Description of Simple Feed-Forward Fully Conntect Network

This implementation is based on the paper Lottery Ticket Hypothesis by Jonathan Frankle and Michael Carbin. It doesn't implement the pruning but only the training of the Fully-Connected Network as described in Figure 2 under LeNet-300-100.

The parameters are set to the default and achieve a similar validation accuracy of ~98% after 50 epochs.

Usage

Make sure to install the requirements.txt file.

pip install -r requirements.txt

To train the network run start the jupyter server and run the notebook.

jupyter notebook

If you're running this on your personal computer and do not have access to a GPU, it might take a while to train the network. You can reduce the number of epochs to speed up the training process. Otherwise, feel free to deploy the notebook on Google Colab with a T4 machine.