Training

From the previous tutorials, you may now have a custom model and data loader.

You are free to create your own optimizer, and write the training logic: it’s usually easy with PyTorch, and allow researchers to see the entire training logic more clearly. One such example is provided in tools/plain_train_net.py.

We also provide a standarized “trainer” abstraction with a minimal hook system that helps simplify the standard types of training.

You can use SimpleTrainer().train() which does single-cost single-optimizer single-data-source training. Or use DefaultTrainer().train() which includes more standard behavior that one might want to opt in.