Starting with Generative Models and Data Distributions

First question

I reframed generation as approximating the distribution of observed (x) and sampling from it, rather than simply producing an image.

Learning a distribution

Tensor shapes

I printed (N, C, H, W) before building the model and checked the feature count after flattening. Silent broadcasting can produce a numeric loss with the wrong meaning.

Plotting the data

I inspected class balance, ranges, missing values, and duplicates. Normalization statistics came only from the training split to avoid leakage.

Next connection

The next note factorizes a difficult joint distribution and converts maximum likelihood into a trainable loss.