use-case-and-architecture/EdgeFLite/architecture/resnet.py

# -*- coding: utf-8 -*-
# @Author: Weisen Pan

import torch
import torch.nn as nn

# Try to import the method to load model weights from a URL, with a fallback in case of ImportError
try:
    from torch.hub import load_state_dict_from_url
except ImportError:
    from torch.utils.model_zoo import load_url as load_state_dict_from_url

# List of available ResNet architectures
__all__ = ['resnet_model_18', 'resnet_model_34', 'resnet_model_50',
           'resnet_model_101', 'resnet_model_152', 'resnet_model_200',
           'resnet110', 'resnet164',
           'resnext29_8x64d', 'resnext29_16x64d',
           'resnext50_32x4d', 'resnext101_32x4d',
           'resnext101_32x8d', 'resnext101_64x4d',
           'wide_resnet_model_50_2', 'wide_resnet_model_50_3', 'wide_resnet_model_101_2',
           'wide_resnet16_8', 'wide_resnet52_8', 'wide_resnet16_12',
           'wide_resnet28_10', 'wide_resnet40_10']

# Pre-trained model URLs for various ResNet variants
model_urls = {
    'resnet_model_18': 'https://download.pytorch.org/models/resnet_model_18-5c106cde.pth',
    'resnet_model_34': 'https://download.pytorch.org/models/resnet_model_34-333f7ec4.pth',
    'resnet_model_50': 'https://download.pytorch.org/models/resnet_model_50-19c8e357.pth',
    'resnet_model_101': 'https://download.pytorch.org/models/resnet_model_101-5d3b4d8f.pth',
    'resnet_model_152': 'https://download.pytorch.org/models/resnet_model_152-b121ed2d.pth',
    'resnext50_32x4d': 'https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth',
    'resnext101_32x8d': 'https://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pth',
    'wide_resnet_model_50_2': 'https://download.pytorch.org/models/wide_resnet_model_50_2-95faca4d.pth',
    'wide_resnet_model_101_2': 'https://download.pytorch.org/models/wide_resnet_model_101_2-32ee1156.pth',
}

# Function for a 3x3 convolution with padding
def apply_3x3_convolution(in_planes, out_planes, stride=1, groups=1, dilation=1):
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=dilation, groups=groups, bias=False, dilation=dilation)

# Function for a 1x1 convolution
def apply_1x1_convolution(in_planes, out_planes, stride=1):
    """1x1 convolution"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)

# BasicBlock class for the ResNet architecture
class BasicBlock(nn.Module):
    expansion = 1  # Expansion factor for the output channels

    def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1,
                 base_width=64, dilation=1, norm_layer=None):
        super(BasicBlock, self).__init__()
        # If norm_layer is not provided, use BatchNorm2d as the default
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        # Ensure BasicBlock is restricted to specific parameters
        if groups != 1 or base_width != 64:
            raise ValueError('BasicBlock is restricted to groups=1 and base_width=64')
        if dilation > 1:
            raise NotImplementedError("BasicBlock does not support dilation greater than 1")
        
        # Define the layers for the BasicBlock
        self.conv1 = apply_3x3_convolution(inplanes, planes, stride)  # First 3x3 convolution
        self.bn1 = norm_layer(planes)  # First BatchNorm layer
        self.relu = nn.ReLU(inplace=True)  # ReLU activation
        self.conv2 = apply_3x3_convolution(planes, planes)  # Second 3x3 convolution
        self.bn2 = norm_layer(planes)  # Second BatchNorm layer
        self.downsample = downsample  # Optional downsample layer
        self.stride = stride

    # Define the forward pass for BasicBlock
    def forward(self, x):
        identity = x  # Save the input for the skip connection

        out = self.conv1(x)  # First convolution
        out = self.bn1(out)  # BatchNorm after first convolution
        out = self.relu(out)  # ReLU activation

        out = self.conv2(out)  # Second convolution
        out = self.bn2(out)  # BatchNorm after second convolution

        # Apply downsample if defined
        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity  # Add the skip connection
        out = self.relu(out)  # Apply ReLU activation again

        return out

# Bottleneck class for the ResNet architecture, a more complex block used in deeper ResNet models
class Bottleneck(nn.Module):
    expansion = 4  # Expansion factor for the output channels

    def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1,
                 base_width=64, dilation=1, norm_layer=None):
        super(Bottleneck, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        width = int(planes * (base_width / 64.)) * groups  # Calculate width based on base width and groups

        # Define the layers for the Bottleneck block
        self.conv1 = apply_1x1_convolution(inplanes, width)  # 1x1 convolution to reduce the dimensions
        self.bn1 = norm_layer(width)  # BatchNorm after 1x1 convolution
        self.conv2 = apply_3x3_convolution(width, width, stride, groups, dilation)  # 3x3 convolution
        self.bn2 = norm_layer(width)  # BatchNorm after 3x3 convolution
        self.conv3 = apply_1x1_convolution(width, planes * self.expansion)  # 1x1 convolution to expand the dimensions
        self.bn3 = norm_layer(planes * self.expansion)  # BatchNorm after final 1x1 convolution
        self.relu = nn.ReLU(inplace=True)  # ReLU activation
        self.downsample = downsample  # Optional downsample layer
        self.stride = stride

    # Define the forward pass for Bottleneck
    def forward(self, x):
        identity = x  # Save the input for the skip connection

        out = self.conv1(x)  # First convolution
        out = self.bn1(out)  # BatchNorm after first convolution
        out = self.relu(out)  # ReLU activation

        out = self.conv2(out)  # Second convolution
        out = self.bn2(out)  # BatchNorm after second convolution
        out = self.relu(out)  # ReLU activation

        out = self.conv3(out)  # Third convolution
        out = self.bn3(out)  # BatchNorm after third convolution

        # Apply downsample if defined
        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity  # Add the skip connection
        out = self.relu(out)  # Apply ReLU activation again

        return out

# Main ResNet class, a customizable deep learning model architecture
class ResNet(nn.Module):

    def __init__(self, arch, block, layers, num_classes=1000, zero_init_residual=True,
                 groups=1, width_per_group=64, replace_stride_with_dilation=None,
                 norm_layer=None, dataset='cifar10', split_factor=1, output_stride=8, dropout_p=None):
        super(ResNet, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d  # Default normalization layer
        self._norm_layer = norm_layer

        self.groups = groups  # Number of groups in convolutions
        self.inplanes = 16 if dataset in ['cifar10', 'cifar100'] else 64  # Adjust initial planes for CIFAR

        # First layer: a combination of convolution, normalization, and ReLU
        self.layer0 = nn.Sequential(
            nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False),
            norm_layer(self.inplanes),
            nn.ReLU(inplace=True),
        )
        
        # Subsequent ResNet layers using the _create_model_layer method
        self.layer1 = self._create_model_layer(block, 16, layers[0])
        self.layer2 = self._create_model_layer(block, 32, layers[1], stride=2)
        self.layer3 = self._create_model_layer(block, 64, layers[2], stride=2)
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))  # Global average pooling
        self.fc = nn.Linear(64 * block.expansion, num_classes)  # Fully connected layer for classification

        # Initialization for model weights
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)
            elif isinstance(m, nn.Linear):
                nn.init.normal_(m.weight, 0, 1e-3)

        # Zero-initialize the last BatchNorm in residual connections if required
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)

    # Helper function to create layers in ResNet
    def _create_model_layer(self, block, planes, blocks, stride=1, dilate=False):
        norm_layer = self._norm_layer  # Set normalization layer
        downsample = None
        # If the stride is not 1 or input/output planes do not match, create a downsample layer
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                apply_1x1_convolution(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )
        layers = [block(self.inplanes, planes, stride, downsample)]  # Create the first block with downsampling
        self.inplanes = planes * block.expansion  # Update inplanes for next blocks
        for _ in range(1, blocks):
            layers.append(block(self.inplanes, planes))  # Add subsequent blocks without downsampling
        return nn.Sequential(*layers)

    # Forward pass through the ResNet architecture
    def forward(self, x):
        x = self.layer0(x)  # Pass input through the first layer
        x = self.layer1(x)  # First ResNet layer
        x = self.layer2(x)  # Second ResNet layer
        x = self.layer3(x)  # Third ResNet layer
        x = self.avgpool(x)  # Global average pooling
        x = torch.flatten(x, 1)  # Flatten the output for the fully connected layer
        x = self.fc(x)  # Pass through the fully connected layer
        return x

# Helper function to instantiate ResNet with pretrained weights if available
def _resnet(arch, block, layers, models_pretrained, progress, **kwargs):
    model = ResNet(arch, block, layers, **kwargs)  # Create a ResNet model
    if models_pretrained:  # Load pretrained weights if requested
        state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)
        model.load_state_dict(state_dict)
    return model

# Functions to create specific ResNet variants
def resnet_model_18(models_pretrained=False, progress=True, **kwargs):
    return _resnet('resnet_model_18', BasicBlock, [2, 2, 2, 2], models_pretrained, progress, **kwargs)

def resnet_model_34(models_pretrained=False, progress=True, **kwargs):
    return _resnet('resnet_model_34', BasicBlock, [3, 4, 6, 3], models_pretrained, progress, **kwargs)

def resnet_model_50(models_pretrained=False, progress=True, **kwargs):
    return _resnet('resnet_model_50', Bottleneck, [3, 4, 6, 3], models_pretrained, progress, **kwargs)

def resnet_model_101(models_pretrained=False, progress=True, **kwargs):
    return _resnet('resnet_model_101', Bottleneck, [3, 4, 23, 3], models_pretrained, progress, **kwargs)

def resnet_model_152(models_pretrained=False, progress=True, **kwargs):
    return _resnet('resnet_model_152', Bottleneck, [3, 8, 36, 3], models_pretrained, progress, **kwargs)

def resnet_model_200(models_pretrained=False, progress=True, **kwargs):
    return _resnet('resnet_model_200', Bottleneck, [3, 24, 36, 3], models_pretrained, progress, **kwargs)