Torch Testing

A collection of assertion methods to compare PyTorch Tensors in tests.

Currently all assertion methods are provided by converting the tensors to numpy arrays and feeding them into an appropriate numpy.testing method. That way, on failure, detailed information is provided as to why the test failed.

Last tested with Python 3.6.4 :: Anaconda, Inc. and PyTorch 0.4.

Installation

You can install this package using pip:

pip install torch_testing

Usage example

You can assert the equality of two torch.tensors like

import unittest
import torch
import torch_testing as tt


class TestSomeClass(unittest.TestCase):

    def test_some_method(self):
        a = torch.tensor([1, 2])
        b = torch.tensor([1, 2])
        tt.assert_equal(a, b)

if __name__ == '__main__':
    unittest.main()

Assertion methods

assert_equal(actual, expected, **kwargs)

Currently this assertion method is provided by converting the tensors to numpy arrays using tensor.numpy() and feeding them to numpy.testing.assert_equal.

assert_allclose(actual, expected, rtol=1e-07, atol=0, equal_nan=True, **kwargs)

Currently this assertion method is provided by converting the tensors to numpy arrays using tensor.numpy() and feeding them to numpy.testing.assert_allclose.

assert_within(tensor, min_val, max_val, rtol=0)

Ensures that all values of the given tensor are greater than or equal to min_val and less than or equal to max_val. Allows to specify a relative tolerance rtol, which behaves as in numpy.testing.assert_allclose.

NOTE: Uses assert_allclose under the hood, hence the fail message might currently be a little confusing.

Development

Unless noted otherwise, all commands are expected to be executed from the root directory of this repository.

Building the package for local development

To make the package available locally while making sure changes to the files are reflected immediately, run

pip install -e .

Test suite

Run all tests using

python -m unittest discover tests