edutorch

Rewritten PyTorch framework designed to help you learn AI/ML

Dynamic Shape-Checking / Type-Checking (Typorch)

I'm interested in shape-checking for tensors. It doesnt need to be static type checking, having fancy asserts would be good enough.

(B, C, H, W)

(B, 1, H, W)

(B, H, W)

If you run a program with the shape-checker, it automatically inserts assert statements that the left side of the variable assignment must have that shape. Letters are tracked throughout (e.g. a new letter introduces a new variable), and a number asserts that that dimension must match exactly.

Tuple shapes maybe, to distinguish a shape comment from a regular comment.

Using this mode, the code is compiled uniquely and increases runtime.

Once you are confident with your shapes, you can simply run your program normally.

Goals

1. Readability. Everything should make it immediately obvious how the layer or mmodel works on its own.

No autograd - if you want a simple autograd implementation, check out Karpathy's micrograd repo.

TODO:

• Convert to dataclasses?