CIFAR-10 CNN

A deep learning pipeline for image classification on the CIFAR-10 dataset, built to explore how architectural choices and training strategies impact performance.

The project progresses through 4 distinct versions: a basic 3-layer CNN (81.43%), a deeper 6-layer CNN (87.43%), a ResNet-18 transfer (90.99%), and a fully optimized ResNet-18 with advanced data augmentation reaching 92.47% accuracy.

Training uses AdamW optimizer with CosineAnnealing learning rate scheduling, label smoothing (0.1), and dropout regularization. Data augmentation includes random crops, horizontal flips, color jittering, and rotation.

The codebase is fully modular — separate files for dataset loading, model definitions, training loops, and evaluation. TensorBoard integration tracks loss curves, accuracy, learning rates, and weight distributions in real time.

Supports multi-backend training: NVIDIA CUDA, CPU, and Apple Silicon MPS. Automatic best-model checkpointing ensures the best weights are always saved.