Pytorch tutorial

• The tutorial has links to Jupyter notebooks and Google Colab notebooks!
• If running in Colab, be sure to enable the GPU (Runtime > Change runtime type)

Summaries / key takeaways

What is Pytorch?

Make sure you know how to do the following:

• How to construct empty, random, all-ones, and all-zeros tensors
• How to specify tensor datatype
• Create a tensor with the same shape as another tensor, but a different datatype
• Look up tensor operations
• Add, subtract and multiply tensors
• Convert between Numpy arrays and tensors
• Move tensors between CPU and GPU

Autograd

• Useful takeaways from this tutorial:
  • with torch.no_grad()
  • requires_grad determines whether the gradient is computed.
• Don’t know how important the other stuff is. Part 3 will go into how weight updates etc. is done in practice.

Neural Networks

• Notice how they define the network as a class.
• Dense layers are called “Linear”
• What the network actually does is in “forward”
  • Layers that don’t need weights (e.g. max_pool) aren’t members of the class
  • E.g. F.max_pool2d is used in forward but not defined in __init__
• .parameters() gets weights
• Using zero_grad is important at the start of each training loop. I copied the training loop from the end of the document, it’s worth remembering.

optimizer.zero_grad() # Zero gradients
output = net(input)
loss = criterion(output, target) # Compute loss
loss.backward() # Backprop
optimizer.step() # Update step

Training a Classifier

• Saving and loading
• torchvision is pretty good for computer vision

Datasets and Parallelization

• DataLoader and Dataset. Basically, create a Dataset that provides the data, and DataLoader can do shuffling and batching on that dataset.
• DataParallel splits batches and distributes them among GPUs. This can be useful since CSUA allows 2 GPUs per person by default!