Custom ResNet-18 for 7-Class Classification

This is a fine-tuned ResNet-18 model designed for a 7-class classification task. The model replaces all ReLU activation functions with PReLU, introduces Dropout2D layers for better generalization, and was trained on RAF-DB DATASET with various augmentations.

📜 Model Details

Base Model: ResNet-18.
Activations: ReLU layers replaced with PReLU.
Dropout: Dropout2D applied to enhance generalization.
Classes: 7 output classes.
Input Size: Images with customizable dimensions (default: [100, 100]).
Normalization: Input images are normalized using the following statistics:
- Mean: [0.485, 0.456, 0.406]
- Std: [0.229, 0.224, 0.225]

📈 Evaluation Metrics on Test Data

Accuracy: 79.92%

Precision: 79.80%

Recall: 79.92%

F1-Score: 79.80%

Classification Report:

          precision    recall  f1-score   support

       1       0.79      0.81      0.80       329
       
       2       0.58      0.47      0.52        74
       
       3       0.51      0.42      0.46       160
       
       4       0.92      0.90      0.91      1185
       
       5       0.74      0.78      0.76       478
       
       6       0.68      0.72      0.70       162
       
       7       0.75      0.78      0.77       680
       
accuracy                           0.80      3068

macro avg       0.71      0.70      0.70      3068

weighted avg       0.80      0.80      0.80      3068

🧑‍💻 How to Use

You can load the model weights and architecture for inference or fine-tuning with the provided files:

Using PyTorch


def get_out_channels(module):
    if isinstance(module, nn.Conv2d):
        return module.out_channels
    elif isinstance(module, nn.BatchNorm2d):
        return module.num_features
    elif isinstance(module, nn.Linear):
        return module.out_features
    return None

def replace_relu_with_prelu_and_dropout(module, inplace=True):
    for name, child in module.named_children():
        replace_relu_with_prelu_and_dropout(child, inplace)
        
        if isinstance(child, nn.ReLU): 
            out_channels = None
            for prev_name, prev_child in module.named_children():
                if prev_name == name:
                    break
                out_channels = get_out_channels(prev_child) or out_channels
            
            if out_channels is None:
                raise ValueError(f"Cannot determine `out_channels` for {child}. Please check the model structure.")
            
            prelu = PReLU(device=device, num_parameters=out_channels) 
            dropout = nn.Dropout2d(p=0.2) 
            setattr(module, name, nn.Sequential(prelu, dropout).to(device))
model = models.resnet18(weights = models.ResNet18_Weights.IMAGENET1K_V1).train(True).to(device)
replace_relu_with_prelu_and_dropout(model)
# print(model.fc.in_features)


number = model.fc.in_features
module  = []

module.append(LazyLinear(7))
model.fc = Sequential(*module).to(device)

state_dict = load_file("model.safetensors")
model.load_state_dict(state_dict)
model.eval()

⚠️ Limitations and Considerations

Input Dimensions: Make sure your input images are resized to the expected dimensions (100x100) before inference.

Number of Classes: The trained model supports exactly 7 classes as defined in the training dataset.

Output: The model output should be a probability of each of the 7 face type labels. Don't forget to use the softmax function to make predictions. Note that softmax is not used in the last layer of this model's architecture.