File size: 6,988 Bytes
2fa5aae |
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 |
from torch.nn.utils.parametrizations import spectral_norm
import torch
from torch.nn import functional as F
import einops
class GBlock(torch.nn.Module):
def __init__(self, in_channel: int, out_channel: int):
super().__init__()
self.in_channel = in_channel
self.out_channel = out_channel
# TODO: close batch
# self.bn1 = torch.nn.BatchNorm2d(in_channel)
# self.bn2 = torch.nn.BatchNorm2d(out_channel)
self.relu = torch.nn.ReLU()
# conv1x1
self.conv1x1 = spectral_norm(
torch.nn.Conv2d(in_channel, out_channel, kernel_size=1)
)
self.conv3x3_1 = spectral_norm(
torch.nn.Conv2d(in_channel, out_channel, kernel_size=3, padding=1)
)
self.conv3x3_2 = spectral_norm(
torch.nn.Conv2d(out_channel, out_channel, kernel_size=3, padding=1)
)
def forward(self, x: torch.Tensor):
# if shape is different then applied
if x.shape[1] != self.out_channel:
res = self.conv1x1(x)
else:
res = x.clone()
# first
# x = self.bn1(x)
x = self.relu(x)
x = self.conv3x3_1(x)
# second
# x = self.bn2(x)
x = self.relu(x)
x = self.conv3x3_2(x)
y = x + res
return y
class Up_GBlock(torch.nn.Module):
def __init__(self, in_channel: int):
super().__init__()
self.in_channel = in_channel
self.out_channel = int(in_channel / 2)
# TODO: close batch
# self.bn1 = torch.nn.BatchNorm2d(in_channel)
# self.bn2 = torch.nn.BatchNorm2d(self.out_channel)
# self.bn2 = torch.nn.BatchNorm2d(in_channel)
self.relu = torch.nn.ReLU()
self.up = torch.nn.Upsample(scale_factor=2, mode='nearest')
self.conv1x1 = spectral_norm(
torch.nn.Conv2d(in_channel, self.out_channel, kernel_size=1)
)
self.conv3x3_1 = spectral_norm(
torch.nn.Conv2d(in_channel, in_channel, kernel_size=3, padding=1)
)
self.conv3x3_2 = spectral_norm(
torch.nn.Conv2d(
in_channel,
self.out_channel,
kernel_size=3,
padding=1))
def forward(self, x):
res = self.up(x)
res = self.conv1x1(res)
# x = self.bn1(x)
x = self.relu(x)
x = self.up(x)
x = self.conv3x3_1(x)
# x = self.bn2(x)
x = self.relu(x)
x = self.conv3x3_2(x)
y = x + res
return y
class DBlock(torch.nn.Module):
def __init__(
self,
in_channel: int,
out_channel: int,
conv_type='2d',
apply_relu=True,
apply_down=False):
super().__init__()
self.in_channel = in_channel
self.out_channel = out_channel
self.apply_relu = apply_relu
self.apply_down = apply_down
# construct layer
if conv_type == '2d':
self.avg_pool = torch.nn.AvgPool2d(kernel_size=2, stride=2)
conv = torch.nn.Conv2d
elif conv_type == '3d':
self.avg_pool = torch.nn.AvgPool3d(kernel_size=2, stride=2)
conv = torch.nn.Conv3d
self.relu = torch.nn.ReLU()
self.conv1x1 = spectral_norm(
conv(in_channel, out_channel, kernel_size=1)
)
self.conv3x3_1 = spectral_norm(
conv(in_channel, out_channel, kernel_size=3, padding=1)
)
self.conv3x3_2 = spectral_norm(
conv(out_channel, out_channel, kernel_size=3, padding=1)
)
def forward(self, x):
# Residual block
if x.shape[1] != self.out_channel:
res = self.conv1x1(x)
else:
res = x.clone()
if self.apply_down:
res = self.avg_pool(res)
##
if self.apply_relu:
x = self.relu(x)
x = self.conv3x3_1(x)
x = self.relu(x)
x = self.conv3x3_2(x)
if self.apply_down:
x = self.avg_pool(x)
# connect
y = res + x
return y
class LBlock(torch.nn.Module):
def __init__(self, in_channel, out_channel):
super().__init__()
self.in_channel = in_channel
self.out_channel = out_channel
self.relu = torch.nn.ReLU()
conv = torch.nn.Conv2d
self.conv1x1 = conv(
in_channel,
(out_channel - in_channel),
kernel_size=1)
self.conv3x3_1 = conv(in_channel, in_channel, kernel_size=3, padding=1)
self.conv3x3_2 = conv(
in_channel,
out_channel,
kernel_size=3,
padding=1)
def forward(self, x):
res = torch.cat([x, self.conv1x1(x)], dim=1)
x = self.relu(x)
x = self.conv3x3_1(x)
x = self.relu(x)
x = self.conv3x3_2(x)
y = x + res
return y
# Attention Layer
def attention_einsum(q, k, v):
"""
Apply self-attention to tensors
"""
# Reshape 3D tensor to 2D tensor with first dimension L = h x w
k = einops.rearrange(k, "h w c -> (h w) c") # [h, w, c] -> [L, c]
v = einops.rearrange(v, "h w c -> (h w) c") # [h, w, c] -> [L, c]
# Einstein summation corresponding to the query * key operation.
beta = F.softmax(torch.einsum("hwc, Lc->hwL", q, k), dim=-1)
# Einstein summation corresponding to the attention * value operation.
out = torch.einsum("hwL, Lc->hwc", beta, v)
return out
class AttentionLayer(torch.nn.Module):
def __init__(self, in_channel, out_channel, ratio_kq=8, ratio_v=8):
super().__init__()
self.ratio_kq = ratio_kq
self.ratio_v = ratio_v
self.in_channel = in_channel
self.out_channel = out_channel
# compute query, key, and value using 1x1 convolution
self.query = torch.nn.Conv2d(
in_channel,
out_channel // ratio_kq,
kernel_size=1,
bias=False
)
self.key = torch.nn.Conv2d(
in_channel,
out_channel // ratio_kq,
kernel_size=1,
bias=False
)
self.value = torch.nn.Conv2d(
in_channel,
out_channel // ratio_v,
kernel_size=1,
bias=False
)
self.conv = torch.nn.Conv2d(
out_channel // 8,
out_channel,
kernel_size=1,
bias=False
)
self.gamma = torch.nn.Parameter(torch.zeros(1))
def forward(self, x):
query = self.query(x)
key = self.key(x)
value = self.value(x)
# apply attention
out = []
for i in range(x.shape[0]):
out.append(attention_einsum(query[i], key[i], value[i]))
out = torch.stack(out, dim=0)
out = self.gamma * self.conv(out)
out = out + x # skip connection
return out
|