import math import torch.nn as nn from collections import OrderedDict import torch.utils.model_zoo as model_zoo from torchvision.models.resnet import model_urls from .base import Backbone class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None, dilation=1): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=dilation, bias=False, dilation=dilation) self.bn2 = nn.BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * 4) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = 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: residual = self.downsample(x) out += residual out = self.relu(out) return out class ResNet(Backbone): """ ResNet network module. Allows extracting specific feature blocks.""" def __init__(self, block, layers, output_layers, num_classes=1000, inplanes=64, dilation_factor=1, frozen_layers=()): self.inplanes = inplanes super(ResNet, self).__init__(frozen_layers=frozen_layers) self.output_layers = output_layers self.conv1 = nn.Conv2d(3, inplanes, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(inplanes) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) stride = [1 + (dilation_factor < l) for l in (8, 4, 2)] self.layer1 = self._make_layer(block, inplanes, layers[0], dilation=max(dilation_factor//8, 1)) self.layer2 = self._make_layer(block, inplanes*2, layers[1], stride=stride[0], dilation=max(dilation_factor//4, 1)) self.layer3 = self._make_layer(block, inplanes*4, layers[2], stride=stride[1], dilation=max(dilation_factor//2, 1)) self.layer4 = self._make_layer(block, inplanes*8, layers[3], stride=stride[2], dilation=dilation_factor) out_feature_strides = {'conv1': 4, 'layer1': 4, 'layer2': 4*stride[0], 'layer3': 4*stride[0]*stride[1], 'layer4': 4*stride[0]*stride[1]*stride[2]} # TODO better way? if isinstance(self.layer1[0], Bottleneck): base_num_channels = 4 * inplanes out_feature_channels = {'conv1': inplanes, 'layer1': base_num_channels, 'layer2': base_num_channels * 2, 'layer3': base_num_channels * 4, 'layer4': base_num_channels * 8} else: raise Exception('block not supported') self._out_feature_strides = out_feature_strides self._out_feature_channels = out_feature_channels # self.avgpool = nn.AvgPool2d(7, stride=1) self.avgpool = nn.AdaptiveAvgPool2d((1,1)) self.fc = nn.Linear(inplanes*8 * block.expansion, num_classes) 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, dilation=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, dilation=dilation)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def _add_output_and_check(self, name, x, outputs, output_layers): if name in output_layers: outputs[name] = x return len(output_layers) == len(outputs) def forward(self, x, output_layers=None): """ Forward pass with input x. The output_layers specify the feature blocks which must be returned """ outputs = OrderedDict() if output_layers is None: output_layers = self.output_layers x = self.conv1(x) x = self.bn1(x) x = self.relu(x) if self._add_output_and_check('conv1', x, outputs, output_layers): return outputs x = self.maxpool(x) x = self.layer1(x) if self._add_output_and_check('layer1', x, outputs, output_layers): return outputs x = self.layer2(x) if self._add_output_and_check('layer2', x, outputs, output_layers): return outputs x = self.layer3(x) if self._add_output_and_check('layer3', x, outputs, output_layers): return outputs x = self.layer4(x) if self._add_output_and_check('layer4', x, outputs, output_layers): return outputs x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.fc(x) if self._add_output_and_check('fc', x, outputs, output_layers): return outputs if len(output_layers) == 1 and output_layers[0] == 'default': return x raise ValueError('output_layer is wrong.') def resnet50(output_layers=None, pretrained=False, **kwargs): """Constructs a ResNet-50 model. """ if output_layers is None: output_layers = ['default'] else: for l in output_layers: if l not in ['conv1', 'layer1', 'layer2', 'layer3', 'layer4', 'fc']: raise ValueError('Unknown layer: {}'.format(l)) model = ResNet(Bottleneck, [3, 4, 6, 3], output_layers, **kwargs) if pretrained: model.load_state_dict(model_zoo.load_url(model_urls['resnet50'])) return model