ML-Intern-lab/mnist-60s-cnn

MNIST CNN — 60-second training run

A small CNN trained from scratch on MNIST for 60 seconds of training compute (wall-clock capped loop), reaching 99.24% test accuracy on the held-out 10k test split.

Details

• Architecture: Conv(1→32) → ReLU → Conv(32→64) → ReLU → MaxPool → Dropout(0.25) → FC(12544→128) → ReLU → Dropout(0.5) → FC(128→10) — 1,625,866 parameters
• Dataset: ylecun/mnist, train split (60,000 images)
• Optimizer: Adam, lr 0.001, batch size 256
• Training steps: 4,691 (78 steps/sec) in 60s
• Seed: 0

Files

• pytorch_model.bin — model state dict (load with torch.load(..., map_location="cpu")["state_dict"])