feat(models,training): multi dataset integration #594
+2,748
−807
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Various changes, see git history. --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
…ure/multi_ds_dict
mchantry
changed the title
This was referenced Oct 9, 2025
JPXKQX
reviewed
Oct 21, 2025
JPXKQX
reviewed
Oct 21, 2025
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
Contributor
|
Hi @mchantry, I'm planning on doing some training with this branch. Could you let me know what I should expect to work and not work? |
MeraX
reviewed
Nov 11, 2025
Comment on lines
+176
to
+178
| # ALWAYS override dataset from dataloader config (ignore dummy in graph config) | ||
| if hasattr(graph_config.nodes.data.node_builder, "dataset"): | ||
| graph_config.nodes.data.node_builder.dataset = dataset_path |
Contributor
There was a problem hiding this comment.
Would a similar mechanism be required for ICON Node builders? see also #627
radiradev
reviewed
Nov 11, 2025
| dataset: aifs-ea-an-oper-0001-mars-${data.resolution}-1979-2023-6h-v8.zarr | ||
|
|
||
| # Secondary dataset for ERA5_copy (using same file for debugging) | ||
| dataset_b: cerra-rr-an-oper-0001-mars-5p5km-1984-2022-6h-v3-hmsi.zarr |
Contributor
There was a problem hiding this comment.
Is this working currently? I've been trying to run with two datasets with non-identical datetimes and getting errors
Contributor
There was a problem hiding this comment.
Here I mean, one is a subset of datetimes of the other as it's the case in the example above.
Member
Author
There was a problem hiding this comment.
What do you mean about non-identical datetimes radi?
The first implementation of multiple datasets aim to use datasets that are fully time aligned, and feature the time datetimes
radiradev
reviewed
Nov 12, 2025
radiradev
reviewed
Nov 12, 2025
for more information, see https://pre-commit.ci
## Description Adaptation of plotting callbacks to multiple datasets. We add a parameter `datasets` to the callback configuration that allows specifying which datasets to plot. All plots will be based on the same configuration, including the choice of parameters to plot. **In order to plot different datasets with different parameters, the user can configure multiple callbacks of the same type with different parameters.** Callbacks included: - PlotLoss - PlotSample - PlotSpectrum - PlotHistogram - GraphTrainableFeaturesPlot Callbacks not included in this PR: - ensemble plots - LongRolloutPlots Here are MLFlow runs for a [single dataset](https://mlflow.ecmwf.int/#/experiments/420/runs/8dca1935a3194edea16a21d4ed4d6a13/artifacts) and a [two dataset](https://mlflow.ecmwf.int/#/experiments/420/runs/ea59f113b0d34061b487038552d9467c/artifacts) use case. ## The config interface for plotting Most datasets will have somewhat different parameters, and hence different parameters to plot. The config interface implemented on this PR means that users will either need to plot parameters shared by all datasets, or configure multiple callbacks of the same type for different datasets. The reasons for this interface as a first draft is not that is it optimal for multiple datasets but rather that: - it keeps plots and plotting configs for the single dataset case backwards compatible. Avoiding regression on plotting for existing use cases was the main requirement. - it requires less changes to the callbacks than making them more configurable - it provides _some functionality_ to plot multiple datasets. - we don't know yet how multiple datasets will mostly be used and it might make sense to delay the design of the interface a bit. This will give us more time to rethink plotting callbacks more generally, rather than settling on an interface now. ## Other decisions - We will adapt pydantic schemas for plotting (and set a default for the new parameter `datasets = ["data"]`) as part of fixing schemas on the multiple datasets branch - I'd suggest to update the remaining plotting callbacks separately but open to thoughts.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
Supports multiple time-aligned datasets as inputs and outputs for training.
e.g.
where inputs/outputs each use their own encoder/decoder.
As a contributor to the Anemoi framework, please ensure that your changes include unit tests, updates to any affected dependencies and documentation, and have been tested in a parallel setting (i.e., with multiple GPUs). As a reviewer, you are also responsible for verifying these aspects and requesting changes if they are not adequately addressed. For guidelines about those please refer to https://anemoi.readthedocs.io/en/latest/
By opening this pull request, I affirm that all authors agree to the Contributor License Agreement.