LeNet5 (PyTorch): MNIST dataset

Another attempt to implement Yann LeCun's LeNet-5 network using PyTorch and test it in MNIST dataset.

Run experiment

1. Install torch and torchvision libs

2. Configure experiment in main.py file. Do not forget to set proper values to DATA_PATH (absolute path to data/datasets/ folder) and MODEL_STORE_PATH (absolute path to data/models/ folder) variables!

3. Run

python main.py

Description

Files

• lenet5.py - contains LeNet5 class that implements LeNet-5 network
• lenet5_trainer.py - contains LeNet5Trainer class. Method train(train_data, learning_rate, num_epochs, print_every=100) trains the network and returns it after.
• net_test.py - contains NetTest class. Method test(test_data) performs testing of the network and returns quantity of correct tests and total quantity of tests.
• utils.py - contains method prepare_mnist_data(data_path, batch_size) that prepares data for the experiment.
• main.py - main file of the experiment.

Folders

• data/datasets/ - MNIST-dataset will be downloaded here (at first run).
• data/models/ - folder to save files of trained models.

Remarks

• I'm new in Python, so be lenient ;)
• The network implementation is very rough, so be very lenient ;)