Online App • Python API • Demo Video

The MNIST LeNet-5 Classifier project implements the LeNet-5 model proposed by Yann LeCun in 1998 to classify handwritten digits from the MNIST dataset.

How it Works

This project is an implementation of the LeNet-5 model, a convolutional neural network (CNN) architecture first proposed by Yann LeCun in 1998 for handwritten character recognition. LeNet-5 was a pioneering architecture in the field of deep learning and is highly effective for image classification tasks such as digit recognition on the MNIST dataset.

The project uses PyTorch for training and testing the model with data augmentation techniques, including random cropping and horizontal flipping, to improve generalization. Real-time training performance is tracked using TensorBoard, and the model checkpoints are saved based on the highest accuracy achieved during training.

Features

• LeNet-5 architecture: A classic CNN model designed for digit classification.
• Data Augmentation: Random cropping and flipping for improving model generalization.
• Training and Testing: Trains on the MNIST dataset and evaluates performance on a separate test dataset.
• TensorBoard Support: Visualize training and testing metrics (loss and accuracy) in real-time.
• Model Checkpointing: Automatically saves the best-performing model based on test accuracy.
• Learning Rate Scheduler: Adapts the learning rate for optimal training efficiency.

Run app locally

To run this project locally, follow these steps:

1. Clone the repository:

   git clone https://github.com/chloelavrat/torchLeNet-5
   cd torchLeNet-5

2. Install the required dependencies:

   pip install -r requirements.txt

3. Run the training script:

   python lenet_5/train.py

4. Monitor training with TensorBoard: Open TensorBoard to view the loss and accuracy plots:

   tensorboard --logdir=runs

Python

The Python API allows you to train, test, and save the LeNet-5 model. Here’s an overview of the key components:

• LeNet Model: The neural network model defined in model.py, following the architecture of the original LeNet-5.
• Train Function: Manages the training loop, logging loss and accuracy during each epoch.
• Test Function: Evaluates the model on the test dataset and logs the test performance.
• TensorBoard Integration: Training metrics are logged for visualization via TensorBoard.
• Model Saving: The best-performing model is saved automatically based on test accuracy.

Training

In this demo, I trained the LeNet-5 model on the MNIST dataset for 15 epochs with a batch size of 64. The learning rate was initially set to 0.001 and decayed by a factor of 0.1 every 5 epochs.

Contributing

This project is open-source, and contributions are always welcome! You can contribute by submitting a pull request or creating an issue on the GitHub repository.

License

This project is licensed under the MIT License. See the LICENSE file for more details.