CNN Image Classification

Designed configurable CNN architectures in PyTorch driven by a layer spec — Conv2d → BatchNorm → ReLU → pooling blocks with automatically computed 'same' padding and Dropout2d regularization.

Built the full training pipeline end to end: dataset loading, the training/validation loop, checkpointing, and a Kaggle-style submission generator.

Iterated from a baseline to an improved network (deeper blocks, batch normalization, dropout) and tracked the resulting gains in validation accuracy.