ksuit.callbacks.base.periodic_callback
======================================

.. py:module:: ksuit.callbacks.base.periodic_callback


Attributes
----------

.. autoapisummary::

   ksuit.callbacks.base.periodic_callback.IntervalType


Classes
-------

.. autoapisummary::

   ksuit.callbacks.base.periodic_callback.PeriodicCallback


Module Contents
---------------

.. py:data:: IntervalType

.. py:class:: PeriodicCallback(callback_config, trainer, model, data_container, tracker, log_writer, checkpoint_writer, metric_property_provider, name = None)

   Bases: :py:obj:`ksuit.callbacks.base.callback_base.CallbackBase`


   Base class for callbacks that are invoked periodically during training. Epoch, update and sample based intervals
   are supported.
   ```

   Initializes the `PeriodicCallback`.

   :param callback_config: Configuration of the `PeriodicCallback`. Implements the `CallBackBaseConfig` schema.
   :param trainer: Trainer of the current run, subclass of `SgdTrainer`.
   :param model: Model of the current run.
   :param data_container: DataContainer instance that provides access to all datasets.
   :param tracker: Tracker instance to log metrics to stdout/disk/online platform.
   :param log_writer: LogWriter instance to log metrics.
   :param checkpoint_writer: CheckpointWriter instance to save checkpoints.
   :param metric_property_provider: MetricPropertyProvider instance to access properties of metrics.
   :param name: Name of the callback.


   .. py:attribute:: every_n_epochs


   .. py:attribute:: every_n_updates


   .. py:attribute:: every_n_samples


   .. py:attribute:: batch_size


   .. py:method:: should_log_after_epoch(training_iteration)

      Checks after every epoch if the `PeriodicCallback` should be invoked.

      :param training_iteration: TrainingIteration to check.



   .. py:method:: should_log_after_update(training_iteration)

      Checks after every update if the `PeriodicCallback` should be invoked.

      :param training_iteration: TrainingIteration to check.



   .. py:method:: should_log_after_sample(training_iteration, effective_batch_size)

      Checks after every sample if the `PeriodicCallback` should be invoked.

      :param training_iteration: TrainingIteration to check.
      :param effective_batch_size: Effective batch size to use for the check (required for checking how many samples
                                   have been processed since the last update).



   .. py:method:: track_after_accumulation_step(*, update_counter, batch, losses, update_outputs, accumulation_steps, accumulation_step)

      Invoked after every gradient accumulation step. May be used to track metrics. Applies `torch.no_grad()`.

      :param update_counter: UpdateCounter instance to access current training progress.
      :param batch: Current batch.
      :param losses: Losses computed for the current batch.
      :param update_outputs: Outputs of the model for the current batch.
      :param accumulation_steps: Total number of accumulation steps.
      :param accumulation_step: Current accumulation step.



   .. py:method:: track_after_update_step(update_counter, times)

      Invoked after every update step. May be used to track metrics. Applies `torch.no_grad()`.

      :param update_counter: UpdateCounter instance to access current training progress.
      :param times: Dictionary containing time measurements.



   .. py:method:: after_epoch(update_counter, **kwargs)

      Invoked after every epoch to check if callback should be invoked. Applies `torch.no_grad()`.

      :param update_counter: UpdateCounter instance to access current training progress.



   .. py:method:: after_update(update_counter, **kwargs)

      Invoked after every update to check if callback should be invoked. Applies `torch.no_grad()`.

      :param update_counter: UpdateCounter instance to access current training progress.



   .. py:method:: updates_till_next_log(update_counter)

      Calculates how many updates remain until this callback is invoked.

      :param update_counter: UpdateCounter instance to access current training progress.

      :returns: Number of updates remaining until the next callback invocation.



   .. py:method:: updates_per_log_interval(update_counter)

      Calculates how many updates are from one invocation of this callback to the next.

      :param update_counter: UpdateCounter instance to access current training progress.

      :returns: Number of updates between callback invocations.



   .. py:method:: get_interval_string_verbose()

      Returns `every_n_epochs`, `every_n_updates` or `every_n_samples` depending on which one is not None.
      :returns: Interval as, e.g., "every_n_epochs=1" for epoch-based intervals.
      :rtype: str



   .. py:method:: to_short_interval_string()

      Returns `every_n_epochs`, `every_n_updates` or `every_n_samples` depending on which one is not None.
      :returns: Interval as, e.g., "E1" if `every_n_epochs=1` for epoch-based intervals.
      :rtype: str