def test_no_numpy(mock_no_numpy):
with pytest.raises(RuntimeError,
match=r"This module requires numpy to be installed."):
FID()
with pytest.raises(RuntimeError,
match=r"fid_score requires numpy to be installed."):
fid_score(0, 0, 0, 0)
def test_compute_fid_sqrtm():
mu1 = torch.tensor([0, 0])
mu2 = torch.tensor([0, 0])
sigma1 = torch.tensor([[-1, 1], [1, 1]], dtype=torch.float64)
sigma2 = torch.tensor([[1, 0], [0, 1]], dtype=torch.float64)
with pytest.raises(ValueError, match=r"Imaginary component "):
fid_score(mu1, mu2, sigma1, sigma2)
sigma1 = torch.ones(
(2, 2), dtype=torch.float64) * torch.finfo(torch.float64).max
sigma2 = torch.tensor([[1, 0.5], [0, 0.5]], dtype=torch.float64)
assert torch.isinf(torch.tensor(fid_score(mu1, mu2, sigma1, sigma2)))
def test_fid_function():
train_samples, test_samples = torch.rand(10, 10), torch.rand(10, 10)
mu1, sigma1 = train_samples.mean(axis=0), cov(train_samples, rowvar=False)
mu2, sigma2 = test_samples.mean(axis=0), cov(test_samples, rowvar=False)
sigma1 = torch.tensor(sigma1, dtype=torch.float64)
sigma2 = torch.tensor(sigma2, dtype=torch.float64)
assert pytest.approx(
fid_score(mu1, mu2, sigma1, sigma2),
rel=1e-5) == pytorch_fid_score.calculate_frechet_distance(
mu1, sigma1, mu2, sigma2)