ResNet_for_CC.py

import torch
import torch.nn as nn
import torchvision.models as models

class ResClassifier(nn.Module):
    """
    A classifier with two fully connected layers followed by a final linear layer.
    Uses BatchNorm, ReLU activations, and Dropout for better generalization.
    """
    def __init__(self, num_classes=14):
        super(ResClassifier, self).__init__()
        
        # First fully connected layer: reduces 128D features to 64D
        self.fc1 = nn.Sequential(
            nn.Linear(128, 64),
            nn.BatchNorm1d(64, affine=True),
            nn.ReLU(inplace=True),
            nn.Dropout()
        )
        
        # Second fully connected layer: retains 64D features
        self.fc2 = nn.Sequential(
            nn.Linear(64, 64),
            nn.BatchNorm1d(64, affine=True),
            nn.ReLU(inplace=True),
            nn.Dropout()
        )
        
        # Final classification layer mapping 64D features to class logits
        self.fc3 = nn.Linear(64, num_classes)

    def forward(self, x):
        """
        Forward pass through the classifier.
        Returns class logits after two hidden layers.
        """
        x = self.fc1(x)  # First FC layer
        x = self.fc2(x)  # Second FC layer
        output = self.fc3(x)  # Final classification layer
        return output


class CC_model(nn.Module):
    """
    Clothing Classification Model based on ResNet50.
    Extracts deep features and uses two independent classifiers for predictions.
    """
    def __init__(self, num_classes1=14, num_classes2=None):
        super(CC_model, self).__init__()
        
        # If num_classes2 is not specified, default to num_classes1
        num_classes2 = num_classes2 if num_classes2 else num_classes1
        assert num_classes1 == num_classes2  # Ensure both classifiers predict the same categories
        
        self.num_classes = num_classes1
        
        # Load a pretrained ResNet-50 model as the feature extractor
        self.model_resnet = models.resnet50(weights='ResNet50_Weights.DEFAULT')
        
        # Remove ResNet's original classification layer to use as a feature extractor
        num_ftrs = self.model_resnet.fc.in_features
        self.model_resnet.fc = nn.Identity()  # Identity layer keeps feature dimensions

        # Additional transformation layer reducing feature size to 128D
        self.dr = nn.Linear(num_ftrs, 128)

        # Two independent classifiers
        self.fc1 = ResClassifier(num_classes1)
        self.fc2 = ResClassifier(num_classes1)

    def forward(self, x, detach_feature=False):
        """
        Forward pass through the model.
        Extracts deep features from ResNet and processes them through classifiers.
        """
        with torch.no_grad():
            # Extract deep features using ResNet-50 (without its original classification head)
            feature = self.model_resnet(x)

        # Generate transformed features (128D) using the custom linear layer
        dr_feature = self.dr(feature)

        if detach_feature:
            dr_feature = dr_feature.detach()  # Detach feature for non-trainable forward pass

        # Pass features through two independent classifiers
        out1 = self.fc1(dr_feature)
        out2 = self.fc2(dr_feature)

        # Compute the mean prediction from both classifiers
        output_mean = (out1 + out2) / 2

        return dr_feature, output_mean  # Returning feature embeddings and final prediction