Custom Convolutional Neural Network Models

This repository contains custom implementations of convolutional neural network (CNN) models including inception, exception, vgg, and resnet designed for binary classification tasks. Each model has been tailored with specific architectures and adapt to various classification challenges.

Models Overview

Custom Inception Model

• Architecture: Custom inception-like model with basic components.
• Key Components:
  • Initial Convolution: Conv2d(3, 16, kernel_size=3, stride=1, padding=1)
  • Inception Block:
    • Branch 1: Conv2d(16, 16, kernel_size=3, stride=1, padding=1), followed by ReLU
    • Branch 2: MaxPool2d(kernel_size=3, stride=1, padding=1) followed by Conv2d(16, 16, kernel_size=1, stride=1) and ReLU
  • Dropout Layers: Dropout(p=0.3) applied twice
  • Flatten Layer: Converts tensor to 1D
  • Fully Connected Layer: Linear(in_features=1605632, out_features=2)

Custom ResNet Model

• Architecture: Custom ResNet with basic residual blocks.
• Key Components:
  • Residual Block:
    • Convolutions: Conv2d(3, 64, kernel_size=3, padding=1), Conv2d(64, 64, kernel_size=3, padding=1)
    • Batch Normalization: Applied after convolutions
    • ReLU Activation: Applied after normalization
    • Shortcut Connection: Adjusted dimensions with Conv2d(3, 64, kernel_size=1, stride=1) if needed
  • Dropout Layers: Dropout(p=0.5) applied twice
  • Pooling: AdaptiveAvgPool2d(output_size=(1, 1))
  • Fully Connected Layer: Linear(in_features=64, out_features=2)

Custom VGG Model

• Architecture: Custom VGG-like model.
• Key Components:
  • Convolutions:
    • Conv2d(3, 64, kernel_size=3, stride=1, padding=1), followed by ReLU and MaxPool2d(kernel_size=2, stride=2)
    • Conv2d(64, 128, kernel_size=3, stride=1, padding=1), followed by ReLU and MaxPool2d(kernel_size=2, stride=2)
  • Dropout Layers: Dropout(p=0.5) applied twice
  • Flatten Layer: Converts tensor to 1D
  • Fully Connected Layer: Linear(in_features=401408, out_features=2)

Custom Xception Model

• Architecture: Simplified version of Xception.
• Key Components:
  • Depthwise Separable Convolution:
    • Depthwise Convolution: Conv2d(3, 3, kernel_size=3, stride=2, padding=1, groups=3)
    • Pointwise Convolution: Conv2d(3, 64, kernel_size=1, stride=1)
  • Activation and Normalization: ReLU followed by BatchNorm2d(64)
  • Pooling: MaxPool2d(kernel_size=3, stride=2, padding=1)
  • Dropout Layers: Dropout(p=0.5) applied twice
  • Fully Connected Layer: Linear(in_features=200704, out_features=2)

Contribution

Feel free to fork the repository, make modifications, and submit pull requests. Contributions and suggestions are welcome.