CNN implementation for image classification on CIFAR-100 using PyTorch.

Test Accuracy: 61%
Training: 20 epochs on Google Colab GPU
Dataset: CIFAR-100 (100 classes, 32×32 RGB images)

Techniques used:

• Batch normalisation for training stability
• Leaky ReLU activation to prevent dying neurons
• Data augmentation with random crops and horizontal flips
• Adam optimiser with 0.001 learning rate
• Dropout regularisation to reduce overfitting

Model accuracy improved from 11% to 61% over 20 epochs through iterative refinements.

Conv2d(3→64) + BatchNorm + LeakyReLU + MaxPool
Conv2d(64→128) + BatchNorm + LeakyReLU + MaxPool
Conv2d(128→256) + BatchNorm + LeakyReLU
Conv2d(256→256) + BatchNorm + LeakyReLU + MaxPool
Linear(4096→512) + LeakyReLU + Dropout(0.3)
Linear(512→100)

jupyter notebook CompVision_A2.ipynb

Requirements: PyTorch, torchvision, CUDA-compatible GPU