Suggestions to `Image Classification` notebook

[yakutovicha]

We will go through the notebook and collect the suggestions here.

[edoardob90]

Apart from fixing some small typos (have not pushed them yet), here are my comments about the structure and content. BTW, it's already a very solid notebook 💪🏻

Augmentation transforms

The "Data Augmentation" section describes transforms to add (A.Affine, A.Rotate, A.HorizontalFlip, etc.) but the train_transform code cell only has Normalize + Resize. Am I missing something here?

Trainer class

The Trainer class (fit + predict) is ~60 lines of training loop infrastructure (gradients, backprop, early stopping). Maybe it's better to give it implemented with the most important/critical parts commented in the code. Because I think most of them will just copy and paste the entire block without even reading it.

Explain (a bit) math formulas

The "Loss function" and "Optimiser function" sections jump straight into formulas. We can leave them, but adding a couple of short paragraphs about the "intuition" behind them would be a good addition.

Add a brief "What is a CNN?"

A short conceptual paragraph would help : what convolution does (sliding a filter to detect patterns), the feature hierarchy (edges -> textures -> shapes), and the overall pipeline (input -> conv blocks -> flatten -> classifier).

Tensor shape conventions

ImageDataset.__getitem__ calls np.transpose(image, (1,2,0)). Why? Maybe explaining that (C, H, W) is Pytorch convetion while (H, W, C) is the convetion of display/image libraries.

[fabioacl]

Normalize and Resize are categorized separately because they are preprocessing steps rather than augmentations. Their purpose is to standardize the input dimensions and scale the data values for the model.