test_checkpoints.py

test_checkpoints.py

import unittest
import tempfile
import os

import torch


def models_eq(a, b):
    """
    Not the most elegant method. Compares two models' parameters.
    """
    sa = a.state_dict()
    sb = b.state_dict()

    sak = set(sa.keys())
    sbk = set(sb.keys())

    eq = len(sak) == len(sbk)

    if not eq:
        return eq

    eq = len(set.intersection(sak, sbk)) == len(sak)

    if not eq:
        return eq

    for k in sak:
        eq = sa[k].shape == sb[k].shape

        if not eq:
            return eq

        eq = sa[k].eq(sb[k]).all()

        if not eq:
            return eq

    return eq


class TestCheckpoints(unittest.TestCase):
    def test_save_load(self):
        shape = (3, 3)
        expected = torch.nn.Linear(*shape)
        actual = torch.nn.Linear(*shape)

        _, path = tempfile.mkstemp()

        try:
            assert models_eq(expected, actual) == False, "Models are randomly initialized and should not be equal."

            torch.save(expected.state_dict(), path)
            state_dict = torch.load(path)
            actual.load_state_dict(state_dict)

            assert models_eq(expected, actual) == True, "Models should be equal after loading checkpoint."
        except AssertionError as e:
            raise e
        except:
            raise RuntimeError('Test failed!')
        finally:
            os.remove(path)


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

python3