detectron/docs/tutorials/training.md

• Training

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.