def test_HyperParameters_iter(params):
hyper_parameters = HyperParameters(sub_param=HyperParameters(**params),
**params)
names = tuple(iter(hyper_parameters))
assert set(names[:-1]) == set(params.keys())
assert names[-1] == "sub_param"
def test_HyperParameters_deepcopy():
x = HyperParameters(param=object(),
sub_param=HyperParameters(param=object()))
y = deepcopy(x)
assert isinstance(y, HyperParameters)
assert y is not x
assert y.param is not x.param
assert y.sub_param is not x.sub_param
assert y.sub_param.param is not x.sub_param.param
def test_HyperParameters_new_similar(subtests, params):
x = HyperParameters(**params)
new = -1
y = x.new_similar(a=new)
assert isinstance(y, HyperParameters)
with subtests.test("old"):
del params["a"]
for name, value in params.items():
assert getattr(y, name) == value
with subtests.test("new"):
assert y.a == new
def test_HyperParameters_named_children(params):
hyper_parameters = HyperParameters(**params)
named_children = tuple(hyper_parameters.named_children())
assert not named_children
name = tuple(params.keys())[0]
val = HyperParameters()
setattr(hyper_parameters, name, val)
named_children = tuple(hyper_parameters.named_children())
assert len(named_children) == 1
assert named_children[0][0] == name
assert named_children[0][1] is val
def hyper_parameters(impl_params: bool = True) -> HyperParameters:
r"""Hyper parameters from :cite:`GEB2016`."""
# https://github.com/pmeier/PytorchNeuralStyleTransfer/blob/master/NeuralStyleTransfer.ipynb
# Cell [8]
style_loss_layers: Tuple[str, ...] = (
"conv1_1",
"conv2_1",
"conv3_1",
"conv4_1",
"conv5_1",
)
if impl_params:
style_loss_layers = tuple(
layer.replace("conv", "relu") for layer in style_loss_layers)
# https://github.com/pmeier/PytorchNeuralStyleTransfer/blob/master/NeuralStyleTransfer.ipynb
# Cell [8]
style_loss_layer_weights = (compute_layer_weights(style_loss_layers)
if impl_params else "mean")
return HyperParameters(
content_loss=HyperParameters(
# https://github.com/pmeier/PytorchNeuralStyleTransfer/blob/master/NeuralStyleTransfer.ipynb
# Cell [8]
layer=f"{'relu' if impl_params else 'conv'}4_2",
# https://github.com/pmeier/PytorchNeuralStyleTransfer/blob/master/NeuralStyleTransfer.ipynb
# Cell [8]
score_weight=1e0,
),
style_loss=HyperParameters(
layers=style_loss_layers,
layer_weights=style_loss_layer_weights,
# https://github.com/pmeier/PytorchNeuralStyleTransfer/blob/master/NeuralStyleTransfer.ipynb
# Cell [8]
score_weight=1e3,
),
nst=HyperParameters(
# https://github.com/pmeier/PytorchNeuralStyleTransfer/blob/master/NeuralStyleTransfer.ipynb
# Cell [9]
num_steps=500,
# https://github.com/pmeier/PytorchNeuralStyleTransfer/blob/master/NeuralStyleTransfer.ipynb
# Cell [6]
starting_point="content" if impl_params else "random",
# https://github.com/pmeier/PytorchNeuralStyleTransfer/blob/master/NeuralStyleTransfer.ipynb
# Cell [4]
image_size=512,
),
)
def hyper_parameters() -> HyperParameters:
r"""Hyper parameters from :cite:`JAL2016`."""
return HyperParameters(
content_loss=HyperParameters(
layer="relu2_2",
# The paper reports no score weight so we go with the default value of the
# implementation instead
# https://github.com/pmeier/fast-neural-style/blob/813c83441953ead2adb3f65f4cc2d5599d735fa7/train.lua#L36
score_weight=1e0,
),
style_loss=HyperParameters(
layers=("relu1_2", "relu2_2", "relu3_3", "relu4_3"),
layer_weights="sum",
# The paper reports no style score weight so we go with the default value
# of the implementation instead
# https://github.com/pmeier/fast-neural-style/blob/813c83441953ead2adb3f65f4cc2d5599d735fa7/train.lua#L43
score_weight=5e0,
),
regularization=HyperParameters(
# The paper reports a range of regularization score weights so we go with
# the default value of the implementation instead
# https://github.com/pmeier/fast-neural-style/blob/813c83441953ead2adb3f65f4cc2d5599d735fa7/train.lua#L33
score_weight=1e-6,
),
content_transform=HyperParameters(image_size=(256, 256)),
style_transform=HyperParameters(edge_size=256),
batch_sampler=HyperParameters(num_batches=40000, batch_size=4),
)
def test_HyperParameters_delattr(params):
hyper_parameters = HyperParameters(**params)
for name in params.keys():
getattr(hyper_parameters, name)
delattr(hyper_parameters, name)
with pytest.raises(AttributeError):
getattr(hyper_parameters, name)
def hyper_parameters(impl_params: bool = True) -> HyperParameters:
r"""Hyper parameters from :cite:`LW2016`.
Args:
impl_params: Switch the behavior and hyper-parameters between the reference
implementation of the original authors and what is described in the paper.
For details see :ref:`here <li_wand_2016-impl_params>`.
"""
return HyperParameters(
content_loss=HyperParameters(
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L57
layer="relu4_1" if impl_params else "relu4_2",
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L58
score_weight=2e1 if impl_params else 1e0,
),
target_transforms=HyperParameters(
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L52
num_scale_steps=0 if impl_params else 3,
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L67
scale_step_width=5e-2,
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L51
num_rotate_steps=0 if impl_params else 2,
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L66
rotate_step_width=7.5,
),
style_loss=HyperParameters(
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L48
layers=("relu3_1", "relu4_1"),
layer_weights="sum",
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L50
patch_size=3,
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L53
stride=2 if impl_params else 1,
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L49
score_weight=1e-4 if impl_params else 1e0,
),
regularization=HyperParameters(
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L59
score_weight=1e-3),
image_pyramid=HyperParameters(
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L40
max_edge_size=384,
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L44
num_steps=100 if impl_params else 200,
# https://github.com/pmeier/CNNMRF/blob/fddcf4d01e2a6ce201059d8bc38597f74a09ba3f/cnnmrf.lua#L43
num_levels=3 if impl_params else None,
min_edge_size=64,
edge="long",
),
nst=HyperParameters(
starting_point="content" if impl_params else "random"),
)
def hyper_parameters(impl_params: bool = True) -> HyperParameters:
r"""Hyper parameters from :cite:`GEB+2017`."""
# https://github.com/pmeier/NeuralImageSynthesis/blob/cced0b978fe603569033b2c7f04460839e4d82c4/ExampleNotebooks/BasicStyleTransfer.ipynb
# Cell [3] / layers['style']
style_loss_layers: Tuple[str, ...] = (
"conv1_1",
"conv2_1",
"conv3_1",
"conv4_1",
"conv5_1",
)
if impl_params:
style_loss_layers = tuple(
layer.replace("conv", "relu") for layer in style_loss_layers
)
style_loss = HyperParameters(
layers=style_loss_layers,
# https://github.com/pmeier/NeuralImageSynthesis/blob/cced0b978fe603569033b2c7f04460839e4d82c4/ExampleNotebooks/BasicStyleTransfer.ipynb
# Cell [3] / weights['style']
layer_weights=compute_layer_weights(style_loss_layers),
# https://github.com/pmeier/NeuralImageSynthesis/blob/cced0b978fe603569033b2c7f04460839e4d82c4/ExampleNotebooks/BasicStyleTransfer.ipynb
# Cell [3] / sw
score_weight=1e3,
)
return HyperParameters(
content_loss=HyperParameters(
# https://github.com/pmeier/NeuralImageSynthesis/blob/cced0b978fe603569033b2c7f04460839e4d82c4/ExampleNotebooks/BasicStyleTransfer.ipynb
# Cell [3] / layers['content']
layer="relu4_2" if impl_params else "conv4_2",
# https://github.com/pmeier/NeuralImageSynthesis/blob/cced0b978fe603569033b2c7f04460839e4d82c4/ExampleNotebooks/BasicStyleTransfer.ipynb
# Cell [3] / cw
score_weight=1e0,
),
style_loss=style_loss,
guided_style_loss=style_loss.new_similar(
# https://github.com/pmeier/NeuralImageSynthesis/blob/cced0b978fe603569033b2c7f04460839e4d82c4/ExampleNotebooks/SpatialControl.ipynb
# TODO: find the cell where this is performed
region_weights="sum"
),
image_pyramid=HyperParameters(
# https://github.com/pmeier/NeuralImageSynthesis/blob/cced0b978fe603569033b2c7f04460839e4d82c4/ExampleNotebooks/BasicStyleTransfer.ipynb
# Cell [3] / img_size, hr_img_size
edge_sizes=(512 if impl_params else 500, 1024),
# https://github.com/pmeier/NeuralImageSynthesis/blob/cced0b978fe603569033b2c7f04460839e4d82c4/ExampleNotebooks/BasicStyleTransfer.ipynb
# Cell [3] / max_iter, hr_max_iter
num_steps=(500, 200),
),
)
def test_HyperParameters_getattr_unknown_attribute():
hyper_parameters = HyperParameters()
with pytest.raises(AttributeError):
getattr(hyper_parameters, "unknown_attribute")
def hyper_parameters(impl_params: bool = True,
instance_norm: bool = True) -> HyperParameters:
r"""Hyper parameters from :cite:`ULVL2016,UVL2017`."""
# https://github.com/pmeier/texture_nets/blob/aad2cc6f8a998fedc77b64bdcfe1e2884aa0fb3e/train.lua#L44
style_loss_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"))
return HyperParameters(
content_loss=HyperParameters(
layer="relu4_2",
# https://github.com/pmeier/texture_nets/blob/b2097eccaec699039038970b191780f97c238816/stylization_train.lua#L22
score_weight=6e-1 if impl_params and not instance_norm else 1e0,
),
style_loss=HyperParameters(
# https://github.com/pmeier/texture_nets/blob/aad2cc6f8a998fedc77b64bdcfe1e2884aa0fb3e/train.lua#L44
layers=style_loss_layers,
# https://github.com/pmeier/texture_nets/blob/b2097eccaec699039038970b191780f97c238816/stylization_train.lua#L23
# The backward pass of the GramLoss'es in style_loss is manipulated and
# this is not independent of the score_weight. For this reason the
# layer_weights are set here to have the correct score_weight in the
# individual GramLoss'es.
layer_weights=(
[1e3 if impl_params and not instance_norm else 1e0] *
len(style_loss_layers)),
score_weight=1e0,
),
content_transform=HyperParameters(edge_size=256, ),
style_transform=HyperParameters(
edge_size=256,
interpolation=InterpolationMode.BICUBIC
# https://github.com/pmeier/texture_nets/blob/aad2cc6f8a998fedc77b64bdcfe1e2884aa0fb3e/train.lua#L152
if impl_params and instance_norm
# https://github.com/pmeier/texture_nets/blob/b2097eccaec699039038970b191780f97c238816/src/descriptor_net.lua#L17
else InterpolationMode.BILINEAR,
),
# The number of iterations is split up into multiple epochs with
# corresponding num_batches:
num_batches=(
# 50000 = 25 * 2000
# https://github.com/pmeier/texture_nets/blob/aad2cc6f8a998fedc77b64bdcfe1e2884aa0fb3e/train.lua#L48
2000 if instance_norm
# 3000 = 10 * 300
# https://github.com/pmeier/texture_nets/blob/b2097eccaec699039038970b191780f97c238816/stylization_train.lua#L30
else 300) if impl_params else 200,
batch_size=((
# https://github.com/pmeier/texture_nets/blob/aad2cc6f8a998fedc77b64bdcfe1e2884aa0fb3e/train.lua#L50
1 if instance_norm
# https://github.com/pmeier/texture_nets/blob/b2097eccaec699039038970b191780f97c238816/stylization_train.lua#L32
else 4) if impl_params else 16),
optimizer=HyperParameters(
# https://github.com/pmeier/texture_nets/blob/b2097eccaec699039038970b191780f97c238816/stylization_train.lua#L29
lr=1e-3 if impl_params and instance_norm else 1e-1, ),
lr_scheduler=HyperParameters(
# https://github.com/pmeier/texture_nets/blob/aad2cc6f8a998fedc77b64bdcfe1e2884aa0fb3e/train.lua#L260
# https://github.com/pmeier/texture_nets/blob/b2097eccaec699039038970b191780f97c238816/stylization_train.lua#L201
lr_decay=0.8 if impl_params else 0.7,
delay=0 if impl_params else 4,
),
# The number of iterations is split up into multiple epochs with
# corresponding num_batches:
num_epochs=(
# 50000 = 25 * 2000
# https://github.com/pmeier/texture_nets/blob/aad2cc6f8a998fedc77b64bdcfe1e2884aa0fb3e/train.lua#L48
25 if impl_params and instance_norm
# 3000 = 10 * 300 / 2000 = 10 * 200
# https://github.com/pmeier/texture_nets/blob/b2097eccaec699039038970b191780f97c238816/stylization_train.lua#L30
else 10),
)
def test_HyperParameters_len(params):
hyper_parameters = HyperParameters(sub_param=HyperParameters(**params),
**params)
assert len(hyper_parameters) == 1 + len(params)
def test_HyperParameters_dict_unpacking(params):
hyper_parameters = HyperParameters(**params)
assert dict(hyper_parameters) == params
assert {**hyper_parameters} == params
def test_HyperParameters_getitem(params):
hyper_parameters = HyperParameters(sub_param=HyperParameters(**params),
**params)
for name in (*params.keys(), "sub_param"):
assert hyper_parameters[name] is getattr(hyper_parameters, name)
def test_HyperParameters_delattr_unknown_attr(params):
hyper_parameters = HyperParameters()
with pytest.raises(AttributeError):
delattr(hyper_parameters, "unknown_attribute")
def test_HyperParameters_properties(params):
hyper_parameters = HyperParameters(**params)
assert dict(hyper_parameters.properties()) == params
def test_HyperParameters_getattr(params):
hyper_parameters = HyperParameters(**params)
for name, val in params.items():
assert getattr(hyper_parameters, name) == val
