MyTorch

A simplified version of PyTorch for constructing and training neural networks. This project was inspired and builds upon Andrej Karpathy's micrograd package to allow for multidimensional Tensors. It also providing functionality for neural network construction, data handling, and optimization.

The components of MyTorch are:

• engine.py: defines the Tensor class, the foundation of MyTorch
• nn.py: implementation of Module base class for MyTorch layers, and a few useful subclasses
• data.py: tools for processing data for use with neural networks
• optim.py: tools for neural network optimization

The notebook experiments.ipynb contains a demonstration of MyTorch in action: we train a neural network to classify MNIST images and compare its performance to an analogous network developed in PyTorch.

Example Usage

Constructing and training an MNIST image classifier:

from mytorch import nn, data, optim

net = nn.Sequential(
    nn.Flatten(),
    nn.Linear(28 * 28, 64),
    nn.ReLU(),
    nn.Linear(64, 10)
)

dataset = data.MNIST(train=True)
dataloader = data.DataLoader(dataset, batch_size=32)

loss_fn = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001)

for images, labels in dataloader:
    optimizer.zero_grad()            # set gradients to zero
    output = net(images)             # output of NN
    loss = loss_fn(output, labels)   # compute loss
    loss.backward()                  # backpropagate gradients
    optimizer.step()                 # optimizer step