File size: 6,400 Bytes
ccfe1ec |
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 |
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def initialize_weights(m):
if isinstance(m, nn.Conv1d):
n = m.kernel_size[0] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2 / n))
if m.bias is not None:
nn.init.constant_(m.bias.data, 0)
elif isinstance(m, nn.BatchNorm1d):
nn.init.constant_(m.weight.data, 1)
nn.init.constant_(m.bias.data, 0)
elif isinstance(m, nn.Linear):
m.weight.data.normal_(0, 0.001)
if m.bias is not None:
nn.init.constant_(m.bias.data, 0)
class SELayer(nn.Module):
def __init__(self, inp, reduction=4):
super(SELayer, self).__init__()
self.fc = nn.Sequential(
nn.Linear(inp, int(inp // reduction)),
nn.SiLU(),
nn.Linear(int(inp // reduction), inp),
nn.Sigmoid()
)
def forward(self, x):
b, c, _, = x.size()
y = x.view(b, c, -1).mean(dim=2)
y = self.fc(y).view(b, c, 1)
return x * y
class EffBlock(nn.Module):
def __init__(self, in_ch, ks, resize_factor, activation, out_ch=None, se_reduction=None):
super().__init__()
self.in_ch = in_ch
self.out_ch = self.in_ch if out_ch is None else out_ch
self.resize_factor = resize_factor
self.se_reduction = resize_factor if se_reduction is None else se_reduction
self.ks = ks
self.inner_dim = self.in_ch * self.resize_factor
block = nn.Sequential(
nn.Conv1d(
in_channels=self.in_ch,
out_channels=self.inner_dim,
kernel_size=1,
padding='same',
bias=False
),
nn.BatchNorm1d(self.inner_dim),
activation(),
nn.Conv1d(
in_channels=self.inner_dim,
out_channels=self.inner_dim,
kernel_size=ks,
groups=self.inner_dim,
padding='same',
bias=False
),
nn.BatchNorm1d(self.inner_dim),
activation(),
SELayer(self.inner_dim, reduction=self.se_reduction),
nn.Conv1d(
in_channels=self.inner_dim,
out_channels=self.in_ch,
kernel_size=1,
padding='same',
bias=False
),
nn.BatchNorm1d(self.in_ch),
activation(),
)
self.block = block
def forward(self, x):
return self.block(x)
class LocalBlock(nn.Module):
def __init__(self, in_ch, ks, activation, out_ch=None):
super().__init__()
self.in_ch = in_ch
self.out_ch = self.in_ch if out_ch is None else out_ch
self.ks = ks
self.block = nn.Sequential(
nn.Conv1d(
in_channels=self.in_ch,
out_channels=self.out_ch,
kernel_size=self.ks,
padding='same',
bias=False
),
nn.BatchNorm1d(self.out_ch),
activation()
)
def forward(self, x):
return self.block(x)
class ResidualConcat(nn.Module):
def __init__(self, fn):
super().__init__()
self.fn = fn
def forward(self, x, **kwargs):
return torch.concat([self.fn(x, **kwargs), x], dim=1)
class MapperBlock(nn.Module):
def __init__(self, in_features, out_features, activation=nn.SiLU):
super().__init__()
self.block = nn.Sequential(
nn.BatchNorm1d(in_features),
nn.Conv1d(in_channels=in_features,
out_channels=out_features,
kernel_size=1),
)
def forward(self, x):
return self.block(x)
class LegNet(nn.Module):
def __init__(self,
in_ch,
stem_ch,
stem_ks,
ef_ks,
ef_block_sizes,
pool_sizes,
resize_factor,
activation=nn.SiLU,
):
super().__init__()
assert len(pool_sizes) == len(ef_block_sizes)
self.in_ch = in_ch
self.stem = LocalBlock(in_ch=in_ch,
out_ch=stem_ch,
ks=stem_ks,
activation=activation)
blocks = []
in_ch = stem_ch
out_ch = stem_ch
for pool_sz, out_ch in zip(pool_sizes, ef_block_sizes):
blc = nn.Sequential(
ResidualConcat(
EffBlock(
in_ch=in_ch,
out_ch=in_ch,
ks=ef_ks,
resize_factor=resize_factor,
activation=activation)
),
LocalBlock(in_ch=in_ch * 2,
out_ch=out_ch,
ks=ef_ks,
activation=activation),
nn.MaxPool1d(pool_sz) if pool_sz != 1 else nn.Identity()
)
in_ch = out_ch
blocks.append(blc)
self.main = nn.Sequential(*blocks)
self.mapper = MapperBlock(in_features=out_ch,
out_features=out_ch * 2)
self.head = nn.Sequential(nn.Linear(out_ch * 2, out_ch * 2),
nn.BatchNorm1d(out_ch * 2),
activation(),
nn.Linear(out_ch * 2, 1))
def forward(self, x):
x = self.stem(x)
x = self.main(x)
x = self.mapper(x)
x = F.adaptive_avg_pool1d(x, 1)
x = x.squeeze(-1)
x = self.head(x)
x = x.squeeze(-1)
return x
|