Spaces:
Sleeping
Sleeping
File size: 7,855 Bytes
b157c29 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
"""
from https://github.com/Separius/SimCLRv2-Pytorch
"""
BATCH_NORM_EPSILON = 1e-5
BATCH_NORM_DECAY = 0.9 # == pytorch's default value as well
class BatchNormRelu(nn.Sequential):
def __init__(self, num_channels, relu=True):
super().__init__(nn.BatchNorm2d(num_channels, eps=BATCH_NORM_EPSILON),
nn.ReLU() if relu else nn.Identity())
def conv(in_channels, out_channels, kernel_size=3, stride=1, bias=False):
return nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size,
stride=stride, padding=(kernel_size - 1) // 2, bias=bias)
class SelectiveKernel(nn.Module):
def __init__(self, in_channels, out_channels, stride, sk_ratio, min_dim=32):
super().__init__()
assert sk_ratio > 0.0
self.main_conv = nn.Sequential(conv(in_channels, 2 * out_channels, stride=stride),
BatchNormRelu(2 * out_channels))
mid_dim = max(int(out_channels * sk_ratio), min_dim)
self.mixing_conv = nn.Sequential(conv(out_channels, mid_dim, kernel_size=1),
BatchNormRelu(mid_dim),
conv(mid_dim, 2 * out_channels, kernel_size=1))
def forward(self, x):
x = self.main_conv(x)
x = torch.stack(torch.chunk(x, 2, dim=1), dim=0) # 2, B, C, H, W
g = x.sum(dim=0).mean(dim=[2, 3], keepdim=True)
m = self.mixing_conv(g)
m = torch.stack(torch.chunk(m, 2, dim=1), dim=0) # 2, B, C, 1, 1
return (x * F.softmax(m, dim=0)).sum(dim=0)
class Projection(nn.Module):
def __init__(self, in_channels, out_channels, stride, sk_ratio=0):
super().__init__()
if sk_ratio > 0:
self.shortcut = nn.Sequential(nn.ZeroPad2d((0, 1, 0, 1)),
nn.AvgPool2d(kernel_size=2, stride=stride, padding=0),
conv(in_channels, out_channels, kernel_size=1))
else:
self.shortcut = conv(in_channels, out_channels, kernel_size=1, stride=stride)
self.bn = BatchNormRelu(out_channels, relu=False)
def forward(self, x):
return self.bn(self.shortcut(x))
class BottleneckBlock(nn.Module):
expansion = 4
def __init__(self, in_channels, out_channels, stride, sk_ratio=0, use_projection=False):
super().__init__()
if use_projection:
self.projection = Projection(in_channels, out_channels * 4, stride, sk_ratio)
else:
self.projection = nn.Identity()
ops = [conv(in_channels, out_channels, kernel_size=1), BatchNormRelu(out_channels)]
if sk_ratio > 0:
ops.append(SelectiveKernel(out_channels, out_channels, stride, sk_ratio))
else:
ops.append(conv(out_channels, out_channels, stride=stride))
ops.append(BatchNormRelu(out_channels))
ops.append(conv(out_channels, out_channels * 4, kernel_size=1))
ops.append(BatchNormRelu(out_channels * 4, relu=False))
self.net = nn.Sequential(*ops)
def forward(self, x):
shortcut = self.projection(x)
return F.relu(shortcut + self.net(x))
class Blocks(nn.Module):
def __init__(self, num_blocks, in_channels, out_channels, stride, sk_ratio=0):
super().__init__()
self.blocks = nn.ModuleList([BottleneckBlock(in_channels, out_channels, stride, sk_ratio, True)])
self.channels_out = out_channels * BottleneckBlock.expansion
for _ in range(num_blocks - 1):
self.blocks.append(BottleneckBlock(self.channels_out, out_channels, 1, sk_ratio))
def forward(self, x):
for b in self.blocks:
x = b(x)
return x
class Stem(nn.Sequential):
def __init__(self, sk_ratio, width_multiplier):
ops = []
channels = 64 * width_multiplier // 2
if sk_ratio > 0:
ops.append(conv(3, channels, stride=2))
ops.append(BatchNormRelu(channels))
ops.append(conv(channels, channels))
ops.append(BatchNormRelu(channels))
ops.append(conv(channels, channels * 2))
else:
ops.append(conv(3, channels * 2, kernel_size=7, stride=2))
ops.append(BatchNormRelu(channels * 2))
ops.append(nn.MaxPool2d(kernel_size=3, stride=2, padding=1))
super().__init__(*ops)
class ResNet(nn.Module):
def __init__(self, layers, width_multiplier, sk_ratio):
super().__init__()
ops = [Stem(sk_ratio, width_multiplier)]
channels_in = 64 * width_multiplier
ops.append(Blocks(layers[0], channels_in, 64 * width_multiplier, 1, sk_ratio))
channels_in = ops[-1].channels_out
ops.append(Blocks(layers[1], channels_in, 128 * width_multiplier, 2, sk_ratio))
channels_in = ops[-1].channels_out
ops.append(Blocks(layers[2], channels_in, 256 * width_multiplier, 2, sk_ratio))
channels_in = ops[-1].channels_out
ops.append(Blocks(layers[3], channels_in, 512 * width_multiplier, 2, sk_ratio))
channels_in = ops[-1].channels_out
self.channels_out = channels_in
self.net = nn.Sequential(*ops)
self.fc = nn.Linear(channels_in, 1000)
def forward(self, x, apply_fc=False):
h = self.net(x).mean(dim=[2, 3])
if apply_fc:
h = self.fc(h)
return h
class ContrastiveHead(nn.Module):
def __init__(self, channels_in, out_dim=128, num_layers=3):
super().__init__()
self.layers = nn.ModuleList()
for i in range(num_layers):
if i != num_layers - 1:
dim, relu = channels_in, True
else:
dim, relu = out_dim, False
self.layers.append(nn.Linear(channels_in, dim, bias=False))
bn = nn.BatchNorm1d(dim, eps=BATCH_NORM_EPSILON, affine=True)
if i == num_layers - 1:
nn.init.zeros_(bn.bias)
self.layers.append(bn)
if relu:
self.layers.append(nn.ReLU())
def forward(self, x):
for b in self.layers:
x = b(x)
return x
def get_resnet(depth=50, width_multiplier=1, sk_ratio=0): # sk_ratio=0.0625 is recommended
layers = {50: [3, 4, 6, 3], 101: [3, 4, 23, 3], 152: [3, 8, 36, 3], 200: [3, 24, 36, 3]}[depth]
resnet = ResNet(layers, width_multiplier, sk_ratio)
return resnet, ContrastiveHead(resnet.channels_out)
def name_to_params(checkpoint):
sk_ratio = 0.0625 if '_sk1' in checkpoint else 0
if 'r50_' in checkpoint:
depth = 50
elif 'r101_' in checkpoint:
depth = 101
elif 'r152_' in checkpoint:
depth = 152
else:
raise NotImplementedError
if '_1x_' in checkpoint:
width = 1
elif '_2x_' in checkpoint:
width = 2
elif '_3x_' in checkpoint:
width = 3
else:
raise NotImplementedError
return depth, width, sk_ratio
class SimCLRv2(nn.Module):
def __init__(self, model, head):
super(SimCLRv2, self).__init__()
self.encoder = model
self.contrastive_head = head
def forward(self, x):
x = self.encoder(x)
x = self.contrastive_head(x)
return x
def get_simclr2_model(ckpt_path):
depth, width, sk_ratio = name_to_params(ckpt_path)
model, head = get_resnet(depth, width, sk_ratio)
checkpoint = torch.load('pretrained_models/simclr2_models/' + ckpt_path)
model.load_state_dict(checkpoint['resnet'])
head.load_state_dict(checkpoint['head'])
del model.fc
simclr2 = SimCLRv2(model, head)
return simclr2.to(device) |