from collections import namedtuple import torch import torch.nn.functional as F from torch.nn import Conv2d, BatchNorm2d, PReLU, ReLU, Sigmoid, MaxPool2d, AdaptiveAvgPool2d, Sequential, Module """ ArcFace implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch) """ class Flatten(Module): def forward(self, input): return input.view(input.size(0), -1) def l2_norm(input, axis=1): norm = torch.norm(input, 2, axis, True) output = torch.div(input, norm) return output class Bottleneck(namedtuple('Block', ['in_channel', 'depth', 'stride'])): """ A named tuple describing a ResNet block. """ def get_block(in_channel, depth, num_units, stride=2): return [Bottleneck(in_channel, depth, stride)] + [Bottleneck(depth, depth, 1) for i in range(num_units - 1)] def get_blocks(num_layers): if num_layers == 50: blocks = [ get_block(in_channel=64, depth=64, num_units=3), get_block(in_channel=64, depth=128, num_units=4), get_block(in_channel=128, depth=256, num_units=14), get_block(in_channel=256, depth=512, num_units=3) ] elif num_layers == 100: blocks = [ get_block(in_channel=64, depth=64, num_units=3), get_block(in_channel=64, depth=128, num_units=13), get_block(in_channel=128, depth=256, num_units=30), get_block(in_channel=256, depth=512, num_units=3) ] elif num_layers == 152: blocks = [ get_block(in_channel=64, depth=64, num_units=3), get_block(in_channel=64, depth=128, num_units=8), get_block(in_channel=128, depth=256, num_units=36), get_block(in_channel=256, depth=512, num_units=3) ] else: raise ValueError("Invalid number of layers: {}. Must be one of [50, 100, 152]".format(num_layers)) return blocks class SEModule(Module): def __init__(self, channels, reduction): super(SEModule, self).__init__() self.avg_pool = AdaptiveAvgPool2d(1) self.fc1 = Conv2d(channels, channels // reduction, kernel_size=1, padding=0, bias=False) self.relu = ReLU(inplace=True) self.fc2 = Conv2d(channels // reduction, channels, kernel_size=1, padding=0, bias=False) seltorch.sigmoid = Sigmoid() def forward(self, x): module_input = x x = self.avg_pool(x) x = self.fc1(x) x = self.relu(x) x = self.fc2(x) x = seltorch.sigmoid(x) return module_input * x class bottleneck_IR(Module): def __init__(self, in_channel, depth, stride): super(bottleneck_IR, self).__init__() if in_channel == depth: self.shortcut_layer = MaxPool2d(1, stride) else: self.shortcut_layer = Sequential( Conv2d(in_channel, depth, (1, 1), stride, bias=False), BatchNorm2d(depth) ) self.res_layer = Sequential( BatchNorm2d(in_channel), Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), PReLU(depth), Conv2d(depth, depth, (3, 3), stride, 1, bias=False), BatchNorm2d(depth) ) def forward(self, x): shortcut = self.shortcut_layer(x) res = self.res_layer(x) return res + shortcut class bottleneck_IR_SE(Module): def __init__(self, in_channel, depth, stride): super(bottleneck_IR_SE, self).__init__() if in_channel == depth: self.shortcut_layer = MaxPool2d(1, stride) else: self.shortcut_layer = Sequential( Conv2d(in_channel, depth, (1, 1), stride, bias=False), BatchNorm2d(depth) ) self.res_layer = Sequential( BatchNorm2d(in_channel), Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), PReLU(depth), Conv2d(depth, depth, (3, 3), stride, 1, bias=False), BatchNorm2d(depth), SEModule(depth, 16) ) def forward(self, x): shortcut = self.shortcut_layer(x) res = self.res_layer(x) return res + shortcut def _upsample_add(x, y): """Upsample and add two feature maps. Args: x: (Variable) top feature map to be upsampled. y: (Variable) lateral feature map. Returns: (Variable) added feature map. Note in PyTorch, when input size is odd, the upsampled feature map with `F.upsample(..., scale_factor=2, mode='nearest')` maybe not equal to the lateral feature map size. e.g. original input size: [N,_,15,15] -> conv2d feature map size: [N,_,8,8] -> upsampled feature map size: [N,_,16,16] So we choose bilinear upsample which supports arbitrary output sizes. """ _, _, H, W = y.size() return F.interpolate(x, size=(H, W), mode='bilinear', align_corners=True) + y