import math import torch.nn as nn import torch.nn.functional as F import torch.utils.model_zoo as model_zoo import torch.utils.checkpoint as cp def conv1x1(in_planes, out_planes, stride=1): return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) def conv3x3(in_planes, out_planes, stride=1): "3x3 convolution with padding" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None, with_cp=False): super(BasicBlock, self).__init__() self.conv1 = conv1x1(inplanes, planes) self.bn1 = nn.BatchNorm2d(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes, stride) self.bn2 = nn.BatchNorm2d(planes) self.downsample = downsample self.stride = stride self.with_cp = with_cp def forward(self, x): def _inner_forward(x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual return out if self.with_cp and x.requires_grad: out = cp.checkpoint(_inner_forward, x) else: out = _inner_forward(x) out = self.relu(out) return out class ResNet(nn.Module): def __init__(self, block, layers, output_channels=512): super(ResNet, self).__init__() channels = [output_channels//(2**i) for i in reversed(range(5))] self.inplanes = channels[0] self.conv1 = nn.Conv2d(3, channels[0], kernel_size=3, stride=1, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(channels[0]) self.relu = nn.ReLU(inplace=True) self.layer1 = self._make_layer(block, channels[0], layers[0], stride=2) self.layer2 = self._make_layer(block, channels[1], layers[1], stride=1) self.layer3 = self._make_layer(block, channels[2], layers[2], stride=2) self.layer4 = self._make_layer(block, channels[3], layers[3], stride=1) self.layer5 = self._make_layer(block, channels[4], layers[4], stride=1) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x, extra_feats=None): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) if extra_feats is not None: if extra_feats[0].shape[1]>0: x = x+F.interpolate(extra_feats[0], x.shape[2:], mode='nearest') x = self.layer1(x) if extra_feats is not None: if extra_feats[1].shape[1]>0: x = x+F.interpolate(extra_feats[1], x.shape[2:], mode='nearest') x = self.layer2(x) if extra_feats is not None: if extra_feats[2].shape[1]>0: x = x+F.interpolate(extra_feats[2], x.shape[2:], mode='nearest') x = self.layer3(x) if extra_feats is not None: if extra_feats[3].shape[1]>0: x = x+F.interpolate(extra_feats[3], x.shape[2:], mode='nearest') x = self.layer4(x) if extra_feats is not None: if extra_feats[4].shape[1]>0: x = x+F.interpolate(extra_feats[4], x.shape[2:], mode='nearest') x = self.layer5(x) if extra_feats is not None: if extra_feats[5].shape[1]>0: x = x+F.interpolate(extra_feats[5], x.shape[2:], mode='nearest') return x def resnet45(alpha_d, output_channels=512): layers = [int(round(x*alpha_d)) for x in [3, 4, 6, 6, 3]] return ResNet(BasicBlock, layers, output_channels=output_channels)