def test_inception_extractor_wrong_inputs():
with pytest.raises(ValueError,
match=r"Inputs should be a tensor of dim 4"):
InceptionModel(return_features=True)(torch.rand(2))
with pytest.raises(ValueError,
match=r"Inputs should be a tensor with 3 channels"):
InceptionModel(return_features=True)(torch.rand(2, 2, 2, 0))
def __init__(
self,
num_features: Optional[int] = None,
feature_extractor: Optional[torch.nn.Module] = None,
output_transform: Callable = lambda x: x,
device: Union[str, torch.device] = torch.device("cpu"),
) -> None:
try:
import numpy as np # noqa: F401
except ImportError:
raise RuntimeError("This module requires numpy to be installed.")
try:
import scipy # noqa: F401
except ImportError:
raise RuntimeError("This module requires scipy to be installed.")
if num_features is None and feature_extractor is None:
num_features = 1000
feature_extractor = InceptionModel(return_features=False,
device=device)
self._eps = 1e-6
super(FID, self).__init__(
num_features=num_features,
feature_extractor=feature_extractor,
output_transform=output_transform,
device=device,
)
def test_device_mismatch_cuda():
images = torch.rand(10, 3, 299, 299)
result = InceptionModel(return_features=False, device="cuda")(images)
assert result.is_cuda
assert result.shape == torch.Size([10, 1000])
result = InceptionModel(return_features=False)(images.cuda())
assert not result.is_cuda
assert result.shape == torch.Size([10, 1000])
images = torch.rand(10, 5)
result = DummyInceptionMetric(num_features=5,
device="cuda")._extract_features(images)
assert result.is_cuda
assert result.shape == torch.Size([10, 5])
result = DummyInceptionMetric(num_features=5)._extract_features(
images.cuda())
assert not result.is_cuda
assert result.shape == torch.Size([10, 5])
def __init__(
self,
num_features: Optional[int] = None,
feature_extractor: Optional[torch.nn.Module] = None,
output_transform: Callable = lambda x: x,
device: Union[str, torch.device] = torch.device("cpu"),
) -> None:
if num_features is None and feature_extractor is None:
num_features = 1000
feature_extractor = InceptionModel(return_features=False, device=device)
self._eps = 1e-16
super(InceptionScore, self).__init__(
num_features=num_features,
feature_extractor=feature_extractor,
output_transform=output_transform,
device=device,
)
def test_inception_model_probability():
x = torch.rand(2, 3, 299, 299)
y = InceptionModel(return_features=False)(x)
assert pytest.approx(torch.sum(y[0]).item()) == 1.0
assert pytest.approx(torch.sum(y[1]).item()) == 1.0
assert torch.all(0 <= y)
def test_no_torchvision(mock_no_torchvision):
with pytest.raises(
RuntimeError,
match=r"This module requires torchvision to be installed."):
InceptionModel(return_features=True)
