File size: 6,683 Bytes
dc6da71 | 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 | """Small, timm-compatible EfficientNetV2-S inference implementation.
The Apache-2.0 timm implementation was reduced to the layers exercised by
``tf_efficientnetv2_s.in21k_ft_in1k`` and modified to remove its runtime
dependency. Module names intentionally match timm so the official UTMOS v2
state dictionary loads; see LICENSE.
"""
from __future__ import annotations
import math
import torch
import torch.nn.functional as F
from torch import nn
class Conv2dSame(nn.Conv2d):
def forward(self, x: torch.Tensor) -> torch.Tensor:
height, width = x.shape[-2:]
stride_h, stride_w = self.stride
kernel_h, kernel_w = self.kernel_size
dilation_h, dilation_w = self.dilation
output_h = math.ceil(height / stride_h)
output_w = math.ceil(width / stride_w)
pad_h = max(
(output_h - 1) * stride_h
+ (kernel_h - 1) * dilation_h
+ 1
- height,
0,
)
pad_w = max(
(output_w - 1) * stride_w
+ (kernel_w - 1) * dilation_w
+ 1
- width,
0,
)
if pad_h or pad_w:
x = F.pad(
x,
(
pad_w // 2,
pad_w - pad_w // 2,
pad_h // 2,
pad_h - pad_h // 2,
),
)
return F.conv2d(
x,
self.weight,
self.bias,
self.stride,
0,
self.dilation,
self.groups,
)
class BatchNormAct2d(nn.BatchNorm2d):
def __init__(self, channels: int, *, activate: bool) -> None:
super().__init__(channels, eps=1e-3, momentum=0.1)
self.activate = activate
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = super().forward(x)
return F.silu(x, inplace=True) if self.activate else x
def conv3x3(
in_channels: int,
out_channels: int,
stride: int,
*,
groups: int = 1,
) -> nn.Conv2d:
if stride == 2:
return Conv2dSame(
in_channels,
out_channels,
3,
stride=2,
groups=groups,
bias=False,
)
return nn.Conv2d(
in_channels,
out_channels,
3,
stride=1,
padding=1,
groups=groups,
bias=False,
)
class ConvBnAct(nn.Module):
def __init__(self) -> None:
super().__init__()
self.conv = conv3x3(24, 24, 1)
self.bn1 = BatchNormAct2d(24, activate=True)
def forward(self, x: torch.Tensor) -> torch.Tensor:
residual = x
x = self.bn1(self.conv(x))
return x + residual
class EdgeResidual(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
expansion: int,
stride: int,
) -> None:
super().__init__()
expanded_channels = in_channels * expansion
self.conv_exp = conv3x3(in_channels, expanded_channels, stride)
self.bn1 = BatchNormAct2d(expanded_channels, activate=True)
self.conv_pwl = nn.Conv2d(expanded_channels, out_channels, 1, bias=False)
self.bn2 = BatchNormAct2d(out_channels, activate=False)
self.has_residual = stride == 1 and in_channels == out_channels
def forward(self, x: torch.Tensor) -> torch.Tensor:
residual = x
x = self.bn1(self.conv_exp(x))
x = self.bn2(self.conv_pwl(x))
return x + residual if self.has_residual else x
class SqueezeExcite(nn.Module):
def __init__(self, expanded_channels: int, reduced_channels: int) -> None:
super().__init__()
self.conv_reduce = nn.Conv2d(expanded_channels, reduced_channels, 1)
self.conv_expand = nn.Conv2d(reduced_channels, expanded_channels, 1)
def forward(self, x: torch.Tensor) -> torch.Tensor:
scale = x.mean((2, 3), keepdim=True)
scale = F.silu(self.conv_reduce(scale), inplace=True)
return x * torch.sigmoid(self.conv_expand(scale))
class InvertedResidual(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
expansion: int,
stride: int,
) -> None:
super().__init__()
expanded_channels = in_channels * expansion
self.conv_pw = nn.Conv2d(in_channels, expanded_channels, 1, bias=False)
self.bn1 = BatchNormAct2d(expanded_channels, activate=True)
self.conv_dw = conv3x3(
expanded_channels,
expanded_channels,
stride,
groups=expanded_channels,
)
self.bn2 = BatchNormAct2d(expanded_channels, activate=True)
self.se = SqueezeExcite(expanded_channels, in_channels // 4)
self.conv_pwl = nn.Conv2d(expanded_channels, out_channels, 1, bias=False)
self.bn3 = BatchNormAct2d(out_channels, activate=False)
self.has_residual = stride == 1 and in_channels == out_channels
def forward(self, x: torch.Tensor) -> torch.Tensor:
residual = x
x = self.bn1(self.conv_pw(x))
x = self.bn2(self.conv_dw(x))
x = self.se(x)
x = self.bn3(self.conv_pwl(x))
return x + residual if self.has_residual else x
def make_stage(
block_type: type[EdgeResidual | InvertedResidual],
in_channels: int,
out_channels: int,
expansion: int,
repeats: int,
stride: int,
) -> nn.Sequential:
blocks = [block_type(in_channels, out_channels, expansion, stride)]
blocks.extend(
block_type(out_channels, out_channels, expansion, 1)
for _ in range(repeats - 1)
)
return nn.Sequential(*blocks)
class EfficientNetV2S(nn.Module):
"""Feature-only TF EfficientNetV2-S with timm-compatible parameter names."""
def __init__(self) -> None:
super().__init__()
self.conv_stem = Conv2dSame(3, 24, 3, stride=2, bias=False)
self.bn1 = BatchNormAct2d(24, activate=True)
self.blocks = nn.Sequential(
nn.Sequential(ConvBnAct(), ConvBnAct()),
make_stage(EdgeResidual, 24, 48, 4, 4, 2),
make_stage(EdgeResidual, 48, 64, 4, 4, 2),
make_stage(InvertedResidual, 64, 128, 4, 6, 2),
make_stage(InvertedResidual, 128, 160, 6, 9, 1),
make_stage(InvertedResidual, 160, 256, 6, 15, 2),
)
self.conv_head = nn.Conv2d(256, 1280, 1, bias=False)
self.bn2 = BatchNormAct2d(1280, activate=True)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.bn1(self.conv_stem(x))
x = self.blocks(x)
x = self.bn2(self.conv_head(x))
return x
|