
173 |
|
174 |
+
# ----- Xception Model Loading -----
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
@st.cache_resource
|
176 |
+
def load_detection_model_xception():
|
177 |
+
"""Loads the Xception model from our module"""
|
178 |
+
with st.spinner("Loading Xception model for deepfake detection..."):
|
179 |
try:
|
180 |
+
model = load_xception_model()
|
181 |
+
# Get the device
|
182 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
183 |
+
model.to(device)
|
|
|
|
|
|
|
184 |
model.eval()
|
185 |
+
return model, device
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
186 |
except Exception as e:
|
187 |
+
st.error(f"Error loading Xception model: {str(e)}")
|
188 |
+
return None, None
|
|
|
|
|
|
|
|
|
189 |
|
190 |
# ----- BLIP Image Captioning -----
|
191 |
|
|
|
405 |
st.error(f"Error during LLM analysis: {str(e)}")
|
406 |
return f"Error analyzing image: {str(e)}"
|
407 |
|
408 |
+
# Preprocess image for Xception
|
409 |
+
def preprocess_image_xception(image):
|
410 |
+
"""Preprocesses image for Xception model input and face detection."""
|
411 |
+
face_detector = cv2.CascadeClassifier(cv2.data.haarcascades + 'haarcascade_frontalface_default.xml')
|
412 |
+
image_np = np.array(image.convert('RGB')) # Ensure RGB
|
413 |
+
gray = cv2.cvtColor(image_np, cv2.COLOR_RGB2GRAY)
|
414 |
+
faces = face_detector.detectMultiScale(gray, 1.1, 5)
|
415 |
+
|
416 |
+
face_img_for_transform = image # Default to whole image
|
417 |
+
face_box_display = None # For drawing on original image
|
418 |
+
|
419 |
+
if len(faces) == 0:
|
420 |
+
st.warning("No face detected, using whole image for prediction/CAM.")
|
421 |
+
else:
|
422 |
+
areas = [w * h for (x, y, w, h) in faces]
|
423 |
+
largest_idx = np.argmax(areas)
|
424 |
+
x, y, w, h = faces[largest_idx]
|
425 |
+
padding_x = int(w * 0.05) # Use percentages as in gradcam_xception
|
426 |
+
padding_y = int(h * 0.05)
|
427 |
+
x1, y1 = max(0, x - padding_x), max(0, y - padding_y)
|
428 |
+
x2, y2 = min(image_np.shape[1], x + w + padding_x), min(image_np.shape[0], y + h + padding_y)
|
429 |
+
|
430 |
+
# Use the padded face region for the model transform
|
431 |
+
face_img_for_transform = Image.fromarray(image_np[y1:y2, x1:x2])
|
432 |
+
# Use the original detected box (without padding) for display rectangle
|
433 |
+
face_box_display = (x, y, w, h)
|
434 |
+
|
435 |
+
# Xception specific transform
|
436 |
+
transform = get_xception_transform()
|
437 |
+
# Apply transform to the selected region (face or whole image)
|
438 |
+
input_tensor = transform(face_img_for_transform).unsqueeze(0)
|
439 |
+
|
440 |
+
# Return tensor, original full image, and the display face box
|
441 |
+
return input_tensor, image, face_box_display
|
442 |
+
|
443 |
# Main app
|
444 |
def main():
|
445 |
# Initialize session state variables
|
446 |
+
if 'xception_model_loaded' not in st.session_state:
|
447 |
+
st.session_state.xception_model_loaded = False
|
448 |
+
st.session_state.xception_model = None
|
449 |
|
450 |
if 'llm_model_loaded' not in st.session_state:
|
451 |
st.session_state.llm_model_loaded = False
|
|
|
468 |
st.write("Please load the models using the buttons below:")
|
469 |
|
470 |
# Button for loading models
|
471 |
+
xception_col, blip_col, llm_col = st.columns(3)
|
472 |
|
473 |
+
with xception_col:
|
474 |
+
if not st.session_state.xception_model_loaded:
|
475 |
+
if st.button("π₯ Load Xception Model for Detection", type="primary"):
|
476 |
+
# Load Xception model
|
477 |
+
model, device = load_detection_model_xception()
|
478 |
if model is not None:
|
479 |
+
st.session_state.xception_model = model
|
480 |
+
st.session_state.device = device
|
481 |
+
st.session_state.xception_model_loaded = True
|
482 |
+
st.success("β
Xception model loaded successfully!")
|
483 |
else:
|
484 |
+
st.error("β Failed to load Xception model.")
|
485 |
else:
|
486 |
+
st.success("β
Xception model loaded and ready!")
|
487 |
|
488 |
with blip_col:
|
489 |
if not st.session_state.blip_model_loaded:
|
|
|
541 |
st.session_state.original_model
|
542 |
)
|
543 |
st.session_state.image_caption = caption
|
|
|
|
|
544 |
|
545 |
+
# Detect with Xception model if loaded
|
546 |
+
if st.session_state.xception_model_loaded:
|
547 |
+
with st.spinner("Analyzing image with Xception model..."):
|
548 |
+
# Preprocess image for Xception
|
549 |
+
input_tensor, original_image, face_box = preprocess_image_xception(image)
|
|
|
|
|
|
|
|
|
550 |
|
551 |
+
# Get device and model
|
552 |
+
device = st.session_state.device
|
553 |
+
model = st.session_state.xception_model
|
|
|
554 |
|
555 |
+
# Move tensor to device
|
556 |
+
input_tensor = input_tensor.to(device)
|
|
|
|
|
|
|
|
|
557 |
|
558 |
# Forward pass
|
559 |
with torch.no_grad():
|
560 |
+
logits = model(input_tensor)
|
561 |
+
probabilities = torch.softmax(logits, dim=1)[0]
|
562 |
+
pred_class = torch.argmax(probabilities).item()
|
563 |
+
confidence = probabilities[pred_class].item()
|
564 |
+
pred_label = "Fake" if pred_class == 0 else "Real" # Check class mapping
|
|
|
565 |
|
566 |
# Display results
|
567 |
with col2:
|
568 |
st.markdown("### Detection Result")
|
569 |
st.markdown(f"**Classification:** {pred_label} (Confidence: {confidence:.2%})")
|
570 |
+
|
571 |
+
# Display face box on image if detected
|
572 |
+
if face_box:
|
573 |
+
img_to_show = original_image.copy()
|
574 |
+
img_draw = np.array(img_to_show)
|
575 |
+
x, y, w, h = face_box
|
576 |
+
cv2.rectangle(img_draw, (x, y), (x + w, y + h), (0, 255, 0), 2)
|
577 |
+
st.image(Image.fromarray(img_draw), caption="Detected Face", width=300)
|
578 |
|
579 |
# GradCAM visualization
|
580 |
st.subheader("GradCAM Visualization")
|
581 |
+
cam, overlay, comparison, detected_face_box = process_image_with_xception_gradcam(
|
582 |
image, model, device, pred_class
|
583 |
)
|
584 |
|
585 |
+
if comparison:
|
586 |
+
# Display GradCAM results (controlled size)
|
587 |
+
st.image(comparison, caption="Original | CAM | Overlay", width=700)
|
588 |
+
|
589 |
+
# Save for later use
|
590 |
+
st.session_state.comparison_image = comparison
|
591 |
|
592 |
# Generate caption for GradCAM overlay image if BLIP model is loaded
|
593 |
+
if st.session_state.blip_model_loaded and overlay:
|
594 |
with st.spinner("Analyzing GradCAM visualization..."):
|
595 |
gradcam_caption = generate_gradcam_caption(
|
596 |
overlay,
|
|
|
598 |
st.session_state.finetuned_model
|
599 |
)
|
600 |
st.session_state.gradcam_caption = gradcam_caption
|
|
|
|
|
601 |
|
602 |
# Save results in session state for LLM analysis
|
603 |
st.session_state.current_image = image
|
|
|
608 |
|
609 |
st.success("β
Initial detection and GradCAM visualization complete!")
|
610 |
else:
|
611 |
+
st.warning("β οΈ Please load the Xception model first to perform initial detection.")
|
612 |
except Exception as e:
|
613 |
st.error(f"Error processing image: {str(e)}")
|
614 |
import traceback
|
|
|
745 |
st.markdown("---")
|
746 |
|
747 |
# Add model version indicator in sidebar
|
748 |
+
st.sidebar.info("Using Xception + deepfake-explainer-2 models")
|
749 |
|
750 |
if __name__ == "__main__":
|
751 |
main()
|