Table of Contents
Shortcuts

Setting the Frequency of Logging, Checkpoint Saving, and Validation

MMEngine supports two training modes, EpochBased based on epochs and IterBased based on the number of iterations. Both of these modes are used in downstream algorithm libraries such as MMDetection, which uses the EpochBased mode by default, and MMSegmentation, which uses the IterBased mode by default.

In different training modes, the semantics of the interval in MMEngine will be different. In EpochBased mode, the interval is in terms of epochs, while in IterBased mode, the interval is in terms of iterations.

Setting the Interval for Training and Validation

To customize the interval for training and validation, set the val_interval parameter in the initialization parameter train_cfg of Runner.

  • EpochBased

In EpochBased mode, the default value of val_interval is 1, which means to validate once after training an epoch.

runner = Runner(
    model=MMResNet50(),
    work_dir='./work_dir',
    train_dataloader=train_dataloader,
    optim_wrapper=dict(optimizer=dict(type=SGD, lr=0.001, momentum=0.9)),
    train_cfg=dict(by_epoch=True, max_epochs=5, val_interval=1),
    val_dataloader=val_dataloader,
    val_cfg=dict(),
    val_evaluator=dict(type=Accuracy),
)
runner.train()

  • IterBased

In IterBased mode, the default value of val_interval is 1000, which means to validate once after training 1000 iterations.

runner = Runner(
    model=MMResNet50(),
    work_dir='./work_dir',
    train_dataloader=train_dataloader,
    optim_wrapper=dict(optimizer=dict(type=SGD, lr=0.001, momentum=0.9)),
    train_cfg=dict(by_epoch=False, max_iters=10000, val_interval=2000),
    val_dataloader=val_dataloader,
    val_cfg=dict(),
    val_evaluator=dict(type=Accuracy),
)
runner.train()

Setting the Interval for Saving Checkpoints

To customize the interval for saving checkpoints, set the interval parameter of CheckpointHook.

  • EpochBased

In EpochBased mode, the default value of interval is 1, which means to save checkpoints once after training for one epoch.

# set the interval to 2, which means to save checkpoints every 2 epochs
default_hooks = dict(checkpoint=dict(type='CheckpointHook', interval=2))
runner = Runner(
    model=MMResNet50(),
    work_dir='./work_dir',
    train_dataloader=train_dataloader,
    optim_wrapper=dict(optimizer=dict(type=SGD, lr=0.001, momentum=0.9)),
    train_cfg=dict(by_epoch=True, max_epochs=5, val_interval=1),
    val_dataloader=val_dataloader,
    val_cfg=dict(),
    val_evaluator=dict(type=Accuracy),
    default_hooks=default_hooks,
)
runner.train()

  • IterBased

By default, checkpoints are saved in terms of epochs. If you want to save checkpoints in terms of iterations, you need to set by_epoch=False.

# set by_epoch=False and interval=500, which means to save checkpoints every 500 iterations
default_hooks = dict(checkpoint=dict(type='CheckpointHook', by_epoch=False, interval=500))
runner = Runner(
    model=MMResNet50(),
    work_dir='./work_dir',
    train_dataloader=train_dataloader,
    optim_wrapper=dict(optimizer=dict(type=SGD, lr=0.001, momentum=0.9)),
    train_cfg=dict(by_epoch=False, max_iters=10000, val_interval=1000),
    val_dataloader=val_dataloader,
    val_cfg=dict(),
    val_evaluator=dict(type=Accuracy),
    default_hooks=default_hooks,
)
runner.train()

For more information on how to use CheckpointHook, please refer to the CheckpointHook tutorial.

Setting the Interval for Printing Logs

By default, logs are printed to the terminal once every 10 iterations. You can set the interval using the interval parameter of the LoggerHook.

# print logs every 20 iterations
default_hooks = dict(logger=dict(type='LoggerHook', interval=20))
runner = Runner(
    model=MMResNet50(),
    work_dir='./work_dir',
    train_dataloader=train_dataloader,
    optim_wrapper=dict(optimizer=dict(type=SGD, lr=0.001, momentum=0.9)),
    train_cfg=dict(by_epoch=True, max_epochs=5, val_interval=1),
    val_dataloader=val_dataloader,
    val_cfg=dict(),
    val_evaluator=dict(type=Accuracy),
    default_hooks=default_hooks,
)
runner.train()

For more information on how to use LoggerHook, please refer to the LoggerHook tutorial.