resnet.py

import torch
import torch.nn as nn

from functools import partial
from dataclasses import dataclass
from collections import OrderedDict

class Conv2dAuto(nn.Conv2d):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.padding =  (self.kernel_size[0] // 2, self.kernel_size[1] // 2)
conv3x3 = partial(Conv2dAuto, kernel_size=3, bias=False)

def activation_func(activation):
    return  nn.ModuleDict([
        ['relu', nn.ReLU(inplace=True)],
        ['leaky_relu', nn.LeakyReLU(negative_slope=0.01, inplace=True)],
        ['selu', nn.SELU(inplace=True)],
        ['none', nn.Identity()]
    ])[activation]
def conv_bn(in_channels, out_channels, conv, *args, **kwargs):
    return nn.Sequential(conv(in_channels, out_channels, *args, **kwargs), nn.BatchNorm2d(out_channels))

class ResidualBlock(nn.Module):
    def __init__(self, in_channels, out_channels, activation='relu'):
        super().__init__()
        self.in_channels, self.out_channels, self.activation = in_channels, out_channels, activation
        self.blocks = nn.Identity()
        self.activate = activation_func(activation)
        self.shortcut = nn.Identity()   
    
    def forward(self, x):
        residual = x
        if self.should_apply_shortcut: residual = self.shortcut(x)
        x = self.blocks(x)
        x += residual
        x = self.activate(x)
        return x
    
    @property
    def should_apply_shortcut(self):
        return self.in_channels != self.out_channels

class ResNetResidualBlock(ResidualBlock):
    def __init__(self, in_channels, out_channels, expansion=1, downsampling=1, conv=conv3x3, *args, **kwargs):
        super().__init__(in_channels, out_channels)
        self.expansion, self.downsampling, self.conv = expansion, downsampling, conv
        self.shortcut = nn.Sequential(
            nn.Conv2d(self.in_channels, self.expanded_channels, kernel_size=1,
                      stride=self.downsampling, bias=False),
            nn.BatchNorm2d(self.expanded_channels)) if self.should_apply_shortcut else None   
        
    @property
    def expanded_channels(self):
        return self.out_channels * self.expansion
    
    @property
    def should_apply_shortcut(self):
        return self.in_channels != self.expanded_channels


class ResNetBasicBlock(ResNetResidualBlock):
    expansion = 1
    def __init__(self, in_channels, out_channels, *args, **kwargs):
        super().__init__(in_channels, out_channels, *args, **kwargs)
        self.blocks = nn.Sequential(
            conv_bn(self.in_channels, self.out_channels, conv=self.conv, bias=False, stride=self.downsampling),
            activation_func(self.activation),
            conv_bn(self.out_channels, self.expanded_channels, conv=self.conv, bias=False),
        )
        

        
class ResNetLayer(nn.Module):
    def __init__(self, in_channels, out_channels, block=ResNetBasicBlock, n=1, *args, **kwargs):
        super().__init__()
        # 'We perform downsampling directly by convolutional layers that have a stride of 2.'
        downsampling = 2 if in_channels != out_channels else 1
        
        self.blocks = nn.Sequential(
            block(in_channels , out_channels, *args, **kwargs, downsampling=downsampling),
            *[block(out_channels * block.expansion, 
                    out_channels, downsampling=1, *args, **kwargs) for _ in range(n - 1)]
        )

    def forward(self, x):
        x = self.blocks(x)
        return x
    
class ResNetEncoder(nn.Module):
    """
    ResNet encoder composed by increasing different layers with increasing features.
    """
    def __init__(self, in_channels=3, blocks_sizes=[32, 32, 64, 64], depths=[2,2,2,2], 
                 activation='relu', block=ResNetBasicBlock, *args, **kwargs):
        super().__init__()
        
        self.blocks_sizes = blocks_sizes
        
        self.gate = nn.Sequential(
            nn.Conv2d(in_channels, self.blocks_sizes[0], kernel_size=(5,7), stride=2, padding=3, bias=False),
            nn.BatchNorm2d(self.blocks_sizes[0]),
            activation_func(activation),
            nn.MaxPool2d(kernel_size=(4,6), stride=(2,3), padding=1)
        )
        
        self.in_out_block_sizes = list(zip(blocks_sizes, blocks_sizes[1:]))
        self.blocks = nn.ModuleList([ 
            ResNetLayer(blocks_sizes[0], blocks_sizes[0], n=depths[0], activation=activation, 
                        block=block,*args, **kwargs),
            *[ResNetLayer(in_channels * block.expansion, 
                          out_channels, n=n, activation=activation, 
                          block=block, *args, **kwargs) 
              for (in_channels, out_channels), n in zip(self.in_out_block_sizes, depths[1:])]       
        ])
        
        
    def forward(self, x):
        x = self.gate(x)
        for block in self.blocks:
            x = block(x)
        return x
    


def ResnetEncoderModel(in_channels,blocks_sizes=[16,32,64,32,16], depths=[2,2,2,2,2], activation='relu'):
    return ResNetEncoder(in_channels,blocks_sizes=blocks_sizes,depths=depths,activation=activation)