From b67731233ac27dc1596f98f919ec567a6a23728c Mon Sep 17 00:00:00 2001 From: lid Date: Sun, 6 Sep 2020 21:25:06 +0800 Subject: [PATCH] add Stand-Alone Self-Attention model --- lib/models/sa_resnet.py | 217 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 217 insertions(+) create mode 100644 lib/models/sa_resnet.py diff --git a/lib/models/sa_resnet.py b/lib/models/sa_resnet.py new file mode 100644 index 0000000..5abf5ab --- /dev/null +++ b/lib/models/sa_resnet.py @@ -0,0 +1,217 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +import math + + +class conv_qkv(nn.Conv2d): + """Conv1d for Q, K, V""" + + +def conv1x1(in_planes, out_planes, stride=1): + """1x1 convolution""" + return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) + + +class SelfAttention2d(nn.Module): + def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, groups=1, bias=False): + super(SelfAttention2d, self).__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = kernel_size + self.stride = stride + self.padding = padding + self.groups = groups + self.group_channels = out_channels // groups + + self.conv_qkv = conv_qkv(in_channels, out_channels * 3, kernel_size=1, bias=False) + self.unfold = nn.Unfold(kernel_size, 1, padding, 1) + self.softmax = nn.Softmax(dim=-1) + + self.relative_h = nn.Parameter(torch.randn(self.group_channels // 2, 1, 1, kernel_size, 1), requires_grad=True) + self.relative_w = nn.Parameter(torch.randn(self.group_channels // 2, 1, 1, 1, kernel_size), requires_grad=True) + + if stride > 1: + self.pooling = nn.AvgPool2d(stride, stride=stride) + + self.reset_parameters() + + def forward(self, x): + N, C, H, W = x.shape + + qkv = self.conv_qkv(x) + q, kv = qkv.split([self.out_channels, self.out_channels * 2], dim=1) + kv = self.unfold(kv).reshape(N, self.out_channels * 2, self.kernel_size**2, H * W).transpose(2, 3).reshape(N, self.out_channels * 2, H, W, self.kernel_size, self.kernel_size) + k, v = kv.split(self.out_channels, dim=1) + k_h, k_w = k.reshape(N, self.groups, self.group_channels, H, W, self.kernel_size, self.kernel_size).split(self.group_channels // 2, dim=2) + kr = torch.cat((k_h + self.relative_h, k_w + self.relative_w), dim=2) + + q = q.view(N, self.groups, self.group_channels, H, W, 1) + kr = kr.view(N, self.groups, self.group_channels, H, W, -1) + v = v.view(N, self.groups, self.group_channels, H, W, -1) # N, g, C//g, H, W, kxk + + qkr = (q * kr).sum(dim=2) + qkr = self.softmax(qkr).unsqueeze(2) # N, g, 1, H, W, kxk + out = (qkr * v).sum(dim=-1).reshape(N, self.out_channels, H, W) + + if self.stride > 1: + out = self.pooling(out) + + return out + + def reset_parameters(self): + n = self.in_channels + self.conv_qkv.weight.data.normal_(0, math.sqrt(1. / n)) + n = self.group_channels + nn.init.normal_(self.relative_h, 0, math.sqrt(1. / n)) + nn.init.normal_(self.relative_w, 0, math.sqrt(1. / n)) + + +class SelfAttentionBlock(nn.Module): + expansion = 4 + + def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1, + base_width=64, dilation=1, norm_layer=None): + super(SelfAttentionBlock, self).__init__() + if norm_layer is None: + norm_layer = nn.BatchNorm2d + width = int(planes * (base_width / 64.)) + # Both self.conv2 and self.downsample layers downsample the input when stride != 1 + self.conv1 = conv1x1(inplanes, width) + self.bn1 = norm_layer(width) + self.conv2 = SelfAttention2d(width, width, kernel_size=7, stride=stride, padding=3, groups=8) + self.bn2 = norm_layer(width) + self.conv3 = conv1x1(width, planes * self.expansion) + self.bn3 = norm_layer(planes * self.expansion) + self.relu = nn.ReLU(inplace=True) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + identity = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample is not None: + identity = self.downsample(x) + + out += identity + out = self.relu(out) + + return out + + +class SelfAttentionNet(nn.Module): + + def __init__(self, block, layers, num_classes=1000, zero_init_residual=True, + groups=8, width_per_group=64, replace_stride_with_dilation=None, + norm_layer=None): + super(SelfAttentionNet, self).__init__() + if norm_layer is None: + norm_layer = nn.BatchNorm2d + self._norm_layer = norm_layer + + self.inplanes = 64 + self.dilation = 1 + if replace_stride_with_dilation is None: + # each element in the tuple indicates if we should replace + # the 2x2 stride with a dilated convolution instead + replace_stride_with_dilation = [False, False, False] + if len(replace_stride_with_dilation) != 3: + raise ValueError("replace_stride_with_dilation should be None " + "or a 3-element tuple, got {}".format(replace_stride_with_dilation)) + self.groups = groups + self.base_width = width_per_group + self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, + bias=False) + self.bn1 = norm_layer(self.inplanes) + self.relu = nn.ReLU(inplace=True) + self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + self.layer1 = self._make_layer(block, 64, layers[0]) + self.layer2 = self._make_layer(block, 128, layers[1], stride=2, + dilate=replace_stride_with_dilation[0]) + self.layer3 = self._make_layer(block, 256, layers[2], stride=2, + dilate=replace_stride_with_dilation[1]) + self.layer4 = self._make_layer(block, 512, layers[3], stride=2, + dilate=replace_stride_with_dilation[2]) + self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) + self.fc = nn.Linear(512 * block.expansion, num_classes) + + for m in self.modules(): + if isinstance(m, nn.Conv2d): + if isinstance(m, conv_qkv): + pass + else: + nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') + elif isinstance(m, (nn.BatchNorm2d, nn.BatchNorm1d, nn.GroupNorm)): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + + # Zero-initialize the last BN in each residual branch, + # so that the residual branch starts with zeros, and each residual block behaves like an identity. + # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 + if zero_init_residual: + for m in self.modules(): + if isinstance(m, SelfAttentionBlock): + nn.init.constant_(m.bn2.weight, 0) + + def _make_layer(self, block, planes, blocks, stride=1, dilate=False): + norm_layer = self._norm_layer + downsample = None + previous_dilation = self.dilation + if dilate: + self.dilation *= stride + stride = 1 + if stride != 1 or self.inplanes != planes * block.expansion: + downsample = nn.Sequential( + conv1x1(self.inplanes, planes * block.expansion, stride), + norm_layer(planes * block.expansion), + ) + + layers = [] + layers.append(block(self.inplanes, planes, stride, downsample, groups=self.groups, + base_width=self.base_width, dilation=previous_dilation, + norm_layer=norm_layer)) + self.inplanes = planes * block.expansion + for _ in range(1, blocks): + layers.append(block(self.inplanes, planes, groups=self.groups, + base_width=self.base_width, dilation=self.dilation, + norm_layer=norm_layer)) + + return nn.Sequential(*layers) + + def _forward_impl(self, x): + # See note [TorchScript super()] + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + x = self.maxpool(x) + + x = self.layer1(x) + x = self.layer2(x) + x = self.layer3(x) + x = self.layer4(x) + + x = self.avgpool(x) + x = torch.flatten(x, 1) + x = self.fc(x) + + return x + + def forward(self, x): + return self._forward_impl(x) + + +def sa_resnet50(**kwargs): + model = SelfAttentionNet(SelfAttentionBlock, [3, 4, 6, 3], **kwargs) + return model +