Trainer Presets and Tasks

Trainer Presets and Tasks#

Trainer presets reduce repetitive fit keyword boilerplate.

trainer = trainer.with_fit_defaults(reduce: :mean, monitor_mode: :min)
trainer.fit(train_data, epochs: 1)

Fit defaults and named presets#

  • with_fit_defaults(**defaults)

  • register_fit_preset(name, **defaults)

  • fit_with(name, dataset, **overrides)

  • fit_report_with(name, dataset, **overrides)

trainer = trainer.with_fit_defaults(reduce: :mean, monitor_mode: :min)
trainer.register_fit_preset(:fast, epochs: 3, limit: 128)
trainer.fit_with(:fast, train_data)

Precedence:

  • explicit call overrides

  • preset defaults

  • trainer defaults

  • trainer intrinsic defaults

Task presets#

  • register_task(name, **defaults)

  • fit_task(task, dataset, **overrides)

  • fit_task_report(task, dataset, **overrides)

trainer.register_task(:image_cls, monitor: :accuracy, monitor_mode: :max)
report = trainer.fit_task_report(
  :image_cls,
  train_data,
  validation_data: validation_data,
  epochs: 10
)
puts report[:best_monitor]

Built-ins include:

  • :classification

  • :regression

  • :language_modeling (includes a perplexity-style monitor metric)

trainer = trainer.with_fit_defaults(reduce: :mean, monitor_mode: :min)
trainer.register_fit_preset(:fast, epochs: 3, limit: 128)

report = trainer.fit_report_with(:fast, train_data, validation_data: val_data)

See implementation:

  • lib/mlx/dsl/trainer.rb

Contents