File size: 4,881 Bytes
bcc0f94 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 |
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)
|