|
import torch |
|
import torch.nn as nn |
|
|
|
|
|
""" |
|
downsampling blocks |
|
(first half of the 'U' in UNet) |
|
[ENCODER] |
|
""" |
|
|
|
|
|
class EncoderLayer(nn.Module): |
|
def __init__( |
|
self, |
|
in_channels=1, |
|
out_channels=64, |
|
n_layers=2, |
|
all_padding=False, |
|
maxpool=True, |
|
): |
|
super(EncoderLayer, self).__init__() |
|
|
|
f_in_channel = lambda layer: in_channels if layer == 0 else out_channels |
|
f_padding = lambda layer: 1 if layer >= 2 or all_padding else 0 |
|
|
|
self.layer = nn.Sequential( |
|
*[ |
|
self._conv_relu_layer( |
|
in_channels=f_in_channel(i), |
|
out_channels=out_channels, |
|
padding=f_padding(i), |
|
) |
|
for i in range(n_layers) |
|
] |
|
) |
|
self.maxpool = maxpool |
|
|
|
def _conv_relu_layer(self, in_channels, out_channels, padding=0): |
|
return nn.Sequential( |
|
nn.Conv2d( |
|
in_channels=in_channels, |
|
out_channels=out_channels, |
|
kernel_size=3, |
|
padding=padding, |
|
), |
|
nn.ReLU(), |
|
) |
|
|
|
def forward(self, x): |
|
return self.layer(x) |
|
|
|
|
|
class Encoder(nn.Module): |
|
def __init__(self, config): |
|
super(Encoder, self).__init__() |
|
self.encoder = nn.ModuleDict( |
|
{ |
|
name: EncoderLayer( |
|
in_channels=block["in_channels"], |
|
out_channels=block["out_channels"], |
|
n_layers=block["n_layers"], |
|
all_padding=block["all_padding"], |
|
maxpool=block["maxpool"], |
|
) |
|
for name, block in config.items() |
|
} |
|
) |
|
self.maxpool = nn.MaxPool2d(2) |
|
|
|
def forward(self, x): |
|
output = dict() |
|
|
|
for i, (block_name, block) in enumerate(self.encoder.items()): |
|
x = block(x) |
|
output[block_name] = x |
|
|
|
if block.maxpool: |
|
x = self.maxpool(x) |
|
|
|
return x, output |
|
|
|
|
|
""" |
|
upsampling blocks |
|
(second half of the 'U' in UNet) |
|
[DECODER] |
|
""" |
|
|
|
|
|
class DecoderLayer(nn.Module): |
|
def __init__( |
|
self, in_channels, out_channels, kernel_size=2, stride=2, padding=[0, 0] |
|
): |
|
super(DecoderLayer, self).__init__() |
|
self.up_conv = nn.ConvTranspose2d( |
|
in_channels=in_channels, |
|
out_channels=in_channels // 2, |
|
kernel_size=kernel_size, |
|
stride=stride, |
|
padding=padding[0], |
|
) |
|
|
|
self.conv = nn.Sequential( |
|
*[ |
|
self._conv_relu_layer( |
|
in_channels=in_channels if i == 0 else out_channels, |
|
out_channels=out_channels, |
|
padding=padding[1], |
|
) |
|
for i in range(2) |
|
] |
|
) |
|
|
|
def _conv_relu_layer(self, in_channels, out_channels, padding=0): |
|
return nn.Sequential( |
|
nn.Conv2d( |
|
in_channels=in_channels, |
|
out_channels=out_channels, |
|
kernel_size=3, |
|
padding=padding, |
|
), |
|
nn.ReLU(), |
|
) |
|
|
|
@staticmethod |
|
def crop_cat(x, encoder_output): |
|
delta = (encoder_output.shape[-1] - x.shape[-1]) // 2 |
|
encoder_output = encoder_output[ |
|
:, :, delta : delta + x.shape[-1], delta : delta + x.shape[-1] |
|
] |
|
return torch.cat((encoder_output, x), dim=1) |
|
|
|
def forward(self, x, encoder_output): |
|
x = self.crop_cat(self.up_conv(x), encoder_output) |
|
return self.conv(x) |
|
|
|
|
|
class Decoder(nn.Module): |
|
def __init__(self, config): |
|
super(Decoder, self).__init__() |
|
self.decoder = nn.ModuleDict( |
|
{ |
|
name: DecoderLayer( |
|
in_channels=block["in_channels"], |
|
out_channels=block["out_channels"], |
|
kernel_size=block["kernel_size"], |
|
stride=block["stride"], |
|
padding=block["padding"], |
|
) |
|
for name, block in config.items() |
|
} |
|
) |
|
|
|
def forward(self, x, encoder_output): |
|
for name, block in self.decoder.items(): |
|
x = block(x, encoder_output[name]) |
|
return x |
|
|
|
|
|
class UNet(nn.Module): |
|
def __init__(self, encoder_config, decoder_config, nclasses): |
|
super(UNet, self).__init__() |
|
self.encoder = Encoder(config=encoder_config) |
|
self.decoder = Decoder(config=decoder_config) |
|
|
|
self.output = nn.Conv2d( |
|
in_channels=decoder_config["block1"]["out_channels"], |
|
out_channels=nclasses, |
|
kernel_size=1, |
|
) |
|
|
|
def forward(self, x): |
|
x, encoder_step_output = self.encoder(x) |
|
x = self.decoder(x, encoder_step_output) |
|
return self.output(x) |
|
|