This project trains a multilayer perceptron on the MNIST dataset to classify handwritten numbers. The code uses the torch library to take use of CUDA to accelerate the learning process. Besides the use of the MNIST dataloader from the torchvision library everything is programmed from scratch.
The MLP network consists of two hidden layers. Both hidden layer use the sigmoid activation function. The first hidden layer consists of 32 neurons, the second hidden layer of 16 neurons. For the output layer the softmax activation is applied. The mean square error is used to calculate the loss.
(784, Input) => (32, Sigmoid) => (16, Sigmoid) => (10, Softmax) => Mean Square Error
To train the network execute the function train() in train.py. After the training the model will be saved to models/model.pt. The model can then be evaluated by running the code inside model_evaluation.ipynb.
