| # 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 a custom dataset with various augmentations. | |
| --- | |
| ## 📜 Model Details | |
| - **Base Model:** ResNet-18 (pre-trained on ImageNet). | |
| - **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 | |
|  | |
| ## 🧑💻 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): | |
| """تابعی برای یافتن تعداد کانالهای خروجی از لایههای کانولوشن و BatchNorm""" | |
| 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. | |