model/model_architectures.py

"""
File: model.py
Author: Elena Ryumina and Dmitry Ryumin
Description: This module provides model architectures.
License: MIT License
"""

import torch
import torch.nn as  nn
import torch.nn.functional as F
import math

class Bottleneck(nn.Module):
    expansion = 4
    def __init__(self, in_channels, out_channels, i_downsample=None, stride=1):
        super(Bottleneck, self).__init__()
        
        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride, padding=0, bias=False)
        self.batch_norm1 = nn.BatchNorm2d(out_channels, eps=0.001, momentum=0.99)
        
        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding='same', bias=False)
        self.batch_norm2 = nn.BatchNorm2d(out_channels, eps=0.001, momentum=0.99)
        
        self.conv3 = nn.Conv2d(out_channels, out_channels*self.expansion, kernel_size=1, stride=1, padding=0, bias=False)
        self.batch_norm3 = nn.BatchNorm2d(out_channels*self.expansion, eps=0.001, momentum=0.99)
        
        self.i_downsample = i_downsample
        self.stride = stride
        self.relu = nn.ReLU()
        
    def forward(self, x):
        identity = x.clone()
        x = self.relu(self.batch_norm1(self.conv1(x)))
        
        x = self.relu(self.batch_norm2(self.conv2(x)))
        
        x = self.conv3(x)
        x = self.batch_norm3(x)
        
        #downsample if needed
        if self.i_downsample is not None:
            identity = self.i_downsample(identity)
        #add identity
        x+=identity
        x=self.relu(x)
        
        return x

class Conv2dSame(torch.nn.Conv2d):

    def calc_same_pad(self, i: int, k: int, s: int, d: int) -> int:
        return max((math.ceil(i / s) - 1) * s + (k - 1) * d + 1 - i, 0)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        ih, iw = x.size()[-2:]

        pad_h = self.calc_same_pad(i=ih, k=self.kernel_size[0], s=self.stride[0], d=self.dilation[0])
        pad_w = self.calc_same_pad(i=iw, k=self.kernel_size[1], s=self.stride[1], d=self.dilation[1])

        if pad_h > 0 or pad_w > 0:
            x = F.pad(
                x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2]
            )
        return F.conv2d(
            x,
            self.weight,
            self.bias,
            self.stride,
            self.padding,
            self.dilation,
            self.groups,
        )

class ResNet(nn.Module):
    def __init__(self, ResBlock, layer_list, num_classes, num_channels=3):
        super(ResNet, self).__init__()
        self.in_channels = 64

        self.conv_layer_s2_same = Conv2dSame(num_channels, 64, 7, stride=2, groups=1, bias=False)
        self.batch_norm1 = nn.BatchNorm2d(64, eps=0.001, momentum=0.99)
        self.relu = nn.ReLU()
        self.max_pool = nn.MaxPool2d(kernel_size = 3, stride=2)
        
        self.layer1 = self._make_layer(ResBlock, layer_list[0], planes=64, stride=1)
        self.layer2 = self._make_layer(ResBlock, layer_list[1], planes=128, stride=2)
        self.layer3 = self._make_layer(ResBlock, layer_list[2], planes=256, stride=2)
        self.layer4 = self._make_layer(ResBlock, layer_list[3], planes=512, stride=2)
        
        self.avgpool = nn.AdaptiveAvgPool2d((1,1))
        self.fc1 = nn.Linear(512*ResBlock.expansion, 512)
        self.relu1 = nn.ReLU()
        self.fc2 = nn.Linear(512, num_classes)

    def extract_features(self, x):
        x = self.relu(self.batch_norm1(self.conv_layer_s2_same(x)))
        x = self.max_pool(x)
        # print(x.shape)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        
        x = self.avgpool(x)
        x = x.reshape(x.shape[0], -1)
        x = self.fc1(x)
        return x
        
    def forward(self, x):
        x = self.extract_features(x)
        x = self.relu1(x)
        x = self.fc2(x)
        return x
        
    def _make_layer(self, ResBlock, blocks, planes, stride=1):
        ii_downsample = None
        layers = []
        
        if stride != 1 or self.in_channels != planes*ResBlock.expansion:
            ii_downsample = nn.Sequential(
                nn.Conv2d(self.in_channels, planes*ResBlock.expansion, kernel_size=1, stride=stride, bias=False, padding=0),
                nn.BatchNorm2d(planes*ResBlock.expansion, eps=0.001, momentum=0.99)
            )
            
        layers.append(ResBlock(self.in_channels, planes, i_downsample=ii_downsample, stride=stride))
        self.in_channels = planes*ResBlock.expansion
        
        for i in range(blocks-1):
            layers.append(ResBlock(self.in_channels, planes))
            
        return nn.Sequential(*layers)
        
def ResNet50(num_classes, channels=3):
    return ResNet(Bottleneck, [3,4,6,3], num_classes, channels)


class LSTMPyTorch(nn.Module):
    def __init__(self):
        super(LSTMPyTorch, self).__init__()
        
        self.lstm1 = nn.LSTM(input_size=512, hidden_size=512, batch_first=True, bidirectional=False)
        self.lstm2 = nn.LSTM(input_size=512, hidden_size=256, batch_first=True, bidirectional=False)
        self.fc = nn.Linear(256, 7)
        self.softmax = nn.Softmax(dim=1)

    def forward(self, x):
        x, _ = self.lstm1(x)
        x, _ = self.lstm2(x)        
        x = self.fc(x[:, -1, :])
        x = self.softmax(x)
        return x