def test_content_transform_grayscale_image(subtests, content_image,
impl_params, instance_norm):
content_image = F.rgb_to_grayscale(content_image)
edge_size = 16
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
hyper_parameters.content_transform.edge_size = edge_size
content_transform = paper.content_transform(
impl_params=impl_params,
instance_norm=instance_norm,
hyper_parameters=hyper_parameters,
)
if instance_norm:
utils.make_reproducible()
actual = content_transform(content_image)
if impl_params:
if instance_norm:
# Since the transform involves an uncontrollable random component, we can't
# do a reference test here
return
transform_image = F.resize(content_image, edge_size)
else:
transform = transforms.CenterCrop(edge_size)
transform_image = transform(content_image)
desired = transform_image.repeat(1, 3, 1, 1)
ptu.assert_allclose(actual, desired)
def test_style_loss(subtests, impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(
impl_params=impl_params, instance_norm=instance_norm).style_loss
style_loss = paper.style_loss(
impl_params=impl_params,
instance_norm=instance_norm,
)
assert isinstance(style_loss, pystiche.loss.MultiLayerEncodingLoss)
with subtests.test("losses"):
assert all(
isinstance(loss, paper.GramLoss) for loss in style_loss.children())
layers, layer_weights = zip(*[(loss.encoder.layer, loss.score_weight)
for loss in style_loss.children()])
with subtests.test("layers"):
assert layers == hyper_parameters.layers
with subtests.test("layer_weights"):
assert layer_weights == pytest.approx(hyper_parameters.layer_weights)
with subtests.test("score_weight"):
assert style_loss.score_weight == pytest.approx(
hyper_parameters.score_weight)
def test_dataset(subtests, mocker, impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
mocker.patch(
"pystiche_papers.ulyanov_et_al_2016._data.ImageFolderDataset._collect_image_files",
return_value=[],
)
dataset = paper.dataset("root",
impl_params=impl_params,
instance_norm=instance_norm)
assert isinstance(dataset, torch.utils.data.IterableDataset)
with subtests.test("dataset"):
assert isinstance(dataset.dataset, ImageFolderDataset)
with subtests.test("content_transform"):
assert isinstance(dataset.transform, type(paper.content_transform()))
with subtests.test("min_size"):
assert dataset.min_size == hyper_parameters.content_transform.edge_size
with subtests.test("length"):
assert (len(dataset) == hyper_parameters.num_batches *
hyper_parameters.batch_size)
def test_stylization_smoke(stylization, postprocessor_mocks, input_image):
hyper_parameters = paper.hyper_parameters()
edge_size = hyper_parameters.content_transform.edge_size
_, _, output_image = stylization(input_image)
ptu.assert_allclose(output_image,
F.resize(input_image, (edge_size, edge_size)) + 0.5,
rtol=1e-6)
def test_hyper_parameters_lr_scheduler(subtests, impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
sub_params = "lr_scheduler"
assert sub_params in hyper_parameters
hyper_parameters = getattr(hyper_parameters, sub_params)
with subtests.test("lr_decay"):
assert hyper_parameters.lr_decay == 0.8 if impl_params else 0.7
def test_hyper_parameters_optimizer(subtests, impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
sub_params = "optimizer"
assert sub_params in hyper_parameters
hyper_parameters = getattr(hyper_parameters, sub_params)
with subtests.test("lr"):
assert hyper_parameters.lr == 1e-3 if impl_params and instance_norm else 1e-1
def test_hyper_parameters_content_transform(subtests, impl_params,
instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
sub_params = "content_transform"
assert sub_params in hyper_parameters
hyper_parameters = getattr(hyper_parameters, sub_params)
with subtests.test("edge_size"):
assert hyper_parameters.edge_size == 256
def test_hyper_parameters_num_images(impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
num_batches_per_epoch = hyper_parameters.num_batches
num_epochs = hyper_parameters.num_epochs
num_images = num_batches_per_epoch * num_epochs
assert num_images == (
50_000 if instance_norm else 3_000) if impl_params else 2_000
def test_style_transform(subtests, impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(
impl_params=impl_params, instance_norm=instance_norm).style_transform
style_transform = paper.style_transform(impl_params=impl_params,
instance_norm=instance_norm)
with subtests.test("edge_size"):
assert style_transform.edge_size == hyper_parameters.edge_size
with subtests.test("interpolation"):
assert style_transform.interpolation == hyper_parameters.interpolation
def test_image_loader(subtests, impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
dataset = ()
image_loader = paper.image_loader(dataset,
impl_params=impl_params,
instance_norm=instance_norm)
assert isinstance(image_loader, DataLoader)
with subtests.test("batch_size"):
assert image_loader.batch_size == hyper_parameters.batch_size
def test_hyper_parameters_content_loss(subtests, impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
sub_params = "content_loss"
assert sub_params in hyper_parameters
hyper_parameters = getattr(hyper_parameters, sub_params)
with subtests.test("layer"):
assert hyper_parameters.layer == "relu4_2"
with subtests.test("score_weight"):
assert hyper_parameters.score_weight == pytest.approx(
6e-1 if impl_params and not instance_norm else 1e0)
def test_hyper_parameters_lr_decay_delay(impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
num_batches = hyper_parameters.num_batches
delay = hyper_parameters.lr_scheduler.delay
num_batches_before_first_decay = num_batches * (delay + 1)
num_batches_between_decays = num_batches
assert (num_batches_before_first_decay == (2_000 if instance_norm else 300)
if impl_params else 1000)
assert (num_batches_between_decays == (2_000 if instance_norm else 300)
if impl_params else 200)
def test_hyper_parameters_style_transform(subtests, impl_params,
instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
sub_params = "style_transform"
assert sub_params in hyper_parameters
hyper_parameters = getattr(hyper_parameters, sub_params)
with subtests.test("edge_size"):
assert hyper_parameters.edge_size == 256
with subtests.test("interpolation"):
assert hyper_parameters.interpolation == (
InterpolationMode.BICUBIC
if impl_params and instance_norm else InterpolationMode.BILINEAR)
def test_content_loss(subtests, impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(
impl_params=impl_params, instance_norm=instance_norm).content_loss
content_loss = paper.content_loss(
impl_params=impl_params,
instance_norm=instance_norm,
)
assert isinstance(content_loss, pystiche.loss.FeatureReconstructionLoss)
with subtests.test("layer"):
assert content_loss.encoder.layer == hyper_parameters.layer
with subtests.test("score_weight"):
assert content_loss.score_weight == pytest.approx(
hyper_parameters.score_weight)
def test_ulyanov_et_al_2016_lr_scheduler(subtests, impl_params, instance_norm):
transformer = nn.Conv2d(3, 3, 1)
optimizer = paper.optimizer(transformer)
hyper_parameters = paper.hyper_parameters(
impl_params=impl_params, instance_norm=instance_norm).lr_scheduler
lr_scheduler = paper.lr_scheduler(optimizer,
impl_params=impl_params,
instance_norm=instance_norm)
assert isinstance(lr_scheduler, paper.DelayedExponentialLR)
with subtests.test("lr_decay"):
assert lr_scheduler.gamma == hyper_parameters.lr_decay
with subtests.test("delay"):
assert lr_scheduler.delay == hyper_parameters.delay
def test_hyper_parameters_style_loss(subtests, impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
sub_params = "style_loss"
assert sub_params in hyper_parameters
hyper_parameters = getattr(hyper_parameters, sub_params)
with subtests.test("layer"):
assert hyper_parameters.layers == (("relu1_1", "relu2_1", "relu3_1",
"relu4_1") if impl_params
and instance_norm else
("relu1_1", "relu2_1", "relu3_1",
"relu4_1", "relu5_1"))
with subtests.test("layer_weights"):
assert hyper_parameters.layer_weights == pytest.approx(
[1e3 if impl_params and not instance_norm else 1e0] *
len(hyper_parameters.layers))
with subtests.test("score_weight"):
assert hyper_parameters.score_weight == pytest.approx(1e0)
def test_ulyanov_et_al_2016_optimizer(subtests, impl_params, instance_norm):
transformer = nn.Conv2d(3, 3, 1)
params = tuple(transformer.parameters())
hyper_parameters = paper.hyper_parameters(
impl_params=impl_params, instance_norm=instance_norm).optimizer
optimizer = paper.optimizer(transformer,
impl_params=impl_params,
instance_norm=instance_norm)
assert isinstance(optimizer, optim.Adam)
assert len(optimizer.param_groups) == 1
param_group = optimizer.param_groups[0]
with subtests.test(msg="optimization params"):
assert len(param_group["params"]) == len(params)
for actual, desired in zip(param_group["params"], params):
assert actual is desired
with subtests.test(msg="optimizer properties"):
assert param_group["lr"] == ptu.approx(hyper_parameters.lr)
def test_hyper_parameters_num_batches(impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
assert hyper_parameters.num_batches == ((2000 if instance_norm else 300)
if impl_params else 200)
def test_hyper_parameters_batch_size(impl_params, instance_norm):
hyper_parameters = paper.hyper_parameters(impl_params=impl_params,
instance_norm=instance_norm)
assert hyper_parameters.batch_size == ((1 if instance_norm else 4)
if impl_params else 16)
def test_hyper_parameters_smoke():
hyper_parameters = paper.hyper_parameters()
assert isinstance(hyper_parameters, HyperParameters)