soutrik's picture
added: model and code and app
29730dd
---
title: Erav2s13
emoji: πŸ”₯
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 4.27.0
app_file: app.py
pinned: false
license: mit
---
# Erav2s13- SOUTRIK πŸ”₯
## Overview
This repository leverages the Hugging Face repository and Gradio for building a user interface (UI). The model training was conducted using Google Colab, and the resulting model files are utilized for inference in the Gradio app.
- **Model Training**: `Main.ipynb` - Colab notebook used to build and train the model.
- **Inference**: The same model structure and files are used in the Gradio app.
## Custom ResNet Model
The `custom_resnet.py` file defines a custom ResNet (Residual Network) model using PyTorch Lightning. This model is specifically designed for image classification tasks, particularly for the CIFAR-10 dataset.
### Model Architecture
The custom ResNet model comprises the following components:
1. **Preparation Layer**: Convolutional layer with 64 filters, followed by batch normalization, ReLU activation, and dropout.
2. **Layer 1**: Convolutional layer with 128 filters, max pooling, batch normalization, ReLU activation, and dropout. Includes a residual block with two convolutional layers (128 filters each), batch normalization, ReLU activation, and dropout.
3. **Layer 2**: Convolutional layer with 256 filters, max pooling, batch normalization, ReLU activation, and dropout.
4. **Layer 3**: Convolutional layer with 512 filters, max pooling, batch normalization, ReLU activation, and dropout. Includes a residual block with two convolutional layers (512 filters each), batch normalization, ReLU activation, and dropout.
5. **Max Pooling**: Max pooling layer with a kernel size of 4.
6. **Fully Connected Layer**: Flattened output passed through a fully connected layer with 10 output units (for CIFAR-10 classes).
7. **Softmax**: Log softmax activation function to obtain predicted class probabilities.
### Training and Evaluation
The model is trained using PyTorch Lightning, which provides a high-level interface for training, validation, and testing. Key components include:
- **Optimizer**: Adam with a learning rate specified by `PREFERRED_START_LR`.
- **Scheduler**: OneCycleLR for learning rate adjustment.
- **Loss and Accuracy**: Cross-entropy loss and accuracy are computed and logged during training, validation, and testing.
### Misclassified Images
During testing, misclassified images are tracked and stored in a dictionary along with their ground truth and predicted labels, facilitating error analysis and model improvement.
### Hyperparameters
Key hyperparameters include:
- `PREFERRED_START_LR`: Initial learning rate.
- `PREFERRED_WEIGHT_DECAY`: Weight decay for regularization.
### Model Summary
The `detailed_model_summary` function prints a comprehensive summary of the model architecture, detailing input size, kernel size, output size, number of parameters, and trainable status of each layer.
## Lightning Dataset Module
The `lightning_dataset.py` file contains the `CIFARDataModule` class, which is a PyTorch Lightning `LightningDataModule` for the CIFAR-10 dataset. This class handles data preparation, splitting, and loading.
### CIFARDataModule Class
#### Parameters
- `data_path`: Directory path for CIFAR-10 dataset.
- `batch_size`: Batch size for data loaders.
- `seed`: Random seed for reproducibility.
- `val_split`: Fraction of training data used for validation (default: 0).
- `num_workers`: Number of worker processes for data loading (default: 0).
#### Methods
- `prepare_data`: Downloads CIFAR-10 dataset if not present.
- `setup`: Defines data transformations and creates training, validation, and testing datasets.
- `train_dataloader`: Returns training data loader.
- `val_dataloader`: Returns validation data loader.
- `test_dataloader`: Returns testing data loader.
#### Utility Methods
- `_split_train_val`: Splits training dataset into training and validation subsets.
- `_init_fn`: Initializes random seed for each worker process to ensure reproducibility.
## License
This project is licensed under the MIT License.
---