why 20 frames?
Frame Extraction Parameter Justification
Dataset Analysis
Statistical analysis of the FaceForensics++ dataset (1,000 real and 1,000 manipulated sequences) reveals the following frame count distribution:
- Minimum: 287 frames
- Median: ~480 frames
- Mean: ~530 frames
- Approximate 80th percentile range: 295–800 frames
Selected Parameter: 20 Frames per Video
After evaluating the dataset characteristics and the architectural requirements of the proposed CNN+LSTM model, a uniform extraction of 20 evenly-spaced frames per video is selected as the optimal parameter.
Justification
1. Dataset Compatibility Every video in the dataset contains a minimum of 287 frames, ensuring that a uniform extraction of 20 frames is feasible across the entire dataset without special-case handling or padding. This guarantees consistent input dimensionality for all samples.
2. Architectural Alignment The proposed model employs a CNN+LSTM architecture where the CNN extracts spatial features per frame and the LSTM models temporal dependencies across the sequence. Deepfake detection via this architecture relies primarily on spatial inconsistencies in facial regions — blending artifacts, texture anomalies, and boundary irregularities — which are present throughout the video rather than being temporally sparse. A sequence length of 20 provides sufficient temporal context for the LSTM to learn inter-frame patterns without requiring high frame density.
3. Computational Feasibility At 20 frames per video across 2,000 videos, the total extracted frame count is 40,000. This is computationally manageable for training on consumer-grade hardware. Increasing this parameter to 40 or 60 frames would double or triple memory consumption and training time with negligible accuracy gain at the current dataset scale of 2,000 videos.
4. Uniform Temporal Coverage Frames are sampled at evenly-spaced intervals across the full video duration rather than sequentially from the beginning. This approach ensures coverage of the entire temporal span of each video, accounting for cases where facial manipulation may intensify or vary throughout the sequence.
Conclusion
A value of 20 uniformly-spaced frames per video satisfies dataset constraints, aligns with the temporal modelling requirements of the CNN+LSTM architecture, and remains computationally tractable for the scope of this project.
PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM> python model_1M_params.py DeepFakeDetector( (cnn): CNNBlock( (block1): Sequential( (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): ReLU() (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) ) (block2): Sequential( (0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): ReLU() (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) ) (block3): Sequential( (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): ReLU() ) (gap): AdaptiveAvgPool2d(output_size=(1, 1)) ) (lstm): LSTM(256, 256, num_layers=2, batch_first=True, dropout=0.3) (classifier): Sequential( (0): Linear(in_features=256, out_features=256, bias=True) (1): ReLU() (2): Dropout(p=0.4, inplace=False) (3): Linear(in_features=256, out_features=128, bias=True) (4): ReLU() (5): Dropout(p=0.3, inplace=False) (6): Linear(in_features=128, out_features=64, bias=True) (7): ReLU() (8): Linear(in_features=64, out_features=2, bias=True) ) )
--- Parameter Count --- Total parameters : 1,530,562 Trainable parameters : 1,530,562
Input shape : torch.Size([2, 20, 3, 224, 224]) Output shape : torch.Size([2, 2]) Architecture check passed. PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM> python train_model.py Total videos found: 2000 Training videos: 1800 Testing videos : 200 Using device: cuda Epoch 1/1: 100%|█████████████████████████████████████████████████████████████████████████████████████| 450/450 [31:44<00:00, 4.23s/it, acc=50.06%, loss=0.694]
Epoch 1/1 Training Loss: 312.2466 Training Accuracy: 50.06%
PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM> python model_3LK_params.py DeepFakeDetector( (cnn): CNNBlock( (block1): Sequential( (0): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): ReLU() (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) ) (block2): Sequential( (0): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): ReLU() (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) ) (block3): Sequential( (0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): ReLU() ) (gap): AdaptiveAvgPool2d(output_size=(1, 1)) ) (lstm): LSTM(128, 128, num_layers=2, batch_first=True, dropout=0.3) (classifier): Sequential( (0): Linear(in_features=128, out_features=128, bias=True) (1): ReLU() (2): Dropout(p=0.4, inplace=False) (3): Linear(in_features=128, out_features=64, bias=True) (4): ReLU() (5): Dropout(p=0.3, inplace=False) (6): Linear(in_features=64, out_features=32, bias=True) (7): ReLU() (8): Linear(in_features=32, out_features=2, bias=True) ) )
--- Parameter Count ---
--- Parameter Count --- Total parameters : 384,354 Trainable parameters : 384,354
Input shape : torch.Size([2, 20, 3, 224, 224]) Output shape : torch.Size([2, 2]) Architecture check passed. Total parameters : 384,354 Trainable parameters : 384,354
Input shape : torch.Size([2, 20, 3, 224, 224]) Output shape : torch.Size([2, 2]) Architecture check passed. PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM> python train_model.py Output shape : torch.Size([2, 2]) Architecture check passed. PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM> python train_model.py Total videos found: 2000 PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM> python train_model.py Total videos found: 2000 Training videos: 1800 Total videos found: 2000 Training videos: 1800 Testing videos : 200 Using device: cuda Training videos: 1800 Testing videos : 200 Using device: cuda Epoch 1/15: 100%|███████████████████████████████████████████████████████████████████| 450/450 [02:36<00:00, 2.88it/s, acc=49.78%, loss=0.703] Testing videos : 200 Using device: cuda Epoch 1/15: 100%|███████████████████████████████████████████████████████████████████| 450/450 [02:36<00:00, 2.88it/s, acc=49.78%, loss=0.703]
Epoch 1/15 Epoch 1/15: 100%|███████████████████████████████████████████████████████████████████| 450/450 [02:36<00:00, 2.88it/s, acc=49.78%, loss=0.703]
Epoch 1/15 Training Loss: 313.7062
Epoch 1/15 Training Loss: 313.7062 Training Loss: 313.7062 Training Accuracy: 49.78%
PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM> python model.py
C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\timm\models_factory.py:138: UserWarning: Mapping deprecated model name xception to current legacy_xception.
model = create_fn(
DeepFakeDetector(
(cnn): CNNBlock(
(xception): Xception(
(conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(2, 2), bias=False)
(bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act1): ReLU(inplace=True)
(conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), bias=False)
(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act2): ReLU(inplace=True)
(block1): Block(
(skip): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
(skipbn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(rep): Sequential(
(0): SeparableConv2d(
(conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=64, bias=False)
(pointwise): Conv2d(64, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
(3): SeparableConv2d(
(conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=128, bias=False)
(pointwise): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(4): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
)
)
(block2): Block(
(skip): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
(skipbn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=128, bias=False)
(pointwise): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256, bias=False)
(pointwise): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
)
)
(block3): Block(
(skip): Conv2d(256, 728, kernel_size=(1, 1), stride=(2, 2), bias=False)
(skipbn): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256, bias=False)
(pointwise): Conv2d(256, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
)
)
(block4): Block(
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): ReLU(inplace=True)
(7): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(block5): Block(
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): ReLU(inplace=True)
(7): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(block6): Block(
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): ReLU(inplace=True)
(7): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(block7): Block(
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): ReLU(inplace=True)
(7): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(block8): Block(
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): ReLU(inplace=True)
(7): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(block9): Block(
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): ReLU(inplace=True)
(7): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(block10): Block(
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): ReLU(inplace=True)
(7): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(block11): Block(
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): ReLU(inplace=True)
(7): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(block12): Block(
(skip): Conv2d(728, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False)
(skipbn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(rep): Sequential(
(0): ReLU()
(1): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): ReLU(inplace=True)
(4): SeparableConv2d(
(conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False)
(pointwise): Conv2d(728, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(5): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
)
)
(conv3): SeparableConv2d(
(conv1): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=1024, bias=False)
(pointwise): Conv2d(1024, 1536, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(bn3): BatchNorm2d(1536, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act3): ReLU(inplace=True)
(conv4): SeparableConv2d(
(conv1): Conv2d(1536, 1536, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=1536, bias=False)
(pointwise): Conv2d(1536, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)
)
(bn4): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act4): ReLU(inplace=True)
(global_pool): SelectAdaptivePool2d(pool_type=avg, flatten=Flatten(start_dim=1, end_dim=-1))
(fc): Linear(in_features=2048, out_features=1000, bias=True)
)
(gap): AdaptiveAvgPool2d(output_size=(1, 1))
(fc): Linear(in_features=2048, out_features=256, bias=True)
)
(lstm): LSTM(256, 256, num_layers=2, batch_first=True, dropout=0.3)
(classifier): Sequential(
(0): Linear(in_features=256, out_features=256, bias=True)
(1): ReLU()
(2): Dropout(p=0.4, inplace=False)
(3): Linear(in_features=256, out_features=128, bias=True)
(4): ReLU()
(5): Dropout(p=0.3, inplace=False)
(6): Linear(in_features=128, out_features=64, bias=True)
(7): ReLU()
(8): Linear(in_features=64, out_features=2, bias=True)
)
)
--- Parameter Count --- Total parameters : 24,540,242 Trainable parameters : 1,684,290
Input shape : torch.Size([2, 20, 3, 224, 224]) Output shape : torch.Size([2, 2]) Architecture check passed. PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM> python train_model.py Total videos found: 2000 Training videos: 1800 Testing videos : 200 Using device: cuda C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\timm\models_factory.py:138: UserWarning: Mapping deprecated model name xception to current legacy_xception. model = create_fn( Epoch 1/15: 100%|███████████████████████████████████████████████████████████████████| 450/450 [04:12<00:00, 1.78it/s, acc=53.17%, loss=0.637]
Epoch 1/15 Training Loss: 311.3695 Training Accuracy: 53.17%
PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM> python model.py C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\timm\models_factory.py:138: UserWarning: Mapping deprecated model name xception to current legacy_xception. model = create_fn( DeepFakeDetector( (cnn): CNNBlock( (xception): Xception( (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(2, 2), bias=False) (bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (act1): ReLU(inplace=True) (conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), bias=False) (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (act2): ReLU(inplace=True) (block1): Block( (skip): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False) (skipbn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (rep): Sequential( (0): SeparableConv2d( (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=64, bias=False) (pointwise): Conv2d(64, 128, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): ReLU(inplace=True) (3): SeparableConv2d( (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=128, bias=False) (pointwise): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (4): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (5): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) ) ) (block2): Block( (skip): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False) (skipbn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=128, bias=False) (pointwise): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256, bias=False) (pointwise): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) ) ) (block3): Block( (skip): Conv2d(256, 728, kernel_size=(1, 1), stride=(2, 2), bias=False) (skipbn): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256, bias=False) (pointwise): Conv2d(256, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) ) ) (block4): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block5): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block6): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block7): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block8): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block9): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block10): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block11): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block12): Block( (skip): Conv2d(728, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False) (skipbn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) ) ) (conv3): SeparableConv2d( (conv1): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=1024, bias=False) (pointwise): Conv2d(1024, 1536, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (bn3): BatchNorm2d(1536, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (act3): ReLU(inplace=True) (conv4): SeparableConv2d( (conv1): Conv2d(1536, 1536, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=1536, bias=False) (pointwise): Conv2d(1536, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (bn4): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (act4): ReLU(inplace=True) (global_pool): SelectAdaptivePool2d(pool_type=avg, flatten=Flatten(start_dim=1, end_dim=-1)) (fc): Linear(in_features=2048, out_features=1000, bias=True) ) (gap): AdaptiveAvgPool2d(output_size=(1, 1)) (fc): Linear(in_features=2048, out_features=256, bias=True) ) (lstm): LSTM(256, 256, num_layers=2, batch_first=True, dropout=0.3) (classifier): Sequential( (0): Linear(in_features=256, out_features=256, bias=True) (1): ReLU() (2): Dropout(p=0.4, inplace=False) (3): Linear(in_features=256, out_features=128, bias=True) (4): ReLU() (5): Dropout(p=0.3, inplace=False) (6): Linear(in_features=128, out_features=64, bias=True) (7): ReLU() (8): Linear(in_features=64, out_features=2, bias=True) ) )
--- Parameter Count --- Total parameters : 24,540,242 Trainable parameters : 1,684,290
Input shape : torch.Size([2, 20, 3, 224, 224]) Output shape : torch.Size([2, 2]) Architecture check passed. PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM> python train_model.py Total videos found: 2000 Training videos: 1800 Testing videos : 200 Using device: cuda C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\timm\models_factory.py:138: UserWarning: Mapping deprecated model name xception to current legacy_xception. model = create_fn( Epoch 1/10: 100%|████████████████████████████████████████████████████████████████████████████████████| 450/450 [04:11<00:00, 1.79it/s, acc=48.72%, loss=0.691]
Epoch 1/10 Training Loss: 312.4615 Training Accuracy: 48.72% Model saved to saved_models/deepfake_model_epoch_10.pth Epoch 2/10: 100%|████████████████████████████████████████████████████████████████████████████████████| 450/450 [04:10<00:00, 1.80it/s, acc=48.89%, loss=0.693]
Epoch 2/10 Training Loss: 312.3502 Training Accuracy: 48.89% Model saved to saved_models/deepfake_model_epoch_10.pth Epoch 3/10: 100%|████████████████████████████████████████████████████████████████████████████████████| 450/450 [04:11<00:00, 1.79it/s, acc=49.17%, loss=0.692]
Epoch 3/10 Training Loss: 312.2474 Training Accuracy: 49.17% Model saved to saved_models/deepfake_model_epoch_10.pth Epoch 4/10: 100%|████████████████████████████████████████████████████████████████████████████████████| 450/450 [04:10<00:00, 1.80it/s, acc=48.67%, loss=0.691]
Epoch 4/10 Training Loss: 312.0824 Training Accuracy: 48.67% Model saved to saved_models/deepfake_model_epoch_10.pth Epoch 5/10: 3%|██▍ | 13/450 [00:22<12:44, 1.75s/it, acc=36.54%, loss=0.701] Traceback (most recent call last): File "C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM\train_model.py", line 265, in main() ~~~~^^ File "C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM\train_model.py", line 193, in main running_loss += loss.item() ~~~~~~~~~^^ KeyboardInterrupt PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN+LSTM>
XceptioNet+LSTM(freeze all) terminal :
PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN_LSTM> python model.py C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\timm\models_factory.py:138: UserWarning: Mapping deprecated model name xception to current legacy_xception. model = create_fn( DeepFakeDetector( (cnn): CNNBlock( (xception): Xception( (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(2, 2), bias=False) (bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (act1): ReLU(inplace=True) (conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), bias=False) (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (act2): ReLU(inplace=True) (block1): Block( (skip): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False) (skipbn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (rep): Sequential( (0): SeparableConv2d( (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=64, bias=False) (pointwise): Conv2d(64, 128, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): ReLU(inplace=True) (3): SeparableConv2d( (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=128, bias=False) (pointwise): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (4): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (5): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) ) ) (block2): Block( (skip): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False) (skipbn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=128, bias=False) (pointwise): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256, bias=False) (pointwise): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) ) ) (block3): Block( (skip): Conv2d(256, 728, kernel_size=(1, 1), stride=(2, 2), bias=False) (skipbn): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256, bias=False) (pointwise): Conv2d(256, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) ) ) (block4): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block5): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block6): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block7): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block8): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block9): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block10): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block11): Block( (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace=True) (7): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (8): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (block12): Block( (skip): Conv2d(728, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False) (skipbn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (rep): Sequential( (0): ReLU() (1): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 728, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (2): BatchNorm2d(728, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (3): ReLU(inplace=True) (4): SeparableConv2d( (conv1): Conv2d(728, 728, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=728, bias=False) (pointwise): Conv2d(728, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (5): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) ) ) (conv3): SeparableConv2d( (conv1): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=1024, bias=False) (pointwise): Conv2d(1024, 1536, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (bn3): BatchNorm2d(1536, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (act3): ReLU(inplace=True) (conv4): SeparableConv2d( (conv1): Conv2d(1536, 1536, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=1536, bias=False) (pointwise): Conv2d(1536, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False) ) (bn4): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (act4): ReLU(inplace=True) (global_pool): SelectAdaptivePool2d(pool_type=avg, flatten=Flatten(start_dim=1, end_dim=-1)) (fc): Linear(in_features=2048, out_features=1000, bias=True) ) (gap): AdaptiveAvgPool2d(output_size=(1, 1)) (fc): Linear(in_features=2048, out_features=256, bias=True) ) (lstm): LSTM(256, 256, num_layers=2, batch_first=True, dropout=0.3) (classifier): Sequential( (0): Linear(in_features=256, out_features=256, bias=True) (1): ReLU() (2): Dropout(p=0.4, inplace=False) (3): Linear(in_features=256, out_features=128, bias=True) (4): ReLU() (5): Dropout(p=0.3, inplace=False) (6): Linear(in_features=128, out_features=64, bias=True) (7): ReLU() (8): Linear(in_features=64, out_features=2, bias=True) ) )
--- Parameter Count --- Total parameters : 24,540,242 Trainable parameters : 1,684,290
Input shape : torch.Size([2, 20, 3, 224, 224]) Output shape : torch.Size([2, 2]) Architecture check passed. PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN_LSTM> python train_model.py
=== Path Verification === Train/Test Root: ✓ Found → E:\dataset\Deepfake Detection Dataset\3. frames_dataset real: 1000 video folders fake: 1000 video folders Validation Root: ✓ Found → E:\dataset\Deepfake Detection Dataset\5. Validation_Dataset_Frames real: 500 video folders fake: 500 video folders
=== Building Splits === Train : 1800 videos Test : 200 videos Val : 1000 videos Train → real: 900, fake: 900 Test → real: 100, fake: 100 Val → real: 500, fake: 500
Train batches : 450 Test batches : 50 Val batches : 250
Using device : cuda GPU : NVIDIA GeForce RTX 2050 C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\timm\models_factory.py:138: UserWarning: Mapping deprecated model name xception to current legacy_xception. Test batches : 50 Val batches : 250
Using device : cuda GPU : NVIDIA GeForce RTX 2050 C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\timm\models_factory.py:138: UserWarning: Mapping deprecated model name xception to current legacy_Test batches : 50 Val batches : 250
Using device : cuda GPU : NVIDIA GeForce RTX 2050 Test batches : 50 Val batches : 250
Using device : cuda Test batches : 50 Val batches : 250
Using device : cuda Test batches : 50 Val batches : 250
Using device : cuda Test batches : 50 Val batches : 250 Test batches : 50 Val batches : 250 Test batches : 50 Val batches : 250
Test batches : 50 Val batches : 250
Test batches : 50 Test batches : 50 Val batches : 250 Test batches : 50 Val batches : 250 Test batches : 50 Val batches : 250 Test batches : 50 Val batches : 250
Using device : cuda GPU : NVIDIA GeForce RTX 2050 C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\timm\models_factory.py:138: UserWarning: Mapping deprecated model name xception to current legacy_xception. model = create_fn( Trainable params : 1,684,290 / 24,540,242
=== Training ===
Epoch 1/20 Train → Loss: 0.6934 Acc: 50.78% Val → Loss: 0.6985 Acc: 50.00% ✓ Best model saved (val_loss: 0.6985)
Epoch 2/20 Train → Loss: 0.6943 Acc: 50.33% Val → Loss: 0.6933 Acc: 49.70% ✓ Best model saved (val_loss: 0.6933)
Epoch 3/20 Train → Loss: 0.6937 Acc: 47.61% Val → Loss: 0.6930 Acc: 50.00% No improvement. Patience: 1/5
Epoch 4/20 Train → Loss: 0.6936 Acc: 49.78% Val → Loss: 0.6931 Acc: 50.00% No improvement. Patience: 2/5
Epoch 5/20 Train → Loss: 0.6935 Acc: 50.39% Val → Loss: 0.6931 Acc: 50.00% No improvement. Patience: 3/5
Epoch 6/20 Train → Loss: 0.6938 Acc: 50.17% Val → Loss: 0.6932 Acc: 50.00% No improvement. Patience: 4/5
Epoch 7/20 Train → Loss: 0.6936 Acc: 49.11% Val → Loss: 0.6931 Acc: 50.00% No improvement. Patience: 5/5
⚡ Early stopping triggered at epoch 7
✓ Training complete — best val loss: 0.6933
=== Loading Best Model === ✓ Loaded from epoch 2 (val_loss: 0.6933)
=== Evaluation === Evaluating Test Set: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████| 50/50 [00:42<00:00, 1.19it/s] Evaluating Val Set: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 250/250 [01:49<00:00, 2.29it/s]
================================================== TEST SET — Classification Report
C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\sklearn\metrics_classification.py:1833: UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, f"{metric.capitalize()} is", result.shape[0])
C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\sklearn\metrics_classification.py:1833: UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, f"{metric.capitalize()} is", result.shape[0])
C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\sklearn\metrics_classification.py:1833: UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, f"{metric.capitalize()} is", result.shape[0])
precision recall f1-score support
Real 0.50 1.00 0.67 100
Fake 0.00 0.00 0.00 100
accuracy 0.50 200
macro avg 0.25 0.50 0.33 200 weighted avg 0.25 0.50 0.33 200
================================================== VALIDATION SET — Classification Report
precision recall f1-score support
Real 0.50 0.99 0.66 500
Fake 0.00 0.00 0.00 500
accuracy 0.50 1000
macro avg 0.25 0.50 0.33 1000 weighted avg 0.25 0.50 0.33 1000
=== Saving Plots === Saved → saved_models\training_curves.png Saved → saved_models\Confusion_Matrix_-Test_Set.png Saved → saved_models\Confusion_Matrix-Val_Set.png Saved → saved_models\ROC_Curve-Test_Set.png Saved → saved_models\ROC_Curve-_Val_Set.png
Test AUC : 0.4625 Val AUC : 0.4310
✓ All plots saved to saved_models
XceptionNet+LSTM (unfreeze some layer) terminals:
PS C:\Users\ASUS\Desktop\Mini Project\Deepfake-video-Detection\ML model\CNN_LSTM> python train_model.py
=== Path Verification === Train/Test Root: ✓ Found → E:\dataset\Deepfake Detection Dataset\3. frames_dataset real: 1000 video folders fake: 1000 video folders Validation Root: ✓ Found → E:\dataset\Deepfake Detection Dataset\5. Validation_Dataset_Frames real: 500 video folders fake: 500 video folders
=== Building Splits === Train : 1800 videos Test : 200 videos Val : 1000 videos Train → real: 900, fake: 900 Test → real: 100, fake: 100 Val → real: 500, fake: 500
Train batches : 450 Test batches : 50 Val batches : 250
Using device : cuda GPU : NVIDIA GeForce RTX 2050 C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\timm\models_factory.py:138: UserWarning: Mapping deprecated model name xception to current legacy_xception. model = create_fn( Xception — frozen: 94, unfrozen: 62 / 156 params groups Trainable params : 15,902,322 / 24,540,242
=== Training ===
Epoch 1/20 Train → Loss: 0.6939 Acc: 49.89% Val → Loss: 0.6935 Acc: 50.00% ✓ Best model saved (val_loss: 0.6935)
Epoch 2/20 Train → Loss: 0.6938 Acc: 49.22% Val → Loss: 0.6927 Acc: 50.00% No improvement. Patience: 1/5
Epoch 3/20 Train → Loss: 0.6934 Acc: 49.83% Val → Loss: 0.6935 Acc: 50.00% No improvement. Patience: 2/5
Epoch 4/20 Train → Loss: 0.6933 Acc: 51.28% Val → Loss: 0.6938 Acc: 50.00% No improvement. Patience: 3/5
Epoch 5/20 Train → Loss: 0.6938 Acc: 50.22% Val → Loss: 0.6938 Acc: 50.00% No improvement. Patience: 4/5
Epoch 6/20 Train → Loss: 0.6938 Acc: 48.72% Val → Loss: 0.6929 Acc: 50.70% No improvement. Patience: 5/5
⚡ Early stopping triggered at epoch 6
✓ Training complete — best val loss: 0.6935
=== Loading Best Model === ✓ Loaded from epoch 1 (val_loss: 0.6935)
=== Evaluation === Evaluating Test Set: 100%|██████████████████████████████████████████████████████████████████████████████████████████████| 50/50 [00:51<00:00, 1.04s/it] Evaluating Val Set: 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 250/250 [01:49<00:00, 2.29it/s]
================================================== TEST SET — Classification Report
C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\sklearn\metrics_classification.py:1833: UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, f"{metric.capitalize()} is", result.shape[0])
C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\sklearn\metrics_classification.py:1833: UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, f"{metric.capitalize()} is", result.shape[0])
C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\sklearn\metrics_classification.py:1833: UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, f"{metric.capitalize()} is", result.shape[0])
precision recall f1-score support
Real 0.50 1.00 0.67 100
Fake 0.00 0.00 0.00 100
accuracy 0.50 200
macro avg 0.25 0.50 0.33 200 weighted avg 0.25 0.50 0.33 200
================================================== VALIDATION SET — Classification Report
C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\sklearn\metrics_classification.py:1833: UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, f"{metric.capitalize()} is", result.shape[0])
C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\sklearn\metrics_classification.py:1833: UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, f"{metric.capitalize()} is", result.shape[0])
C:\Users\ASUS\AppData\Roaming\Python\Python313\site-packages\sklearn\metrics_classification.py:1833: UndefinedMetricWarning: Precision is ill-defined and being set to 0.0 in labels with no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, f"{metric.capitalize()} is", result.shape[0])
precision recall f1-score support
Real 0.50 1.00 0.67 500
Fake 0.00 0.00 0.00 500
accuracy 0.50 1000
macro avg 0.25 0.50 0.33 1000 weighted avg 0.25 0.50 0.33 1000
=== Saving Plots === Saved → saved_models\training_curves.png Saved → saved_models\Confusion_Matrix_-Test_Set.png Saved → saved_models\Confusion_Matrix-Val_Set.png Saved → saved_models\ROC_Curve-Test_Set.png Saved → saved_models\ROC_Curve-_Val_Set.png
Test AUC : 0.5139 Val AUC : 0.3664
✓ All plots saved to saved_models
EfficientNet-B1 training terminal :-
=== Path Verification === Train/Test Root: ✓ Found → E:\dataset\Deepfake Detection Dataset\3. frames_dataset real: 1000 video folders fake: 1000 video folders Validation Root: ✓ Found → E:\dataset\Deepfake Detection Dataset\5. Validation_Dataset_Frames real: 500 video folders fake: 500 video folders
=== Building Splits === Train : 1800 videos Test : 200 videos Val : 1000 videos Train → real: 900, fake: 900 Test → real: 100, fake: 100 Val → real: 500, fake: 500
Train batches : 450 Test batches : 50 Val batches : 250
Using device : cuda GPU : NVIDIA GeForce RTX 2050 EfficientNet-B1 — frozen: 180, unfrozen: 119 / 299 Test → real: 100, fake: 100 Val → real: 500, fake: 500
Train batches : 450 Test batches : 50 Val batches : 250
Using device : cuda GPU : NVIDIA GeForce RTX 2050 EfficientNet-B1 — frozen: 180, unfrozen: 119 / 299
Train batches : 450 Test batches : 50 Val batches : 250
Using device : cuda GPU : NVIDIA GeForce RTX 2050 EfficientNet-B1 — frozen: 180, unfrozen: 119 / 299 Test batches : 50 Val batches : 250
Using device : cuda GPU : NVIDIA GeForce RTX 2050 EfficientNet-B1 — frozen: 180, unfrozen: 119 / 299
Using device : cuda GPU : NVIDIA GeForce RTX 2050 EfficientNet-B1 — frozen: 180, unfrozen: 119 / 299 Using device : cuda GPU : NVIDIA GeForce RTX 2050 EfficientNet-B1 — frozen: 180, unfrozen: 119 / 299 GPU : NVIDIA GeForce RTX 2050 EfficientNet-B1 — frozen: 180, unfrozen: 119 / 299 EfficientNet-B1 — frozen: 180, unfrozen: 119 / 299 Trainable params : 7,136,854 / 8,000,866
=== Training ===
Epoch 1/20 Train → Loss: 0.6965 Acc: 50.00% Val → Loss: 0.6962 Acc: 50.00% ✓ Best model saved (val_loss: 0.6962)
Epoch 2/20 Train → Loss: 0.6949 Acc: 49.94% Val → Loss: 0.6875 Acc: 50.00% ✓ Best model saved (val_loss: 0.6875)
Epoch 3/20 Train → Loss: 0.6622 Acc: 64.33% Val → Loss: 0.5967 Acc: 77.60% ✓ Best model saved (val_loss: 0.5967)
Epoch 4/20 Train → Loss: 0.6276 Acc: 69.94% Val → Loss: 0.5649 Acc: 79.30% ✓ Best model saved (val_loss: 0.5649)
Epoch 5/20 Train → Loss: 0.6048 Acc: 72.11% Val → Loss: 0.5258 Acc: 78.30% ✓ Best model saved (val_loss: 0.5258)
Epoch 6/20 Train → Loss: 0.5571 Acc: 74.06% Val → Loss: 0.4761 Acc: 79.30% ✓ Best model saved (val_loss: 0.4761)
Epoch 7/20 Train → Loss: 0.5250 Acc: 76.17% Val → Loss: 0.5161 Acc: 76.20% No improvement. Patience: 1/5
Epoch 8/20 Train → Loss: 0.5162 Acc: 75.39% Val → Loss: 0.4988 Acc: 76.00% No improvement. Patience: 2/5
Epoch 9/20 Train → Loss: 0.4903 Acc: 78.56% Val → Loss: 0.4555 Acc: 79.50% ✓ Best model saved (val_loss: 0.4555)
Epoch 10/20 Train → Loss: 0.4533 Acc: 79.56% Val → Loss: 0.5358 Acc: 73.60% No improvement. Patience: 1/5
Epoch 11/20 Train → Loss: 0.4353 Acc: 81.56% Val → Loss: 0.5443 Acc: 72.40% No improvement. Patience: 2/5
Epoch 12/20 Train → Loss: 0.3911 Acc: 83.78% Val → Loss: 0.4415 Acc: 77.60% ✓ Best model saved (val_loss: 0.4415)
Epoch 13/20 Train → Loss: 0.3830 Acc: 83.17% Val → Loss: 0.5870 Acc: 69.40% No improvement. Patience: 1/5
Epoch 14/20 Train → Loss: 0.3429 Acc: 86.22% Val → Loss: 0.7223 Acc: 64.20% No improvement. Patience: 2/5
Epoch 15/20 Train → Loss: 0.3190 Acc: 86.83% Val → Loss: 0.6290 Acc: 68.20% No improvement. Patience: 3/5
Epoch 16/20 Train → Loss: 0.3195 Acc: 86.83% Val → Loss: 0.5972 Acc: 71.40% No improvement. Patience: 4/5
Epoch 17/20 Train → Loss: 0.2944 Acc: 87.50% Val → Loss: 0.6218 Acc: 68.70% No improvement. Patience: 5/5
⚡ Early stopping triggered at epoch 17
✓ Training complete — best val loss: 0.4415
=== Loading Best Model === ✓ Loaded from epoch 12 (val_loss: 0.4415)
=== Evaluation === Evaluating Test Set: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████| 50/50 [00:36<00:00, 1.36it/s] Evaluating Val Set: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 250/250 [01:17<00:00, 3.22it/s]
================================================== TEST SET — Classification Report
precision recall f1-score support
Real 0.59 0.68 0.63 100
Fake 0.62 0.53 0.57 100
accuracy 0.60 200
macro avg 0.61 0.60 0.60 200 weighted avg 0.61 0.60 0.60 200
================================================== VALIDATION SET — Classification Report
precision recall f1-score support
Real 0.73 0.88 0.80 500
Fake 0.85 0.67 0.75 500
accuracy 0.78 1000
macro avg 0.79 0.78 0.77 1000 weighted avg 0.79 0.78 0.77 1000
=== Saving Plots === Saved → saved_models\training_curves.png Saved → saved_models\Confusion_Matrix_-Test_Set.png Saved → saved_models\Confusion_Matrix-Val_Set.png Saved → saved_models\ROC_Curve-Test_Set.png Saved → saved_models\ROC_Curve-_Val_Set.png
Test AUC : 0.6472 Val AUC : 0.8974
✓ All plots saved to saved_models
🔍 So What Did the Original Author Do?
There are only 3 possible ways they handled videos:
🔹 1. Frame-wise Prediction + Aggregation (MOST COMMON ✅)
Pipeline:
Video → Frames → CNN → Predictions per frame → Aggregate → Final output Example: Frame1 → 0.91 (fake) Frame2 → 0.85 (fake) Frame3 → 0.40 (real) ...
Then:
Average Majority vote Max confidence
👉 Final decision = video-level prediction
🔹 2. Frame Sampling (Single Frame) Video → pick 1 frame → CNN → output
❌ Less accurate ✔ Simple
🔹 3. Feature Aggregation (Advanced) Frames → CNN features → LSTM/Pooling → Classifier
👉 This is what YOU did with:
EfficientNet + LSTM ✅