This is simple WIP hacky testing/assertion/logging library for Pytorch stuff.

It was mostly intended for personal use. As I continue add to it I'll hopefully eventually clean it up so others could more easily use.

Nice assertions on tensors

Combustable provides utilites for making assertions on tensors and better outputs on failure.

import torch
from combustable.equality import AssertTensor

a = torch.tensor([[1.0, 2, 3]])

# Make assertions on our tensor
AssertTensor(a).is_close_to(torch.tensor([[1.0, 2, 3]]))
# List-like's are cast into tensors for you. More concise.
AssertTensor(a).is_close_to([[1, 2, 3]])
# Depending on what you're doing might need to adjust epsilon
AssertTensor(a).is_close_to([[1.01, 2, 3]], epsilon=1e-1)
# Other kinds of assertions
AssertTensor(a).has_shape(1, 3)

The autocasting when comparing with lists was one of the main motivation for Combustable. Right now Combustable's output on failure is ok (better than a naked assertion, at least), but also want to eventually make it much more pretty/debug-friendly.

Assertions on assign.

A common idiom in pytorch is unpacking the shape of something and then verify the shape is as expected.

Combustable introduces some syntactic sugar for doing this more concisely.

# Without this sugar
def my_cool_layer_func(a: torch.Tensor, b: torch.Tensor):
    batch_size, seq_len, hidden = a.shape
    b_batch_size, class_ind = b.shape
    assert batch_size == b_batch_size
    # Having these variables unpacked is nice as it is self documenting
    # and gives us access to the axis-size info if we need them in the method.
    # However, it'd be nice if we didn't have to have this `b_batch_size` temp var.
    # We only really expect one batch_size.

# With assertions on assign
from combustable.assert_on_assign import eq
def my_cool_layer_func(a: torch.Tensor, b: torch.Tensor):
    batch_size, seq_len, hidden = a.shape
    eq[batch_size], class_ind = b.shape
    # If the first dim of `b` != batch_size we will get an error.
    # The assertion is automatic. This is concise and avoids polluting our namespace
    # with a temp variable.

Install

pip install git+https://github.com/DNGros/combustable.git
# OR
git clone https://github.com/DNGros/combustable.git
cd combustable
pip install .

Like I mentioned, it's mostly WIP/for personal use. So no pypi package yet.