torch-shapeguard

ShapeGuard

ShapeGuard allows you to very succinctly assert the expected shapes of tensors in a dynamic, einsum inspired way

Turn this:

def batch_outer_product(x, y):
    # x has shape (batch, x_channels)
    # y has shape (batch, y_channels)
    # return has shape (batch, x_channels, y_channels)

    return x.unsqueeze(-1) * y.unsqueeze(-2)

Into this:

def batch_outer_product(x, y):        
    x.sg(("batch", "x_channels"))
    y.sg(("batch", "y_channels"))

    return (x.unsqueeze(-1) * y.unsqueeze(-2)).sg(("batch", "x_channels", "y_channels"))

Installation

pip install torch-shapeguard

Motivation

It’s easy to make bugs in ml. One particular rich source of bugs is due to the flexibility of the operators: a*b works whether a and b are vectors, scalar vector, vector vector, etc. Similarly .sum() will work regardless of the shape of your tensor. Since we're doing optimization whatever computation we end up performing, we can probably optimize it to work reasonably, even if it's not doing what we intended. So our algorithm might "work" even if we have bugs (just less well). This makes bugs super hard to discover.

The best way I’ve found to avoid bugs is to religiously check the shapes of all my tensors, all the time, so I end up spending a lot of time debugging and writing comments like #(bs, n_samples, z_size) all over the place.

So why not algorithmically check the shapes then? Well it gets ugly fast.

You have to add assert foo.shape == (bs, n_samples, x_size) everywhere, which essentially doubles your linecount and you have to define all your dimensional sizes (bs, etc.), which might vary across train/test, batches, etc. So I made a small helper that makes it much nicer. I call it ShapeGuard.

Usage

When you import shapeguard, It adds the sg method to torch.Tensor and torch.distributions.Distribution.

You can use the sg method like an assert:

def forward(self, x, y):
    x.sg("bchw")
    y.sg("by")

This will verify that x has 4 dimensions, y has 2 dimensions and that x and y have the same size in the first dimension 'b'.

If the assert passes, the tensor is returned. This means you can also chain it inline on results of operations:

z = f(x).sg("bnz").mean(axis=1).sg("bz")

If the assert fails it produces a nice error message:

AssertionError: expected 'b' to be 2 but was 4

If you want to verify an exact dimension you can pass an int as the shape e.g.

def forward(self, x, y):
    x.sg(("b", 1, "h", "w"))
    y.sg("by")

The special shape '*' is reserved for shapes that should not be asserted, e.g. x.sg("*chw") will assert all shapes except the first.

How it works

The first time sg is called for an unseen shape, the size of the tensor for that shape is saved in the ShapeGuard.shapes global dict. Subsequent calls are checked against this stored shape.

You can call ShapeGuard.reset(shape) to reset a specific shape. This can be useful if e.g. your batch size varies between runs. ShapeGuard.reset() resets all shapes.