PyTorch
Specific best practices and architectural patterns when working with PyTorch.
Details
Language / Topic
PythonCategory
framework
Rules
balanced
PyTorch
- Always use `torch.no_grad()` during inference to save memory. Move tensors to the same device explicitly — `tensor.to(device)`. Use `DataLoader` with `num_workers>0` and `pin_memory=True` for GPU training.
PyTorch
- Always use `torch.no_grad()` during inference to save memory. Move tensors to the same device explicitly — `tensor.to(device)`. Use `DataLoader` with `num_workers>0` and `pin_memory=True` for GPU training.
- Use `nn.Module` subclasses for all models — implement `forward()` only, never `__call__`. Use `torch.amp.autocast` and `GradScaler` for mixed-precision training. Save model state dicts (`model.state_dict()`), not entire models. Set `model.eval()` before inference and `model.train()` before training. Use `torch.utils.data.Dataset` for custom data — implement `__len__` and `__getitem__`.