Spaces:
Configuration error
Configuration error
File size: 6,269 Bytes
d01f62c |
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 |
import torch
import torch.nn.functional as F
from torch import nn
class ResBlock(nn.Module):
def __init__(self, indim, outdim=None, stride=1):
super(ResBlock, self).__init__()
if outdim == None:
outdim = indim
if indim == outdim and stride == 1:
self.downsample = None
else:
self.downsample = nn.Conv2d(indim, outdim, kernel_size=3, padding=1, stride=stride)
self.gn3 = nn.GroupNorm(8, outdim)
self.conv1 = nn.Conv2d(indim, outdim, kernel_size=3, padding=1, stride=stride)
self.conv2 = nn.Conv2d(outdim, outdim, kernel_size=3, padding=1)
self.gn1 = nn.GroupNorm(8, outdim)
self.gn2 = nn.GroupNorm(8, outdim)
def forward(self, x):
r = self.conv1(F.relu(x))
r = self.gn1(r)
r = self.conv2(F.relu(r))
r = self.gn2(r)
if self.downsample is not None:
x = self.gn3(self.downsample(x))
return x + r
class ResGN(nn.Module):
def __init__(self, indim, outdim):
super().__init__()
self.res1 = ResBlock(indim, outdim)
self.res2 = ResBlock(outdim, outdim)
def forward(self, x):
return self.res2(self.res1(x))
class GroupNorm1D(nn.Module):
def __init__(self, indim, groups=8):
super().__init__()
self.gn = nn.GroupNorm(groups, indim)
def forward(self, x):
return self.gn(x.permute(1, 2, 0)).permute(2, 0, 1)
class GNActDWConv2d(nn.Module):
def __init__(self, indim, gn_groups=32):
super().__init__()
self.gn = nn.GroupNorm(gn_groups, indim)
self.conv = nn.Conv2d(indim,
indim,
5,
dilation=1,
padding=2,
groups=indim,
bias=False)
def forward(self, x, size_2d):
h, w = size_2d
_, bs, c = x.size()
x = x.view(h, w, bs, c).permute(2, 3, 0, 1)
x = self.gn(x)
x = F.gelu(x)
x = self.conv(x)
x = x.view(bs, c, h * w).permute(2, 0, 1)
return x
class DWConv2d(nn.Module):
def __init__(self, indim, dropout=0.1):
super().__init__()
self.conv = nn.Conv2d(indim,
indim,
5,
dilation=1,
padding=2,
groups=indim,
bias=False)
self.dropout = nn.Dropout2d(p=dropout, inplace=True)
def forward(self, x, size_2d):
h, w = size_2d
_, bs, c = x.size()
x = x.view(h, w, bs, c).permute(2, 3, 0, 1)
x = self.conv(x)
x = self.dropout(x)
x = x.view(bs, c, h * w).permute(2, 0, 1)
return x
class ScaleOffset(nn.Module):
def __init__(self, indim):
super().__init__()
self.gamma = nn.Parameter(torch.ones(indim))
# torch.nn.init.normal_(self.gamma, std=0.02)
self.beta = nn.Parameter(torch.zeros(indim))
def forward(self, x):
if len(x.size()) == 3:
return x * self.gamma + self.beta
else:
return x * self.gamma.view(1, -1, 1, 1) + self.beta.view(
1, -1, 1, 1)
class ConvGN(nn.Module):
def __init__(self, indim, outdim, kernel_size, gn_groups=8):
super().__init__()
self.conv = nn.Conv2d(indim,
outdim,
kernel_size,
padding=kernel_size // 2)
self.gn = nn.GroupNorm(gn_groups, outdim)
def forward(self, x):
return self.gn(self.conv(x))
def seq_to_2d(tensor, size_2d):
h, w = size_2d
_, n, c = tensor.size()
tensor = tensor.view(h, w, n, c).permute(2, 3, 0, 1).contiguous()
return tensor
def drop_path(x, drop_prob: float = 0., training: bool = False):
if drop_prob == 0. or not training:
return x
keep_prob = 1 - drop_prob
shape = (
x.shape[0],
x.shape[1],
) + (1, ) * (x.ndim - 2
) # work with diff dim tensors, not just 2D ConvNets
random_tensor = keep_prob + torch.rand(
shape, dtype=x.dtype, device=x.device)
random_tensor.floor_() # binarize
output = x.div(keep_prob) * random_tensor
return output
def mask_out(x, y, mask_rate=0.15, training=False):
if mask_rate == 0. or not training:
return x
keep_prob = 1 - mask_rate
shape = (
x.shape[0],
x.shape[1],
) + (1, ) * (x.ndim - 2
) # work with diff dim tensors, not just 2D ConvNets
random_tensor = keep_prob + torch.rand(
shape, dtype=x.dtype, device=x.device)
random_tensor.floor_() # binarize
output = x * random_tensor + y * (1 - random_tensor)
return output
class DropPath(nn.Module):
def __init__(self, drop_prob=None, batch_dim=0):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
self.batch_dim = batch_dim
def forward(self, x):
return self.drop_path(x, self.drop_prob)
def drop_path(self, x, drop_prob):
if drop_prob == 0. or not self.training:
return x
keep_prob = 1 - drop_prob
shape = [1 for _ in range(x.ndim)]
shape[self.batch_dim] = x.shape[self.batch_dim]
random_tensor = keep_prob + torch.rand(
shape, dtype=x.dtype, device=x.device)
random_tensor.floor_() # binarize
output = x.div(keep_prob) * random_tensor
return output
class DropOutLogit(nn.Module):
def __init__(self, drop_prob=None):
super(DropOutLogit, self).__init__()
self.drop_prob = drop_prob
def forward(self, x):
return self.drop_logit(x, self.drop_prob)
def drop_logit(self, x, drop_prob):
if drop_prob == 0. or not self.training:
return x
random_tensor = drop_prob + torch.rand(
x.shape, dtype=x.dtype, device=x.device)
random_tensor.floor_() # binarize
mask = random_tensor * 1e+8 if (
x.dtype == torch.float32) else random_tensor * 1e+4
output = x - mask
return output
|