Masterdqqq/Facial_Expression_Recognition

1"""2File: model.py3Author: Elena Ryumina and Dmitry Ryumin4Description: This module provides model architectures.5License: MIT License6"""7 8import torch9import torch.nn as  nn10import torch.nn.functional as F11import math12 13class Bottleneck(nn.Module):14    expansion = 415    def __init__(self, in_channels, out_channels, i_downsample=None, stride=1):16        super(Bottleneck, self).__init__()17        18        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride, padding=0, bias=False)19        self.batch_norm1 = nn.BatchNorm2d(out_channels, eps=0.001, momentum=0.99)20        21        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding='same', bias=False)22        self.batch_norm2 = nn.BatchNorm2d(out_channels, eps=0.001, momentum=0.99)23        24        self.conv3 = nn.Conv2d(out_channels, out_channels*self.expansion, kernel_size=1, stride=1, padding=0, bias=False)25        self.batch_norm3 = nn.BatchNorm2d(out_channels*self.expansion, eps=0.001, momentum=0.99)26        27        self.i_downsample = i_downsample28        self.stride = stride29        self.relu = nn.ReLU()30        31    def forward(self, x):32        identity = x.clone()33        x = self.relu(self.batch_norm1(self.conv1(x)))34        35        x = self.relu(self.batch_norm2(self.conv2(x)))36        37        x = self.conv3(x)38        x = self.batch_norm3(x)39        40        #downsample if needed41        if self.i_downsample is not None:42            identity = self.i_downsample(identity)43        #add identity44        x+=identity45        x=self.relu(x)46        47        return x48 49class Conv2dSame(torch.nn.Conv2d):50 51    def calc_same_pad(self, i: int, k: int, s: int, d: int) -> int:52        return max((math.ceil(i / s) - 1) * s + (k - 1) * d + 1 - i, 0)53 54    def forward(self, x: torch.Tensor) -> torch.Tensor:55        ih, iw = x.size()[-2:]56 57        pad_h = self.calc_same_pad(i=ih, k=self.kernel_size[0], s=self.stride[0], d=self.dilation[0])58        pad_w = self.calc_same_pad(i=iw, k=self.kernel_size[1], s=self.stride[1], d=self.dilation[1])59 60        if pad_h > 0 or pad_w > 0:61            x = F.pad(62                x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2]63            )64        return F.conv2d(65            x,66            self.weight,67            self.bias,68            self.stride,69            self.padding,70            self.dilation,71            self.groups,72        )73 74class ResNet(nn.Module):75    def __init__(self, ResBlock, layer_list, num_classes, num_channels=3):76        super(ResNet, self).__init__()77        self.in_channels = 6478 79        self.conv_layer_s2_same = Conv2dSame(num_channels, 64, 7, stride=2, groups=1, bias=False)80        self.batch_norm1 = nn.BatchNorm2d(64, eps=0.001, momentum=0.99)81        self.relu = nn.ReLU()82        self.max_pool = nn.MaxPool2d(kernel_size = 3, stride=2)83        84        self.layer1 = self._make_layer(ResBlock, layer_list[0], planes=64, stride=1)85        self.layer2 = self._make_layer(ResBlock, layer_list[1], planes=128, stride=2)86        self.layer3 = self._make_layer(ResBlock, layer_list[2], planes=256, stride=2)87        self.layer4 = self._make_layer(ResBlock, layer_list[3], planes=512, stride=2)88        89        self.avgpool = nn.AdaptiveAvgPool2d((1,1))90        self.fc1 = nn.Linear(512*ResBlock.expansion, 512)91        self.relu1 = nn.ReLU()92        self.fc2 = nn.Linear(512, num_classes)93 94    def extract_features(self, x):95        x = self.relu(self.batch_norm1(self.conv_layer_s2_same(x)))96        x = self.max_pool(x)97        # print(x.shape)98        x = self.layer1(x)99        x = self.layer2(x)100        x = self.layer3(x)101        x = self.layer4(x)102        103        x = self.avgpool(x)104        x = x.reshape(x.shape[0], -1)105        x = self.fc1(x)106        return x107        108    def forward(self, x):109        x = self.extract_features(x)110        x = self.relu1(x)111        x = self.fc2(x)112        return x113        114    def _make_layer(self, ResBlock, blocks, planes, stride=1):115        ii_downsample = None116        layers = []117        118        if stride != 1 or self.in_channels != planes*ResBlock.expansion:119            ii_downsample = nn.Sequential(120                nn.Conv2d(self.in_channels, planes*ResBlock.expansion, kernel_size=1, stride=stride, bias=False, padding=0),121                nn.BatchNorm2d(planes*ResBlock.expansion, eps=0.001, momentum=0.99)122            )123            124        layers.append(ResBlock(self.in_channels, planes, i_downsample=ii_downsample, stride=stride))125        self.in_channels = planes*ResBlock.expansion126        127        for i in range(blocks-1):128            layers.append(ResBlock(self.in_channels, planes))129            130        return nn.Sequential(*layers)131        132def ResNet50(num_classes, channels=3):133    return ResNet(Bottleneck, [3,4,6,3], num_classes, channels)134 135 136class LSTMPyTorch(nn.Module):137    def __init__(self):138        super(LSTMPyTorch, self).__init__()139        140        self.lstm1 = nn.LSTM(input_size=512, hidden_size=512, batch_first=True, bidirectional=False)141        self.lstm2 = nn.LSTM(input_size=512, hidden_size=256, batch_first=True, bidirectional=False)142        self.fc = nn.Linear(256, 7)143        self.softmax = nn.Softmax(dim=1)144 145    def forward(self, x):146        x, _ = self.lstm1(x)147        x, _ = self.lstm2(x)        148        x = self.fc(x[:, -1, :])149        x = self.softmax(x)150        return x