Pet Breeds

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

Misc

FastAI: what is notable about lists

Uses a special L class with additional features

Setup

FastAI: what is a DataBlock?

A container for building Datasets and DataLoaders

FastAI: what class enables easy building of Datasets and DataLoaders

DataBlock

FastAI: what arguments must be passed to DataBlock

DataBlock(blocks, get_items, splitter)

FastAI: what additional arguments may be useful to pass to DataBlock?

DataBlock(..., get_x, get_y, item_tfms, batch_tfms)

FastAI: what is the key difference between a DataBlock and a Datasets/DataLoaders?

The DataBlock doesn't specify the location of the data, only what to do with it

FastAI: resizing image transform

Resize(side_len) (square) or Resize(x_len, y_len)

FastAI: multi-transform utility function for images

aug_transforms

FastAI: create a Datasets/ Dataloaders from a DataBlock

data_block.datasets(path) or data_block.dataloaders(path)

FastAI: debug DataBlock

data_block.summary(path)

Training

FastAI: base helper class for analysing model predictions

Interpretation (e.g. ClassificationInterpretation)

FastAI: how do you use the learning rate finder?

learner.lr_find()

FastAI: what does the learning rate finder do

Start from a tiny lr and increase it until the loss begins to increase (then plots the graph)

FastAI: given the learning rate finder graph, how should one pick a learning rate?

Either:

One order of magnitude (base 10) less than where the minimum loss was achieved

The last point where the loss was clearly decreasing

FastAI: discriminative learning rates

learn.fit_one_cycle(..., max_lr=slice(low, high))

Each layer has a different learning rate, spread equally across the given range

At what point should you stop training

When your validation metrics are getting worse, not the loss

What should you do if validation loss starts getting worse

Check validation metrics ➡️ only if they start getting worse should you stop training

What should you do if validation metrics start getting worse

Stop training

Why might validation loss begin to get worse, but validation metrics continue to improve?

The model is becoming overconfident in its predictions (loss ⬆️)

(even when the actual accuracy, as measured by the validation metrics, is still improving)

If overfitting is detected, what should we do instead of early stopping (to get best accuracy)

Retrain from scratch with a lower learning rate

At what point should you train a model during the development process

As soon as possible!