pytorch-training

PyTorch Training Best Practice

1. Use torch.manual_seed() for reproducibility (set for torch, numpy, random)
2. Use DataLoader with num_workers>0 and pin_memory=True for GPU
3. Enable cudnn.benchmark=True for fixed input sizes
4. Use learning rate schedulers (CosineAnnealingLR or OneCycleLR)
5. Implement early stopping based on validation metric
6. Log metrics every epoch, save best model checkpoint
7. Use torch.no_grad() for evaluation
8. Clear gradients with optimizer.zero_grad(set_to_none=True) for efficiency