Spaces:
Sleeping
Sleeping
File size: 6,423 Bytes
d7b0f75 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from os.path import join as pjoin
from collections import OrderedDict
def weight_standardize(w, dim, eps):
"""Subtracts mean and divides by standard deviation."""
w = w - torch.mean(w, dim=dim)
w = w / (torch.std(w, dim=dim) + eps)
return w
def np2th(weights, conv=False):
"""Possibly convert HWIO to OIHW."""
if conv:
weights = weights.transpose([3, 2, 0, 1])
return torch.from_numpy(weights)
class StdConv2d(nn.Conv2d):
def forward(self, x):
w = weight_standardize(self.weight, [0, 1, 2], 1e-5)
return F.conv2d(x, w, self.bias, self.stride, self.padding,
self.dilation, self.groups)
def conv3x3(in_channels, out_channels, stride=1, groups=1, bias=False):
return StdConv2d(in_channels,
out_channels,
kernel_size=3,
stride=stride,
padding=1,
bias=bias,
groups=groups)
def conv1x1(in_channels, out_channels, stride=1, bias=False):
return StdConv2d(in_channels,
out_channels,
kernel_size=1,
stride=stride,
padding=0,
bias=bias)
class PreActBottleneck(nn.Module):
"""Pre-activation (v2) bottleneck block.
"""
def __init__(self,
in_channels,
out_channels=None,
mid_channels=None,
stride=1):
super().__init__()
out_channels = out_channels or in_channels
mid_channels = mid_channels or out_channels // 4
self.gn1 = nn.GroupNorm(32, mid_channels, eps=1e-6)
self.conv1 = conv1x1(in_channels, mid_channels, bias=False)
self.gn2 = nn.GroupNorm(32, mid_channels, eps=1e-6)
self.conv2 = conv3x3(mid_channels, mid_channels, stride,
bias=False) # Original code has it on conv1!!
self.gn3 = nn.GroupNorm(32, out_channels, eps=1e-6)
self.conv3 = conv1x1(mid_channels, out_channels, bias=False)
self.relu = nn.ReLU(inplace=True)
if (stride != 1 or in_channels != out_channels):
# Projection also with pre-activation according to paper.
self.downsample = conv1x1(in_channels,
out_channels,
stride,
bias=False)
self.gn_proj = nn.GroupNorm(out_channels, out_channels)
def forward(self, x):
# Residual branch
residual = x
if hasattr(self, 'downsample'):
residual = self.downsample(x)
residual = self.gn_proj(residual)
# Unit's branch
y = self.relu(self.gn1(self.conv1(x)))
y = self.relu(self.gn2(self.conv2(y)))
y = self.gn3(self.conv3(y))
y = self.relu(residual + y)
return y
class ResNetV2(nn.Module):
"""Implementation of Pre-activation (v2) ResNet mode."""
def __init__(self, block_units, width_factor):
super().__init__()
width = int(64 * width_factor)
self.width = width
self.downsample = 16 # four stride=2 conv2d layer
# The following will be unreadable if we split lines.
# pylint: disable=line-too-long
self.root = nn.Sequential(
OrderedDict([('conv',
StdConv2d(3,
width,
kernel_size=7,
stride=2,
bias=False,
padding=3)),
('gn', nn.GroupNorm(32, width, eps=1e-6)),
('relu', nn.ReLU(inplace=True)),
('pool',
nn.MaxPool2d(kernel_size=3, stride=2, padding=0))]))
self.body = nn.Sequential(
OrderedDict([
('block1',
nn.Sequential(
OrderedDict([('unit1',
PreActBottleneck(in_channels=width,
out_channels=width * 4,
mid_channels=width))] +
[(f'unit{i:d}',
PreActBottleneck(in_channels=width * 4,
out_channels=width * 4,
mid_channels=width))
for i in range(2, block_units[0] + 1)], ))),
('block2',
nn.Sequential(
OrderedDict([('unit1',
PreActBottleneck(in_channels=width * 4,
out_channels=width * 8,
mid_channels=width * 2,
stride=2))] +
[(f'unit{i:d}',
PreActBottleneck(in_channels=width * 8,
out_channels=width * 8,
mid_channels=width * 2))
for i in range(2, block_units[1] + 1)], ))),
('block3',
nn.Sequential(
OrderedDict([('unit1',
PreActBottleneck(in_channels=width * 8,
out_channels=width * 16,
mid_channels=width * 4,
stride=2))] +
[(f'unit{i:d}',
PreActBottleneck(in_channels=width * 16,
out_channels=width * 16,
mid_channels=width * 4))
for i in range(2, block_units[2] + 1)], ))),
]))
def forward(self, x):
x = self.root(x)
x = self.body(x)
return x
def resnet50():
return ResNetV2(block_units=(3, 4, 9), width_factor=1)
|