Add custom Train loop

VISSL implements a default training loop (single iteration step) that is used for self-supervised training of all VISSL reference approaches, for feature extraction and for supervised workflows. Users can implement their own training loop.

The training loop performs: data read, forward, loss computation, backward, optimizer step, parameter updates.

Various intermediate steps are also performed:

  • logging the training loss, training eta, LR, etc to loggers.

  • logging metrics to tensorboard.

  • performing any self-supervised method specific operations (like in MoCo approach, the momentum encoder is updated), computing the scores in swav.

  • checkpointing model if user wants to checkpoint in the middle of an epoch.

Users can implement their custom training loop by following the steps:

  • Step1: Create your my_new_training_loop module under vissl/trainer/train_steps/ following the template:

from vissl.trainer.train_steps import register_train_step

def my_new_training_loop(task):
    add documentation on what this training loop does and how it varies from
    standard training loop in vissl.
    # implement the training loop. It should take care of running the dataloader
    # iterator to get the input sample

    return task
  • Step2: New train loop is ready to use. Set the TRAINER.TRAIN_STEP_NAME=my_new_training_loop