MNIST Basics

Anki

Colab link: https://colab.research.google.com/github/fastai/fastbook/blob/master/04_mnist_basics.ipynb#scrollTo=kKgM7c63yIbI

Aim: create a model that can classify an image as a 3 or a 7

Setup

Download dataset

FastAI: download (library-provided) dataset

path = untar_data(URLs.<dataset_name>)

FastAI: set global base path

Path.BASE_PATH = path

FastAI: get files at path

path.ls()

FastAI: get subdirectory from path

path/'sub1'/'fileA'

Analysis

FastAI: path to PIL image

Image.open(path)

FastAI: PIL image to tensor

tensor(image)

FastAI: view tensor as image

show_image(image_tensor)

Training

General

Pytorch: track gradients

tensor.requires_grad_()

Pytorch: backprop

loss.backward()

Pytorch: get calculated gradients

params.grad

MNIST

FastAI: tensor to batched iterator

DataLoader(tensor, batch_size)

FastAI: what is a DataLoader

A wrapper around a tensor enabling batched iteration

FastAI: group training and validation datasets (transforms can be applied to both)

DataLoaders(train, valid)

FastAI: what is a DataLoaders

A wrapper grouping training and validation datasets (transforms can be applied to both)

Pytorch: linear layer

nn.Linear(dims)

Pytorch: get module params

module.parameters()

FastAI: base class combining everything needed for training (also, what are the key components?)

Learner(dataloaders, model, loss_func, opt_func, ...)

FastAI: where are statistics from training run stored?

learner.recorder

Pytorch: combine modules/layers

nn.Sequential(m1, m2, ...)

Misc

Pytorch: cast to e.g. float

tensor.float()

Pytorch: torch.stack vs torch.cat?

torch.stack: concatenates a sequence of tensors along a new dimension

torch.cat: concatenates a sequence of tensors along a in the given dimension

Pytorch: stack tensors

torch.stack([tensorA, ...])

Pytorch: what is "concatenating"

Concatenating a sequence of tensors in the given dimension

Pytorch: what does .view(...) do on a tensor?

Changes the shape to the one specified.

Pytorch: how to change a tensor's shape (within reason)?

.view(...)

Pytorch: what does .unsqueeze(dim) do on a tensor?

Returns a new tensor with a dimension of size one inserted at the specified position.

Pytorch: how to insert a dimension of size one at a specified position?

.unsqueeze(dim)

Pytorch: conditional computation

torch.where(condition, x, y)