Spaces:
Runtime error
Runtime error
File size: 9,013 Bytes
01df1d6 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 |
'''Module for making resnet encoders.
'''
import torch
import torch.nn as nn
from cortex_DIM.nn_modules.convnet import Convnet
from cortex_DIM.nn_modules.misc import Fold, Unfold, View
_nonlin_idx = 6
class ResBlock(Convnet):
'''Residual block for ResNet
'''
def create_layers(self, shape, conv_args=None):
'''Creates layers
Args:
shape: Shape of input.
conv_args: Layer arguments for block.
'''
# Move nonlinearity to a separate step for residual.
final_nonlin = conv_args[-1][_nonlin_idx]
conv_args[-1] = list(conv_args[-1])
conv_args[-1][_nonlin_idx] = None
conv_args.append((None, 0, 0, 0, False, False, final_nonlin, None))
super().create_layers(shape, conv_args=conv_args)
if self.conv_shape != shape:
dim_x, dim_y, dim_in = shape
dim_x_, dim_y_, dim_out = self.conv_shape
stride = dim_x // dim_x_
next_x, _ = self.next_size(dim_x, dim_y, 1, stride, 0)
assert next_x == dim_x_, (self.conv_shape, shape)
self.downsample = nn.Sequential(
nn.Conv2d(dim_in, dim_out, kernel_size=1, stride=stride, padding=0, bias=False),
nn.BatchNorm2d(dim_out),
)
else:
self.downsample = None
def forward(self, x: torch.Tensor):
'''Forward pass
Args:
x: Input.
Returns:
torch.Tensor or list of torch.Tensor.
'''
if self.downsample is not None:
residual = self.downsample(x)
else:
residual = x
x = self.conv_layers[-1](self.conv_layers[:-1](x) + residual)
return x
class ResNet(Convnet):
def create_layers(self, shape, conv_before_args=None, res_args=None, conv_after_args=None, fc_args=None):
'''Creates layers
Args:
shape: Shape of the input.
conv_before_args: Arguments for convolutional layers before residuals.
res_args: Residual args.
conv_after_args: Arguments for convolutional layers after residuals.
fc_args: Fully-connected arguments.
'''
dim_x, dim_y, dim_in = shape
shape = (dim_x, dim_y, dim_in)
self.conv_before_layers, self.conv_before_shape = self.create_conv_layers(shape, conv_before_args)
self.res_layers, self.res_shape = self.create_res_layers(self.conv_before_shape, res_args)
self.conv_after_layers, self.conv_after_shape = self.create_conv_layers(self.res_shape, conv_after_args)
dim_x, dim_y, dim_out = self.conv_after_shape
dim_r = dim_x * dim_y * dim_out
self.reshape = View(-1, dim_r)
self.fc_layers, _ = self.create_linear_layers(dim_r, fc_args)
def create_res_layers(self, shape, block_args=None):
'''Creates a set of residual blocks.
Args:
shape: input shape.
block_args: Arguments for blocks.
Returns:
nn.Sequential: sequence of residual blocks.
'''
res_layers = nn.Sequential()
block_args = block_args or []
for i, (conv_args, n_blocks) in enumerate(block_args):
block = ResBlock(shape, conv_args=conv_args)
res_layers.add_module('block_{}_0'.format(i), block)
for j in range(1, n_blocks):
shape = block.conv_shape
block = ResBlock(shape, conv_args=conv_args)
res_layers.add_module('block_{}_{}'.format(i, j), block)
shape = block.conv_shape
return res_layers, shape
def forward(self, x: torch.Tensor, return_full_list=False):
'''Forward pass
Args:
x: Input.
return_full_list: Optional, returns all layer outputs.
Returns:
torch.Tensor or list of torch.Tensor.
'''
if return_full_list:
conv_before_out = []
for conv_layer in self.conv_before_layers:
x = conv_layer(x)
conv_before_out.append(x)
else:
conv_before_out = self.conv_layers(x)
x = conv_before_out
if return_full_list:
res_out = []
for res_layer in self.res_layers:
x = res_layer(x)
res_out.append(x)
else:
res_out = self.res_layers(x)
x = res_out
if return_full_list:
conv_after_out = []
for conv_layer in self.conv_after_layers:
x = conv_layer(x)
conv_after_out.append(x)
else:
conv_after_out = self.conv_after_layers(x)
x = conv_after_out
x = self.reshape(x)
if return_full_list:
fc_out = []
for fc_layer in self.fc_layers:
x = fc_layer(x)
fc_out.append(x)
else:
fc_out = self.fc_layers(x)
return conv_before_out, res_out, conv_after_out, fc_out
class FoldedResNet(ResNet):
'''Resnet with strided crop input.
'''
def create_layers(self, shape, crop_size=8, conv_before_args=None, res_args=None,
conv_after_args=None, fc_args=None):
'''Creates layers
Args:
shape: Shape of the input.
crop_size: Size of the crops.
conv_before_args: Arguments for convolutional layers before residuals.
res_args: Residual args.
conv_after_args: Arguments for convolutional layers after residuals.
fc_args: Fully-connected arguments.
'''
self.crop_size = crop_size
dim_x, dim_y, dim_in = shape
self.final_size = 2 * (dim_x // self.crop_size) - 1
self.unfold = Unfold(dim_x, self.crop_size)
self.refold = Fold(dim_x, self.crop_size)
shape = (self.crop_size, self.crop_size, dim_in)
self.conv_before_layers, self.conv_before_shape = self.create_conv_layers(shape, conv_before_args)
self.res_layers, self.res_shape = self.create_res_layers(self.conv_before_shape, res_args)
self.conv_after_layers, self.conv_after_shape = self.create_conv_layers(self.res_shape, conv_after_args)
self.conv_after_shape = self.res_shape
dim_x, dim_y, dim_out = self.conv_after_shape
dim_r = dim_x * dim_y * dim_out
self.reshape = View(-1, dim_r)
self.fc_layers, _ = self.create_linear_layers(dim_r, fc_args)
def create_res_layers(self, shape, block_args=None):
'''Creates a set of residual blocks.
Args:
shape: input shape.
block_args: Arguments for blocks.
Returns:
nn.Sequential: sequence of residual blocks.
'''
res_layers = nn.Sequential()
block_args = block_args or []
for i, (conv_args, n_blocks) in enumerate(block_args):
block = ResBlock(shape, conv_args=conv_args)
res_layers.add_module('block_{}_0'.format(i), block)
for j in range(1, n_blocks):
shape = block.conv_shape
block = ResBlock(shape, conv_args=conv_args)
res_layers.add_module('block_{}_{}'.format(i, j), block)
shape = block.conv_shape
dim_x, dim_y = shape[:2]
if dim_x != dim_y:
raise ValueError('dim_x and dim_y do not match.')
if dim_x == 1:
shape = (self.final_size, self.final_size, shape[2])
return res_layers, shape
def forward(self, x: torch.Tensor, return_full_list=False):
'''Forward pass
Args:
x: Input.
return_full_list: Optional, returns all layer outputs.
Returns:
torch.Tensor or list of torch.Tensor.
'''
x = self.unfold(x)
conv_before_out = []
for conv_layer in self.conv_before_layers:
x = conv_layer(x)
if x.size(2) == 1:
x = self.refold(x)
conv_before_out.append(x)
res_out = []
for res_layer in self.res_layers:
x = res_layer(x)
res_out.append(x)
if x.size(2) == 1:
x = self.refold(x)
res_out[-1] = x
conv_after_out = []
for conv_layer in self.conv_after_layers:
x = conv_layer(x)
if x.size(2) == 1:
x = self.refold(x)
conv_after_out.append(x)
x = self.reshape(x)
if return_full_list:
fc_out = []
for fc_layer in self.fc_layers:
x = fc_layer(x)
fc_out.append(x)
else:
fc_out = self.fc_layers(x)
if not return_full_list:
conv_before_out = conv_before_out[-1]
res_out = res_out[-1]
conv_after_out = conv_after_out[-1]
return conv_before_out, res_out, conv_after_out, fc_out
|