File size: 4,897 Bytes
6c9ac8f |
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 |
import math
from typing import Tuple, Union
import torch
import torch.nn as nn
from mmcv.cnn.bricks import Swish, build_norm_layer
from torch.nn import functional as F
from torch.nn.init import _calculate_fan_in_and_fan_out, trunc_normal_
from mmdet.registry import MODELS
from mmdet.utils import OptConfigType
def variance_scaling_trunc(tensor, gain=1.):
fan_in, _ = _calculate_fan_in_and_fan_out(tensor)
gain /= max(1.0, fan_in)
std = math.sqrt(gain) / .87962566103423978
return trunc_normal_(tensor, 0., std)
@MODELS.register_module()
class Conv2dSamePadding(nn.Conv2d):
def __init__(self,
in_channels: int,
out_channels: int,
kernel_size: Union[int, Tuple[int, int]],
stride: Union[int, Tuple[int, int]] = 1,
padding: Union[int, Tuple[int, int]] = 0,
dilation: Union[int, Tuple[int, int]] = 1,
groups: int = 1,
bias: bool = True):
super().__init__(in_channels, out_channels, kernel_size, stride, 0,
dilation, groups, bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
img_h, img_w = x.size()[-2:]
kernel_h, kernel_w = self.weight.size()[-2:]
extra_w = (math.ceil(img_w / self.stride[1]) -
1) * self.stride[1] - img_w + kernel_w
extra_h = (math.ceil(img_h / self.stride[0]) -
1) * self.stride[0] - img_h + kernel_h
left = extra_w // 2
right = extra_w - left
top = extra_h // 2
bottom = extra_h - top
x = F.pad(x, [left, right, top, bottom])
return F.conv2d(x, self.weight, self.bias, self.stride, self.padding,
self.dilation, self.groups)
class MaxPool2dSamePadding(nn.Module):
def __init__(self,
kernel_size: Union[int, Tuple[int, int]] = 3,
stride: Union[int, Tuple[int, int]] = 2,
**kwargs):
super().__init__()
self.pool = nn.MaxPool2d(kernel_size, stride, **kwargs)
self.stride = self.pool.stride
self.kernel_size = self.pool.kernel_size
if isinstance(self.stride, int):
self.stride = [self.stride] * 2
if isinstance(self.kernel_size, int):
self.kernel_size = [self.kernel_size] * 2
def forward(self, x):
h, w = x.shape[-2:]
extra_h = (math.ceil(w / self.stride[1]) -
1) * self.stride[1] - w + self.kernel_size[1]
extra_v = (math.ceil(h / self.stride[0]) -
1) * self.stride[0] - h + self.kernel_size[0]
left = extra_h // 2
right = extra_h - left
top = extra_v // 2
bottom = extra_v - top
x = F.pad(x, [left, right, top, bottom])
x = self.pool(x)
return x
class DepthWiseConvBlock(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
apply_norm: bool = True,
conv_bn_act_pattern: bool = False,
norm_cfg: OptConfigType = dict(type='BN', momentum=1e-2, eps=1e-3)
) -> None:
super(DepthWiseConvBlock, self).__init__()
self.depthwise_conv = Conv2dSamePadding(
in_channels,
in_channels,
kernel_size=3,
stride=1,
groups=in_channels,
bias=False)
self.pointwise_conv = Conv2dSamePadding(
in_channels, out_channels, kernel_size=1, stride=1)
self.apply_norm = apply_norm
if self.apply_norm:
self.bn = build_norm_layer(norm_cfg, num_features=out_channels)[1]
self.apply_activation = conv_bn_act_pattern
if self.apply_activation:
self.swish = Swish()
def forward(self, x):
x = self.depthwise_conv(x)
x = self.pointwise_conv(x)
if self.apply_norm:
x = self.bn(x)
if self.apply_activation:
x = self.swish(x)
return x
class DownChannelBlock(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
apply_norm: bool = True,
conv_bn_act_pattern: bool = False,
norm_cfg: OptConfigType = dict(type='BN', momentum=1e-2, eps=1e-3)
) -> None:
super(DownChannelBlock, self).__init__()
self.down_conv = Conv2dSamePadding(in_channels, out_channels, 1)
self.apply_norm = apply_norm
if self.apply_norm:
self.bn = build_norm_layer(norm_cfg, num_features=out_channels)[1]
self.apply_activation = conv_bn_act_pattern
if self.apply_activation:
self.swish = Swish()
def forward(self, x):
x = self.down_conv(x)
if self.apply_norm:
x = self.bn(x)
if self.apply_activation:
x = self.swish(x)
return x
|