Refactor Callbacks (#60)

* Refactor Callbacks
- Split into seperate files
- Use list in config to add callbacks
- Split out plotting callbacks config

* Refactor rollout (#87)
- New rollout central function

---------

Co-authored-by: Mario Santa Cruz <[email protected]>
Co-authored-by: Sara Hahner <[email protected]>

Author: 3 people

CHANGELOG.md

@@ -12,6 +12,10 @@ Keep it human-readable, your future self will thank you!
 
 ## [0.2.2 - Maintenance: pin python <3.13](https://github.com/ecmwf/anemoi-training/compare/0.2.1...0.2.2) - 2024-10-28
 
+### Fixed
+- Refactored callbacks. [#60](https://github.com/ecmwf/anemoi-training/pulls/60)
+- Refactored rollout [#87](https://github.com/ecmwf/anemoi-training/pulls/87)
+    - Enable longer validation rollout than training
 ### Added
 - Included more loss functions and allowed configuration [#70](https://github.com/ecmwf/anemoi-training/pull/70)
 
@@ -29,7 +33,9 @@ Keep it human-readable, your future self will thank you!
 - Feature: New `Boolean1DMask` class. Enables rollout training for limited area models. [#79](https://github.com/ecmwf/anemoi-training/pulls/79)
 
 ### Fixed
-
+- Mlflow-sync to handle creation of new experiments in the remote server [#83] (https://github.com/ecmwf/anemoi-training/pull/83)
+- Fix for multi-gpu when using mlflow due to refactoring of _get_mlflow_run_params function [#99] (https://github.com/ecmwf/anemoi-training/pull/99)
+- ci: fix pyshtools install error (#100) https://github.com/ecmwf/anemoi-training/pull/100
 - Mlflow-sync to handle creation of new experiments in the remote server [#83](https://github.com/ecmwf/anemoi-training/pull/83)
 - Fix for multi-gpu when using mlflow due to refactoring of _get_mlflow_run_params function [#99](https://github.com/ecmwf/anemoi-training/pull/99)
 - ci: fix pyshtools install error [#100](https://github.com/ecmwf/anemoi-training/pull/100)

docs/modules/diagnostics.rst

@@ -21,23 +21,94 @@ functionality to use both Weights & Biases and Tensorboard.
 
 The callbacks can also be used to evaluate forecasts over longer
 rollouts beyond the forecast time that the model is trained on. The
-number of rollout steps (or forecast iteration steps) is set using
-``config.eval.rollout = *num_of_rollout_steps*``.
-
-Note the user has the option to evaluate the callbacks asynchronously
-(using the following config option
-``config.diagnostics.plot.asynchronous``, which means that the model
-training doesn't stop whilst the callbacks are being evaluated).
-However, note that callbacks can still be slow, and therefore the
-plotting callbacks can be switched off by setting
-``config.diagnostics.plot.enabled`` to ``False`` or all the callbacks
-can be completely switched off by setting
-``config.diagnostics.eval.enabled`` to ``False``.
+number of rollout steps for verification (or forecast iteration steps)
+is set using ``config.dataloader.validation_rollout =
+*num_of_rollout_steps*``.
+
+Callbacks are configured in the config file under the
+``config.diagnostics`` key.
+
+For regular callbacks, they can be provided as a list of dictionaries
+underneath the ``config.diagnostics.callbacks`` key. Each dictionary
+must have a ``_target`` key which is used by hydra to instantiate the
+callback, any other kwarg is passed to the callback's constructor.
+
+.. code:: yaml
+
+   callbacks:
+      - _target_: anemoi.training.diagnostics.callbacks.evaluation.RolloutEval
+      rollout: ${dataloader.validation_rollout}
+      frequency: 20
+
+Plotting callbacks are configured in a similar way, but they are
+specified underneath the ``config.diagnostics.plot.callbacks`` key.
+
+This is done to ensure seperation and ease of configuration between
+experiments.
+
+``config.diagnostics.plot`` is a broader config file specifying the
+parameters to plot, as well as the plotting frequency, and
+asynchronosity.
+
+Setting ``config.diagnostics.plot.asynchronous``, means that the model
+training doesn't stop whilst the callbacks are being evaluated)
+
+.. code:: yaml
+
+   plot:
+      asynchronous: True # Whether to plot asynchronously
+      frequency: # Frequency of the plotting
+      batch: 750
+      epoch: 5
+
+      # Parameters to plot
+         parameters:
+         - z_500
+         - t_850
+         - u_850
+
+         # Sample index
+         sample_idx: 0
+
+         # Precipitation and related fields
+         precip_and_related_fields: [tp, cp]
+
+         callbacks:
+         - _target_: anemoi.training.diagnostics.callbacks.plot.PlotLoss
+            # group parameters by categories when visualizing contributions to the loss
+            # one-parameter groups are possible to highlight individual parameters
+            parameter_groups:
+               moisture: [tp, cp, tcw]
+               sfc_wind: [10u, 10v]
+         - _target_: anemoi.training.diagnostics.callbacks.plot.PlotSample
+            sample_idx: ${diagnostics.plot.sample_idx}
+            per_sample : 6
+            parameters: ${diagnostics.plot.parameters}
 
 Below is the documentation for the default callbacks provided, but it is
 also possible for users to add callbacks using the same structure:
 
-.. automodule:: anemoi.training.diagnostics.callbacks
+.. automodule:: anemoi.training.diagnostics.callbacks.checkpoint
+   :members:
+   :no-undoc-members:
+   :show-inheritance:
+
+.. automodule:: anemoi.training.diagnostics.callbacks.evaluation
+   :members:
+   :no-undoc-members:
+   :show-inheritance:
+
+.. automodule:: anemoi.training.diagnostics.callbacks.optimiser
+   :members:
+   :no-undoc-members:
+   :show-inheritance:
+
+.. automodule:: anemoi.training.diagnostics.callbacks.plot
+   :members:
+   :no-undoc-members:
+   :show-inheritance:
+
+.. automodule:: anemoi.training.diagnostics.callbacks.provenance
    :members:
    :no-undoc-members:
    :show-inheritance:

docs/user-guide/configuring.rst

@@ -21,7 +21,7 @@ settings at the top as follows:
    defaults:
    - data: zarr
    - dataloader: native_grid
-   - diagnostics: eval_rollout
+   - diagnostics: evaluation
    - hardware: example
    - graph: multi_scale
    - model: gnn
@@ -100,7 +100,7 @@ match the dataset you provide.
    defaults:
    - data: zarr
    - dataloader: native_grid
-   - diagnostics: eval_rollout
+   - diagnostics: evaluation
    - hardware: example
    - graph: multi_scale
    - model: transformer # Change from default group

docs/user-guide/tracking.rst

@@ -33,7 +33,7 @@ the same experiment.
 Within the MLflow experiments tab, it is possible to define different
 namespaces. To create a new namespace, the user just needs to pass an
 'experiment_name'
-(``config.diagnostics.eval_rollout.log.mlflow.experiment_name``) to the
+(``config.diagnostics.evaluation.log.mlflow.experiment_name``) to the
 mlflow logger.
 
 **Parent-Child Runs**

src/anemoi/training/config/config.yaml

@@ -1,7 +1,7 @@
 defaults:
 - data: zarr
 - dataloader: native_grid
-- diagnostics: eval_rollout
+- diagnostics: evaluation
 - hardware: example
 - graph: multi_scale
 - model: gnn

src/anemoi/training/config/dataloader/native_grid.yaml

@@ -45,6 +45,8 @@ training:
   frequency: ${data.frequency}
   drop:  []
 
+validation_rollout: 1 # number of rollouts to use for validation, must be equal or greater than rollout expected by callbacks
+
 validation:
   dataset: ${dataloader.dataset}
   start: 2021

src/anemoi/training/config/debug.yaml

@@ -1,7 +1,7 @@
 defaults:
 - data: zarr
 - dataloader: native_grid
-- diagnostics: eval_rollout
+- diagnostics: evaluation
 - hardware: example
 - graph: multi_scale
 - model: gnn
@@ -18,7 +18,7 @@ defaults:
 
 diagnostics:
   plot:
-    enabled: False
+    callbacks: []
 hardware:
   files:
     graph: ???

src/anemoi/training/config/diagnostics/callbacks/pretraining.yaml

@@ -0,0 +1 @@
+# Add callbacks here

src/anemoi/training/config/diagnostics/callbacks/rollout_eval.yaml

@@ -0,0 +1,4 @@
+# Add callbacks here
+- _target_: anemoi.training.diagnostics.callbacks.evaluation.RolloutEval
+  rollout: ${dataloader.validation_rollout}
+  frequency: 20

src/anemoi/training/config/diagnostics/eval_rollout.yaml renamed to src/anemoi/training/config/diagnostics/evaluation.yaml

@@ -1,53 +1,8 @@
 ---
-eval:
-  enabled: False
-  # use this to evaluate the model over longer rollouts, every so many validation batches
-  rollout: 12
-  frequency: 20
-plot:
-  enabled: True
-  asynchronous: True
-  frequency: 750
-  sample_idx: 0
-  per_sample: 6
-  parameters:
-  - z_500
-  - t_850
-  - u_850
-  - v_850
-  - 2t
-  - 10u
-  - 10v
-  - sp
-  - tp
-  - cp
-  #Defining the accumulation levels for precipitation related fields and the colormap
-  accumulation_levels_plot: [0, 0.05, 0.1, 0.25, 0.5, 1, 1.5, 2, 3, 4, 5, 6, 7, 100] # in mm
-  cmap_accumulation: ["#ffffff", "#04e9e7", "#019ff4", "#0300f4", "#02fd02", "#01c501", "#008e00", "#fdf802", "#e5bc00", "#fd9500", "#fd0000", "#d40000", "#bc0000", "#f800fd"]
-  precip_and_related_fields: [tp, cp]
-  # Histogram and Spectrum plots
-  parameters_histogram:
-  - z_500
-  - tp
-  - 2t
-  - 10u
-  - 10v
-  parameters_spectrum:
-  - z_500
-  - tp
-  - 2t
-  - 10u
-  - 10v
-  # group parameters by categories when visualizing contributions to the loss
-  # one-parameter groups are possible to highlight individual parameters
-  parameter_groups:
-    moisture: [tp, cp, tcw]
-    sfc_wind: [10u, 10v]
-  learned_features: False
-  longrollout:
-    enabled: False
-    rollout: [60]
-    frequency: 20 # every X epochs
+defaults:
+  - plot: detailed
+  - callbacks: pretraining
+
 
 debug:
   # this will detect and trace back NaNs / Infs etc. but will slow down training
@@ -57,6 +12,7 @@ debug:
 # remember to also activate the tensorboard logger (below)
 profiler: False
 
+enable_checkpointing: True
 checkpoint:
   every_n_minutes:
     save_frequency: 30 # Approximate, as this is checked at the end of training steps