Handwritten Character Recognition

Simple MNIST trained handwriting recognition model using CNN with webpage playground. The network consists of two convolutional layers followed by two fully-connected layers for classification with softmax. It uses the negative log likelihood loss. It is a quite simple architecture. I opted for a modern pytorch lightning implementation so this architecture is easily adaptable. Furthermore I added a simple webpage with a flask backend to interactively check the model. It is therefore easy to interactively experiment with differing model architectures.

Run main.py to train model which is then checkpointed

Run test.py to see example predictions on randomly drawn digits on trained model (set version to match the version you want to load).

Run flask --app server run in the web directory to launch an interactive website that allows to draw a digit that gets predicted by the model (set version to match the version you want to load).

Example