Configurable weight initialization with lightning>=2.6.

jakob-schloer · jakob-schloer · commit 89d10ca5b4b7 · 2025-12-15T09:18:05.000Z
diff --git a/training/docs/user-guide/training.rst b/training/docs/user-guide/training.rst
@@ -570,3 +570,50 @@ frozen and only the encoder and decoder will be trained:
 Freezing can be particularly beneficial in scenarios such as fine-tuning
 when only specific components (e.g., the encoder, the decoder) need to
 adapt to a new task while keeping others (e.g., the processor) fixed.
+
+******************
+ Weight Averaging
+******************
+
+Weight averaging is a technique to improve model generalization by
+averaging model weights during training. Anemoi Training supports weight
+averaging methods through PyTorch Lightning callbacks:
+
+-  **Exponential Moving Average (EMA)**: Maintains an exponential moving
+      average of model weights, which can lead to smoother convergence
+      and better generalization.
+
+      .. code:: yaml
+
+         weight_averaging:
+            _target_: pytorch_lightning.callbacks.EMAWeightAveraging
+            decay: 0.999
+
+      The ``decay`` parameter (typically between 0.99 and 0.9999)
+      controls the smoothing factor. Higher values give more weight to
+      historical weights, resulting in a more stable average. By
+      default, the decay is set to 0.999.
+
+-  **Stochastic Weight Averaging (SWA)**: Averages weights from multiple
+      points along the training trajectory, typically resulting in wider
+      optima and improved generalization.
+
+      .. code:: yaml
+
+         weight_averaging:
+            _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
+            swa_lrs: 1.e-4
+
+      The ``swa_lrs`` parameter specifies the learning rate to use
+      during the SWA phase. By default, the learning rate is set to
+      1e-4. Additional parameters can be configured as described in the
+      [PyTorch Lightning
+      documentation](https://lightning.ai/docs/pytorch/latest/api/lightning.pytorch.callbacks.StochasticWeightAveraging.html#lightning.pytorch.callbacks.StochasticWeightAveraging)
+
+By default, weight averaging is disabled. To explicitly disable it or to
+override a parent configuration, set ``weight_averaging`` to null.
+
+.. note::
+
+   Weight averaging is only supported in PyTorch Lightning 2.6 and later
+   versions.
diff --git a/training/src/anemoi/training/config/training/default.yaml b/training/src/anemoi/training/config/training/default.yaml
@@ -31,11 +31,14 @@ gradient_clip:
   val: 32.
   algorithm: value
 
-# stochastic weight averaging
-# https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/
-swa:
-  enabled: False
-  lr: 1.e-4
+# weight averaging only supported in Lightning 2.6+
+weight_averaging: null
+# For EMA:
+#  _target_: pytorch_lightning.callbacks.EMAWeightAveraging
+#  decay: 0.999
+# For SWA:
+#  _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
+#  swa_lrs: 1.e-4
 
 # =====================================================================
 # Optimizer configuration
diff --git a/training/src/anemoi/training/config/training/diffusion.yaml b/training/src/anemoi/training/config/training/diffusion.yaml
@@ -31,11 +31,14 @@ gradient_clip:
   val: 1.
   algorithm: norm
 
-# stochastic weight averaging
-# https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/
-swa:
-  enabled: False
-  lr: 1.e-4
+# weight averaging only supported in Lightning 2.6+
+weight_averaging: null
+# For EMA:
+#  _target_: pytorch_lightning.callbacks.EMAWeightAveraging
+#  decay: 0.999
+# For SWA:
+#  _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
+#  swa_lrs: 1.e-4
 
 # Optimizer settings
 optimizer:
diff --git a/training/src/anemoi/training/config/training/ensemble.yaml b/training/src/anemoi/training/config/training/ensemble.yaml
@@ -31,11 +31,14 @@ gradient_clip:
   val: 32.
   algorithm: value
 
-# stochastic weight averaging
-# https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/
-swa:
-  enabled: False
-  lr: 1.e-4
+# weight averaging only supported in Lightning 2.6+
+weight_averaging: null
+# For EMA:
+#  _target_: pytorch_lightning.callbacks.EMAWeightAveraging
+#  decay: 0.999
+# For SWA:
+#  _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
+#  swa_lrs: 1.e-4
 
 # Optimizer settings
 optimizer:
diff --git a/training/src/anemoi/training/config/training/interpolator.yaml b/training/src/anemoi/training/config/training/interpolator.yaml
@@ -31,11 +31,14 @@ gradient_clip:
   val: 32.
   algorithm: value
 
-# stochastic weight averaging
-# https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/
-swa:
-  enabled: False
-  lr: 1.e-4
+# weight averaging only supported in Lightning 2.6+
+weight_averaging: null
+# For EMA:
+#  _target_: pytorch_lightning.callbacks.EMAWeightAveraging
+#  decay: 0.999
+# For SWA:
+#  _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
+#  swa_lrs: 1.e-4
 
 # Optimizer settings
 optimizer:
diff --git a/training/src/anemoi/training/config/training/lam.yaml b/training/src/anemoi/training/config/training/lam.yaml
@@ -31,11 +31,14 @@ gradient_clip:
   val: 32.
   algorithm: value
 
-# stochastic weight averaging
-# https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/
-swa:
-  enabled: False
-  lr: 1.e-4
+# weight averaging only supported in Lightning 2.6+
+weight_averaging: null
+# For EMA:
+#  _target_: pytorch_lightning.callbacks.EMAWeightAveraging
+#  decay: 0.999
+# For SWA:
+#  _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
+#  swa_lrs: 1.e-4
 
 # Optimizer settings
 optimizer:
diff --git a/training/src/anemoi/training/config/training/stretched.yaml b/training/src/anemoi/training/config/training/stretched.yaml
@@ -31,11 +31,14 @@ gradient_clip:
   val: 32.
   algorithm: value
 
-# stochastic weight averaging
-# https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/
-swa:
-  enabled: False
-  lr: 1.e-4
+# weight averaging only supported in Lightning 2.6+
+weight_averaging: null
+# For EMA:
+#  _target_: pytorch_lightning.callbacks.EMAWeightAveraging
+#  decay: 0.999
+# For SWA:
+#  _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
+#  swa_lrs: 1.e-4
 
 # Optimizer settings
 optimizer:
diff --git a/training/src/anemoi/training/diagnostics/callbacks/__init__.py b/training/src/anemoi/training/diagnostics/callbacks/__init__.py
@@ -23,7 +23,6 @@
 
 from anemoi.training.diagnostics.callbacks.checkpoint import AnemoiCheckpoint
 from anemoi.training.diagnostics.callbacks.optimiser import LearningRateMonitor
-from anemoi.training.diagnostics.callbacks.optimiser import StochasticWeightAveraging
 from anemoi.training.diagnostics.callbacks.provenance import ParentUUIDCallback
 from anemoi.training.diagnostics.callbacks.sanity import CheckVariableOrder
 from anemoi.training.schemas.base_schema import BaseSchema
@@ -49,7 +48,6 @@ def nestedget(config: DictConfig, key: str, default: Any) -> Any:
 # Callbacks to add according to flags in the config
 # Can be function to check status from config
 CONFIG_ENABLED_CALLBACKS: list[tuple[list[str] | str | Callable[[DictConfig], bool], type[Callback]]] = [
-    ("training.swa.enabled", StochasticWeightAveraging),
     (
         lambda config: nestedget(config, "diagnostics.log.wandb.enabled", False)
         or nestedget(config, "diagnostics.log.mlflow.enabled", False),
@@ -58,6 +56,39 @@ def nestedget(config: DictConfig, key: str, default: Any) -> Any:
 ]
 
 
+def _get_weight_averaging_callback(config: DictConfig) -> list[Callback]:
+    """Get weight averaging callback.
+
+    Supported are ExponentialMovingAverage and StochasticWeightAveraging.
+
+    Example config:
+        weight_averaging:
+          _target_: anemoi.training.diagnostics.callbacks.optimiser.ExponentialMovingAverage
+          decay: 0.999
+
+    Parameters
+    ----------
+    config : DictConfig
+        Job configuration
+
+    Returns
+    -------
+    list[Callback]
+        List containing the weight averaging callback, or empty list if not configured.
+    """
+    weight_averaging_config = nestedget(config, "training.weight_averaging", None)
+
+    if weight_averaging_config is not None:
+        try:
+            weight_averaging = instantiate(weight_averaging_config)
+            LOGGER.info("Using weight averaging: %s", type(weight_averaging))
+        except InstantiationException:
+            LOGGER.warning("Failed to instantiate weight averaging callback from config: %s", weight_averaging_config)
+        else:
+            return [weight_averaging]
+    return []
+
+
 def _get_checkpoint_callback(config: BaseSchema) -> list[AnemoiCheckpoint]:
     """Get checkpointing callbacks."""
     if not config.diagnostics.enable_checkpointing:
@@ -226,6 +257,9 @@ def get_callbacks(config: DictConfig) -> list[Callback]:
     # Plotting callbacks
     trainer_callbacks.extend(instantiate(callback, config) for callback in config.diagnostics.plot.callbacks)
 
+    # Weight averaging callback (SWA or EMA)
+    trainer_callbacks.extend(_get_weight_averaging_callback(config))
+
     # Extend with config enabled callbacks
     trainer_callbacks.extend(_get_config_enabled_callbacks(config))
 
diff --git a/training/src/anemoi/training/diagnostics/callbacks/optimiser.py b/training/src/anemoi/training/diagnostics/callbacks/optimiser.py
@@ -12,7 +12,6 @@
 
 from omegaconf import DictConfig
 from pytorch_lightning.callbacks import LearningRateMonitor as pl_LearningRateMonitor
-from pytorch_lightning.callbacks.stochastic_weight_avg import StochasticWeightAveraging as pl_StochasticWeightAveraging
 
 LOGGER = logging.getLogger(__name__)
 
@@ -28,46 +27,3 @@ def __init__(
     ) -> None:
         super().__init__(logging_interval=logging_interval, log_momentum=log_momentum)
         self.config = config
-
-
-class StochasticWeightAveraging(pl_StochasticWeightAveraging):
-    """Provide StochasticWeightAveraging from pytorch_lightning as a callback."""
-
-    def __init__(
-        self,
-        config: DictConfig,
-        swa_lrs: int | None = None,
-        swa_epoch_start: int | None = None,
-        annealing_epochs: int | None = None,
-        annealing_strategy: str | None = None,
-        device: str | None = None,
-        **kwargs,
-    ) -> None:
-        """Stochastic Weight Averaging Callback.
-
-        Parameters
-        ----------
-        config : OmegaConf
-            Full configuration object
-        swa_lrs : int, optional
-            Stochastic Weight Averaging Learning Rate, by default None
-        swa_epoch_start : int, optional
-            Epoch start, by default 0.75 * config.training.max_epochs
-        annealing_epochs : int, optional
-            Annealing Epoch, by default 0.25 * config.training.max_epochs
-        annealing_strategy : str, optional
-            Annealing Strategy, by default 'cos'
-        device : str, optional
-            Device to use, by default None
-        """
-        kwargs["swa_lrs"] = swa_lrs or config.training.swa.lr
-        kwargs["swa_epoch_start"] = swa_epoch_start or min(
-            int(0.75 * config.training.max_epochs),
-            config.training.max_epochs - 1,
-        )
-        kwargs["annealing_epochs"] = annealing_epochs or max(int(0.25 * config.training.max_epochs), 1)
-        kwargs["annealing_strategy"] = annealing_strategy or "cos"
-        kwargs["device"] = device
-
-        super().__init__(**kwargs)
-        self.config = config
diff --git a/training/src/anemoi/training/schemas/training.py b/training/src/anemoi/training/schemas/training.py
@@ -39,16 +39,42 @@ class GradientClip(BaseModel):
     "The gradient clipping algorithm to use"
 
 
-class SWA(BaseModel):
-    """Stochastic weight averaging configuration.
+class WeightAveragingSchema(BaseModel):
+    """Weight averaging configuration (SWA or EMA).
 
+    Uses Hydra instantiate pattern with _target_ to specify the callback class.
     See https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/
+
+    Example:
+        weight_averaging:
+          _target_: pytorch_lightning.callbacks.EMAWeightAveraging
+          decay: 0.999
+          update_starting_at_step: 1000
     """
 
-    enabled: bool = Field(example=False)
-    "Enable stochastic weight averaging."
-    lr: NonNegativeFloat = Field(example=1.0e-4)
-    "Learning rate for SWA."
+    target_: Literal[
+        "pytorch_lightning.callbacks.EMAWeightAveraging",
+        "pytorch_lightning.callbacks.StochasticWeightAveraging",
+    ] = Field(..., alias="_target_")
+    "Target callback class for weight averaging. Either EMAWeightAveraging or StochasticWeightAveraging."
+    # EMA specific
+    decay: NonNegativeFloat | None = Field(default=None)
+    "EMA decay rate (only used for EMAWeightAveraging)."
+    update_every_n_steps: PositiveInt = Field(default=1)
+    "Update every n steps (only used for EMAWeightAveraging)."
+    update_starting_at_step: PositiveInt | None = Field(default=None)
+    "Update starting at step (only used for EMAWeightAveraging)."
+    update_starting_at_epoch: PositiveInt | None = Field(default=None)
+    "Update starting at epoch (only used for EMAWeightAveraging)."
+    # SWA specific
+    swa_lrs: NonNegativeFloat | list[NonNegativeFloat] = Field(default=0.8)
+    "SWA learning rate (only used for StochasticWeightAveraging)."
+    swa_epoch_start: NonNegativeFloat = Field(default=0.8)
+    "SWA epoch start (only used for StochasticWeightAveraging)."
+    annealing_epochs: PositiveInt = Field(default=10)
+    "Annealing epochs (only used for StochasticWeightAveraging)."
+    annealing_strategy: Literal["cos", "linear"] = Field(default="cos")
+    "Annealing strategy (only used for StochasticWeightAveraging)."
 
 
 class Rollout(BaseModel):
@@ -334,8 +360,8 @@ class BaseTrainingSchema(BaseModel):
     "Config for gradient clipping."
     strategy: StrategySchemas
     "Strategy to use."
-    swa: SWA = Field(default_factory=SWA)
-    "Config for stochastic weight averaging."
+    weight_averaging: WeightAveragingSchema | None = Field(default=None)
+    "Config for weight averaging (SWA or EMA). Set to null to disable."
     training_loss: LossSchemas
     "Training loss configuration."
     loss_gradient_scaling: bool = False
diff --git a/training/tests/unit/diagnostics/callbacks/test_weight_averaging.py b/training/tests/unit/diagnostics/callbacks/test_weight_averaging.py
@@ -0,0 +1,47 @@
+# (C) Copyright 2024 Anemoi contributors.
+#
+# This software is licensed under the terms of the Apache Licence Version 2.0
+# which can be obtained at http://www.apache.org/licenses/LICENSE-2.0.
+#
+# In applying this licence, ECMWF does not waive the privileges and immunities
+# granted to it by virtue of its status as an intergovernmental organisation
+# nor does it submit to any jurisdiction.
+
+"""Unit tests for weight averaging callback functionality."""
+
+import omegaconf
+import pytest
+import yaml
+
+from anemoi.training.diagnostics.callbacks import _get_weight_averaging_callback
+
+default_config = """
+training:
+  weight_averaging: null
+"""
+
+
+def test_weight_averaging_disabled_when_null() -> None:
+    """Test that weight averaging is disabled when set to null."""
+    config = omegaconf.OmegaConf.create(yaml.safe_load(default_config))
+    callbacks = _get_weight_averaging_callback(config)
+    assert callbacks == []
+
+
+def test_ema_callback_available() -> None:
+    """Test that EMA weight averaging callback can be instantiated."""
+    pytest.importorskip("pytorch_lightning.callbacks", reason="EMA requires PyTorch Lightning 2.6+")
+
+    try:
+        from pytorch_lightning.callbacks import EMAWeightAveraging
+    except ImportError:
+        pytest.skip("EMAWeightAveraging not available in this PyTorch Lightning version")
+
+    config = omegaconf.OmegaConf.create(yaml.safe_load(default_config))
+    config.training.weight_averaging = {
+        "_target_": "pytorch_lightning.callbacks.EMAWeightAveraging",
+        "decay": 0.999,
+    }
+    callbacks = _get_weight_averaging_callback(config)
+    assert len(callbacks) == 1
+    assert isinstance(callbacks[0], EMAWeightAveraging)
