from torch import nn
from radionets.architecture.layers import (
ComplexConv2d,
ComplexInstanceNorm2d,
ComplexPReLU,
)
__all__ = [
"NNBlock",
"SRBlock",
"ComplexSRBlock",
"BottleneckResBlock",
"Encoder",
"Decoder",
]
[docs]
class NNBlock(nn.Module):
"""Base block for neural networks.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
*,
stride: int | tuple = 1,
padding: int = 0,
groups: int = 1,
dropout: bool | float = False,
):
"""Base block for neural networks.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
"""
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.groups = groups
self.dropout = dropout
self.idconv = (
nn.Identity()
if self.in_channels == self.out_channels
else nn.Conv2d(self.in_channels, self.out_channels, 1)
)
self.pool = (
nn.Identity()
if self.stride == 1
else nn.AvgPool2d(kernel_size=2, ceil_mode=True)
)
def _conv_block(self):
pass
[docs]
def forward(self, x):
pass
[docs]
class SRBlock(NNBlock):
"""Default SRResNet building block.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
*,
stride: int = 1,
padding: int = 0,
groups: int = 1,
dropout: bool | int = False,
):
super().__init__(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
groups=groups,
dropout=dropout,
)
"""Default SRResNet building block.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
"""
self.convs = nn.Sequential(*self._conv_block())
def _conv_block(self):
blocks = [
nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
stride=self.stride,
padding=1,
bias=False,
padding_mode="reflect",
),
nn.InstanceNorm2d(num_features=self.out_channels),
nn.PReLU(),
nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
stride=self.stride,
padding=1,
bias=False,
padding_mode="reflect",
),
nn.InstanceNorm2d(num_features=self.out_channels),
]
# NOTE: This will be included directly in the blocks
# list in a future release and is only here for
# compatibility reasons
if self.dropout:
blocks.insert(1, nn.Dropout(p=self.dropout))
blocks.insert(4, nn.Rropout(p=self.dropout))
return blocks
[docs]
def forward(self, x):
return self.convs(x) + self.idconv(self.pool(x))
[docs]
class ComplexSRBlock(NNBlock):
"""Complex SRResNet building block.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
*,
stride: int = 1,
padding: int = 0,
groups: int = 1,
dropout: int = 0,
):
super().__init__(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
groups=groups,
dropout=dropout,
)
"""Complex SRResNet building block.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
"""
self.convs = nn.Sequential(*self._conv_block())
def _conv_block(self):
blocks = [
ComplexConv2d(
in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
stride=self.stride,
padding=1,
bias=False,
),
nn.Dropout(p=self.dropout),
ComplexInstanceNorm2d(num_features=self.out_channels),
ComplexPReLU(num_parameters=2),
ComplexConv2d(
in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
stride=self.stride,
padding=1,
bias=False,
),
nn.Dropout(p=self.dropout),
ComplexInstanceNorm2d(num_features=self.out_channels),
]
return blocks
[docs]
def forward(self, x):
return self.convs(x) + self.idconv(self.pool(x))
[docs]
class BottleneckResBlock(NNBlock):
"""Three-convolution layer deep residual neural network
building block.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
*,
stride: int = 1,
padding: int = 0,
groups: int = 1,
dropout: bool | int = False,
downsample: bool = False,
):
"""Three-convolution layer deep residual neural network
building block.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
"""
super().__init__(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
groups=groups,
dropout=dropout,
)
self.prelu = nn.PReLU()
self.convs = self._conv_block()
def _conv_block(self):
block = nn.Sequential(
nn.Conv2d(
self.in_channels,
self.out_channels // 4,
kernel_size=1,
stride=1,
bias=False,
),
nn.Dropout(p=self.dropout),
nn.BatchNorm2d(self.out_channels // 4),
self.prelu,
nn.Conv2d(
self.out_channels // 4,
self.out_channels // 4,
kernel_size=3,
stride=self.stride,
padding=1,
bias=False,
),
nn.Dropout(p=self.dropout),
nn.BatchNorm2d(self.out_channels // 4),
self.prelu,
nn.Conv2d(
self.out_channels // 4,
self.out_channels,
kernel_size=1,
stride=1,
bias=False,
),
nn.Dropout(p=self.dropout),
nn.BatchNorm2d(self.out_channels),
)
return block
[docs]
def forward(self, x):
x0 = x
x = self.convs(x)
x += x0
x = self.prelu(x)
return x
[docs]
class Encoder(NNBlock):
"""Encoder block for UNets.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
bias : bool
Whether to apply bias. Default: False
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
batchnorm : bool, optional
If ``True``, add a batchnorm layer to the
encoder block. Default: False
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
*,
stride: int = 1,
padding: int = 1,
groups: int = 1,
bias: bool = False,
dropout: bool | int = False,
batchnorm: bool = False,
):
"""Encoder block for UNets.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
bias : bool
Whether to apply bias. Default: False
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
batchnorm : bool, optional
If ``True``, add a batchnorm layer to the
encoder block. Default: False
"""
super().__init__(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
groups=groups,
dropout=dropout,
)
self.bias = bias
self.batchnorm = batchnorm
self.encoder_block = nn.Sequential(self.__encoder_block())
def __encoder_block(self):
block = [
nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
stride=self.stride,
padding=self.padding,
bias=self.bias,
),
nn.PReLU(),
]
if self.batchnorm:
block.insert(1, nn.BatchNorm2d(self.out_channels))
return block
[docs]
def forward(self, x):
return self.encoder_block(x)
[docs]
class Decoder(NNBlock):
"""Decoder block for UNets.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
output_padding : int, optional
Controls the padding applied to the output.
See :class:`~torch.nn.ConvTranspose2d`. Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
bias : bool
Whether to apply bias. Default: False
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
stride: int = 1,
padding: int = 0,
output_padding: int = 0,
groups: int = 1,
bias: bool = False,
dropout: bool | int = False,
):
"""Decoder block for UNets.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int, optional
Size of the convolution kernel. Default: 3
stride : int or tuple, optional
Stride for the cross-correlation. Default: 1
padding : int, optional
The amount of padding applied to the input.
Default: 0
output_padding : int, optional
Controls the padding applied to the output.
See :class:`~torch.nn.ConvTranspose2d`. Default: 0
groups : int, optional
Controls the behavior of input and output groups.
See :class:`~torch.nn.Conv2d`. Default: 1
bias : bool
Whether to apply bias. Default: False
dropout : bool or float, optional
Wether to apply dropout. If float > 0 this is
the dropout percentage. Default: False
"""
super().__init__(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
groups=groups,
dropout=dropout,
)
self.padding = padding
self.output_padding = output_padding
self.bias = bias
self.decoder_block = nn.Sequential(self.__decoder_block())
def __decoder_block(self):
block = [
nn.ConvTranspose2d(
in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=self.kernel_size,
stride=self.stride,
padding=self.padding,
output_padding=self.output_padding,
bias=self.bias,
),
nn.PReLU(),
]
return block
[docs]
def forward(self, x):
return self.decoder_block(x)
