ausjahk/animal-classifier-cnn-from-scratch

Animal Image Classifier (CNN Trained From Scratch)

Overview

Custom 4-Block Convolutional Neural Network (CNN) designed and trained entirely from scratch (random weight initialization) on the Kaggle Cat & Dog dataset.

Architecture

• 4 Convolutional Blocks (32 → 64 → 128 → 256 filters)
• BatchNorm + ReLU + MaxPool2d
• Dense Classifier Head with Dropout (0.5)
• Explicit Kaiming Normal Weight Initialization
• Total Parameters: 4,584,450 (Zero pre-trained weights)

Dataset & Performance

• Dataset: Kaggle Cat and Dog Dataset
• Best Test Accuracy: 67.82%