noether.core.callbacks.checkpoint.ema

Classes

EmaCallbackConfig

Internal base class for all registry-based configs.

EmaCallback

Callback for exponential moving average (EMA) of model weights.

Module Contents

class noether.core.callbacks.checkpoint.ema.EmaCallbackConfig(/, **data)

Bases: noether.core.callbacks.base.CallBackBaseConfig

Internal base class for all registry-based configs.

Provides auto-registration via __init_subclass__. Not meant to be used directly - use specific config base classes instead.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

Parameters:

data (Any)

name: Literal['EmaCallback'] = None
target_factors: list[float] = None

The factors for the EMA.

model_paths: list[str | None] | None = None

The paths to the models to apply the EMA to (i.e., composite_model.encoder/composite_model.decoder, path of the PyTorch nn.Modules in the checkpoint). If None, the EMA is applied to the whole model. When training with a CompositeModel, the paths on the submodules (i.e., ‘encoder’, ‘decoder’, etc.) should be provided via this field, otherwise the EMA will be applied to the CompositeModel as a whole which is not possible to restore later on.

save_weights: bool = None

Whether to save the EMA weights.

save_last_weights: bool = None

Save the weights of the model when training is over (i.e., at the end of training, save the EMA weights).

save_latest_weights: bool = None

Save the latest EMA weights. Note that the latest weights are always overwritten on the next invocation of this callback.

eval_callbacks: list[Annotated[Any, Discriminated(CallBackBaseConfig)]] | None = None

Optional nested periodic callbacks to run against EMA weights. Each child retains its own schedule (every_n_epochs etc.); the EMA callback swaps its stored EMA parameters into the live model around eval-time hooks (after_epoch, after_update, at_eval) and restores the live weights on exit. Children are dispatched once per target_factor and their metric keys are automatically prefixed with ema=<factor>/ to avoid collisions with live-model metrics. Note: before_training and after_training are forwarded without swapping, so EMA initialization and the final save see live weights.

class noether.core.callbacks.checkpoint.ema.EmaCallback(callback_config, **kwargs)

Bases: noether.core.callbacks.periodic.PeriodicCallback

Callback for exponential moving average (EMA) of model weights.

In addition to maintaining and checkpointing EMA weights, this callback can optionally own a list of child evaluation callbacks via eval_callbacks. At each eval-time hook (after_epoch, after_update, at_eval) the EMA weights are swapped into the live model, the children are dispatched, and the live weights are restored. Children are dispatched once per target_factor and their metric keys are automatically prefixed with ema=<factor>/ to avoid collisions with live-model metrics.

Example config:

- kind: noether.core.callbacks.EmaCallback
  every_n_epochs: 10
  save_weights: false
  save_last_weights: false
  save_latest_weights: true
  target_factors:
    - 0.9999
  name: EmaCallback
  eval_callbacks:
    - kind: noether.training.callbacks.OfflineLossCallback
      every_n_epochs: 1
      dataset_key: val
Parameters:

  • callback_config (EmaCallbackConfig) – Configuration for the callback. See EmaCallbackConfig for available options.

  • **kwargs – Additional arguments passed to the parent class.

model_paths
target_factors
save_weights
save_last_weights
save_latest_weights
parameters: dict[tuple[str | None, float], dict[str, torch.Tensor]]
buffers: dict[str | None, dict[str, torch.Tensor]]
eval_callbacks: dict[float, list[noether.core.callbacks.base.CallbackBase]]
get_children()

Flat list of child eval callbacks (across all target_factors).

Exposed to the trainer so nested PeriodicDataIteratorCallback instances have their samplers registered on the shared data loader. The EMA callback remains responsible for dispatching lifecycle hooks to its children.

Return type:

list[noether.core.callbacks.base.CallbackBase]

resume_from_checkpoint(resumption_paths, model)

Resume EMA state from a checkpoint.

Tries cp=latest first (written by periodic saves), then cp=last (written by after_training, e.g. on graceful signal interrupt). If neither exists, falls back to initializing EMA from the current model weights.

Parameters:
Return type:

None

before_training(**kwargs)

Hook called once before the training loop starts.

This method is intended to be overridden by derived classes to perform initialization tasks before training begins. Common use cases include:

  • Initializing experiment tracking (e.g., logging hyperparameters)

  • Printing model summaries or architecture details

  • Initializing specific data structures or buffers needed during training

  • Performing sanity checks on the data or configuration

Note

This method is executed within a torch.no_grad() context.

Parameters:

update_counter – UpdateCounter instance to access current training progress.

Return type:

None

apply_ema(cur_model, model_path, target_factor)

fused in-place implementation

track_after_update_step(**_)

Hook called after each optimizer update step.

This method is invoked after a successful optimizer step and parameter update. It is typically used for tracking metrics that should be recorded once per update cycle, such as:

  • Latest loss values

  • Learning rates

  • Model parameter statistics (norms, etc.)

  • Training throughput and timing measurements

Unlike periodic_callback(), this hook is called on every update step, making it suitable for maintaining running averages or high-frequency telemetry.

Note

This method is executed within a torch.no_grad() context.

Parameters:

  • update_counter – UpdateCounter instance to access current training progress.

  • times – Dictionary containing time measurements for various parts of the training step (e.g., ‘data_time’, ‘forward_time’, ‘backward_time’, ‘update_time’).

Return type:

None

after_epoch(update_counter, **kwargs)

Invoked after every epoch to check if callback should be invoked.

Applies torch.no_grad() context.

Parameters:
Return type:

None

after_update(update_counter, **kwargs)

Invoked after every update to check if callback should be invoked.

Applies torch.no_grad() context.

Parameters:
Return type:

None

at_eval(update_counter, **kwargs)
Parameters:

update_counter (noether.core.utils.training.counter.UpdateCounter)

Return type:

None

periodic_callback(*, interval_type, update_counter, **_)

Hook called periodically based on the configured intervals.

This method is the primary entry point for periodic actions in subclasses. It is triggered when any of the configured intervals (every_n_epochs, every_n_updates, or every_n_samples) are reached.

Subclasses should override this method to implement periodic logic such as:

  • Calculating and logging expensive validation metrics

  • Saving specific model checkpoints or artifacts

  • Visualizing training progress (e.g., plotting samples)

  • Adjusting training hyperparameters or model state

Note

This method is executed within a torch.no_grad() context.

Parameters:

  • interval_type (noether.core.callbacks.periodic.IntervalType) – “epoch”, “update”, “sample” or “eval” indicating which interval triggered this callback.

  • update_counter – UpdateCounter instance providing details about the current training progress (epoch, update, sample counts).

  • **kwargs – Additional keyword arguments passed from the triggering hook (e.g., from after_epoch() or after_update()).

Return type:

None

after_training(**kwargs)

Hook called once after the training loop finishes.

This method is intended to be overridden by derived classes to perform cleanup or final reporting tasks after training is complete. Common use cases include:

  • Performing a final evaluation on the test set

  • Saving final model weights or artifacts

  • Sending notifications (e.g., via Slack or email) about the completed run

  • Closing or finalizing experiment tracking sessions

Note

This method is executed within a torch.no_grad() context.

Parameters:

update_counter – UpdateCounter instance to access current training progress.

Return type:

None