def test_main(self):
modules = (nn.Conv2d(3, 3, 3), nn.ReLU())
model = pystiche.SequentialModule(*modules)
for idx, module in enumerate(modules):
actual = getattr(model, str(idx))
desired = module
assert actual is desired
def test_call(self):
torch.manual_seed(0)
modules = (nn.Conv2d(3, 3, 3), nn.ReLU())
input = torch.rand(1, 3, 256, 256)
pystiche_model = pystiche.SequentialModule(*modules)
torch_model = nn.Sequential(*modules)
actual = pystiche_model(input)
desired = torch_model(input)
ptu.assert_allclose(actual, desired)
def johnson_alahi_li_2016_transformer_decoder(
impl_params: bool = True,
instance_norm: bool = True,
) -> pystiche.SequentialModule:
def get_value_range_delimiter() -> nn.Module:
if impl_params:
def value_range_delimiter(x: torch.Tensor) -> torch.Tensor:
return 150.0 * torch.tanh(x)
else:
def value_range_delimiter(x: torch.Tensor) -> torch.Tensor:
# sgm(2*x) == (tanh(x) + 1) / 2
return torch.sigmoid(2.0 * x)
class ValueRangeDelimiter(nn.Module):
def forward(self, x: torch.Tensor) -> torch.Tensor:
return value_range_delimiter(x)
return ValueRangeDelimiter()
modules = (
johnson_alahi_li_2016_conv_block(
in_channels=64 if instance_norm else 128,
out_channels=32 if instance_norm else 64,
kernel_size=3,
stride=2,
upsample=True,
instance_norm=instance_norm,
),
johnson_alahi_li_2016_conv_block(
in_channels=32 if instance_norm else 64,
out_channels=16 if instance_norm else 32,
kernel_size=3,
stride=2,
upsample=True,
instance_norm=instance_norm,
),
nn.Conv2d(
in_channels=16 if instance_norm else 32,
out_channels=3,
kernel_size=9,
padding=same_size_padding(kernel_size=9),
),
get_value_range_delimiter(),
)
return pystiche.SequentialModule(*modules)
def johnson_alahi_li_2016_transformer_encoder(
instance_norm: bool = True, ) -> pystiche.SequentialModule:
modules = (
nn.ReflectionPad2d(40),
johnson_alahi_li_2016_conv_block(
in_channels=3,
out_channels=16 if instance_norm else 32,
kernel_size=9,
instance_norm=instance_norm,
),
johnson_alahi_li_2016_conv_block(
in_channels=16 if instance_norm else 32,
out_channels=32 if instance_norm else 64,
kernel_size=3,
stride=2,
instance_norm=instance_norm,
),
johnson_alahi_li_2016_conv_block(
in_channels=32 if instance_norm else 64,
out_channels=64 if instance_norm else 128,
kernel_size=3,
stride=2,
instance_norm=instance_norm,
),
johnson_alahi_li_2016_residual_block(
channels=64 if instance_norm else 128,
instance_norm=instance_norm,
),
johnson_alahi_li_2016_residual_block(
channels=64 if instance_norm else 128,
instance_norm=instance_norm,
),
johnson_alahi_li_2016_residual_block(
channels=64 if instance_norm else 128,
instance_norm=instance_norm,
),
johnson_alahi_li_2016_residual_block(
channels=64 if instance_norm else 128,
instance_norm=instance_norm,
),
johnson_alahi_li_2016_residual_block(
channels=64 if instance_norm else 128,
instance_norm=instance_norm,
),
)
return pystiche.SequentialModule(*modules)
def encoder(
instance_norm: bool = True,
) -> pystiche.SequentialModule:
r"""Encoder part of the :class:`Transformer` from :cite:`JAL2016` .
Args:
instance_norm: If ``True``, use :class:`~torch.nn.InstanceNorm2d` rather than
:class:`~torch.nn.BatchNorm2d` as described in the paper. In addition, the
number of channels of the convolution layers is reduced by half.
"""
modules = (
nn.ReflectionPad2d(40),
conv_block(
in_channels=3,
out_channels=maybe_fix_num_channels(32, instance_norm),
kernel_size=9,
instance_norm=instance_norm,
),
conv_block(
in_channels=maybe_fix_num_channels(32, instance_norm),
out_channels=maybe_fix_num_channels(64, instance_norm),
kernel_size=3,
stride=2,
instance_norm=instance_norm,
),
conv_block(
in_channels=maybe_fix_num_channels(64, instance_norm),
out_channels=maybe_fix_num_channels(128, instance_norm),
kernel_size=3,
stride=2,
instance_norm=instance_norm,
),
residual_block(channels=maybe_fix_num_channels(128, instance_norm)),
residual_block(channels=maybe_fix_num_channels(128, instance_norm)),
residual_block(channels=maybe_fix_num_channels(128, instance_norm)),
residual_block(channels=maybe_fix_num_channels(128, instance_norm)),
residual_block(channels=maybe_fix_num_channels(128, instance_norm)),
)
return pystiche.SequentialModule(*modules)
def decoder(
impl_params: bool = True,
instance_norm: bool = True,
) -> pystiche.SequentialModule:
r"""Decoder part of the :class:`Transformer` from :cite:`JAL2016`.
Args:
impl_params: If ``True``, the output of the is not externally pre-processed
before being fed into the
:func:`~pystiche_papers.johnson_alahi_li_2016.perceptual_loss`. Since this
step is necessary to get meaningful encodings from the
:func:`~pystiche_papers.johnson_alahi_li_2016.multi_layer_encoder`, the
pre-processing transform has to be learned within the output layer of the
decoder. To make this possible, ``150 * tanh(input)`` is used as activation
in contrast to the ``(tanh(input) + 1) / 2`` given in the paper.
instance_norm: If ``True``, use :class:`~torch.nn.InstanceNorm2d` rather than
:class:`~torch.nn.BatchNorm2d` as described in the paper. In addition, the
number of channels of the convolution layers is reduced by half.
"""
def get_value_range_delimiter() -> nn.Module:
if impl_params:
# https://github.com/pmeier/fast-neural-style/blob/813c83441953ead2adb3f65f4cc2d5599d735fa7/train.lua#L25
# https://github.com/pmeier/fast-neural-style/blob/813c83441953ead2adb3f65f4cc2d5599d735fa7/fast_neural_style/models.lua#L137-L138
# A tanh with a constant factor of 150 is used instead of the
# (tanh(x) + 1) / 2 in the paper.
def value_range_delimiter(x: torch.Tensor) -> torch.Tensor:
return 150.0 * torch.tanh(x)
else:
def value_range_delimiter(x: torch.Tensor) -> torch.Tensor:
# (tanh(x) + 1) / 2 == sgm(2*x)
return torch.sigmoid(2.0 * x)
class ValueRangeDelimiter(nn.Module):
def forward(self, x: torch.Tensor) -> torch.Tensor:
return value_range_delimiter(x)
return ValueRangeDelimiter()
modules = (
conv_block(
in_channels=maybe_fix_num_channels(128, instance_norm),
out_channels=maybe_fix_num_channels(64, instance_norm),
kernel_size=3,
stride=2,
upsample=True,
instance_norm=instance_norm,
),
conv_block(
in_channels=maybe_fix_num_channels(64, instance_norm),
out_channels=maybe_fix_num_channels(32, instance_norm),
kernel_size=3,
stride=2,
upsample=True,
instance_norm=instance_norm,
),
AutoPadConv2d(
in_channels=maybe_fix_num_channels(32, instance_norm),
out_channels=3,
kernel_size=9,
),
get_value_range_delimiter(),
)
return pystiche.SequentialModule(*modules)
