[DATASET ROADMAP] convert public bases

[bstaber]

Let's try to make a curated list of datasets that we could convert to PLAID. Here's a first draft:

• Vortex shedding (MGN paper)
• Airfoil dataset (MGN paper)
• Deforming plate (MGN paper)
• Flag dataset (MGN paper)
• DrivAerNet [Elrefaie et al., 2024]
• Ahmed body
• Datasets from the GEO-FNO paper available here: elasticity, plasticity, airfoil naca, pipeflow.
• Datasets used in OTNO paper
• The FlowBench 3D Lid-Driven Cavity dataset hosted on HuggingFace
• Datasets used in this benchmark paper
• PDEArena
• BubbleML
• BLASTNet
• PDEBench
• The Well
• PINNacle
• May be some interesting cases here
• CFDBench
• Dataset from Fourier Neural Operator with Learned Deformations for PDEs on General Geometries

Add @fabiencasenave list as checklist

I think we should converge on the following points before mass conversion:

MANDATORY:

• globals in CGNS trees to replace scalars and time series -> PR ♻️ introduce global in CGNS trees to replace scalars and time_series mechanism #229
• Assess the links/paths mechanism -> PR ♻️feat(links/paths) remove links/paths and CGNS linking capacities #239
• Use of Hugging Face's native split -> PR ♻️ Modernize the Hugging Face bridge #240 , still time variable samples to implement.
• Problem_definition: indicate in and out split concerned by the regression task and name the score function for the moment (maybe later we should find a way to define an implementation), see next comment.

OPTIONAL

• Clarify the status of the time in feature_identifiers. time=all means all time steps are considered by the feature ? Or no mention for trying to access all existing time steps for this feature ? What about sample.get_all_features_identifiers() ?
• Enforce the use of feature_identifiers in problem_definitions (in the yaml as well) and replace node feature by 'mesh' (the complete support, with nodes, elements and tags, at a particular base and zone). Proposal: flattened keys (already implemented in the Huggingface bridge) for feature ids ?
• [ARCHITECTURE UPDATE] update data organisation #241 should also be addressed for the plaid part - with the mandatory part above, it will be implemented totally for the HF format

Important note

We just tneed to converge on a representation on HuggingFace to start converting other bases.

[casenave]

I'm setting the milestone to 1.0.0, not as an objective to include all the list, but a first batch

I think we should converge on the following points before mass conversion:

MANDATORY:

• globals in CGNS trees to replace scalars and time series -> PR ♻️ introduce global in CGNS trees to replace scalars and time_series mechanism #229
• Assess the links/paths mechanism -> PR ♻️feat(links/paths) remove links/paths and CGNS linking capacities #239
• Use of Hugging Face's native split -> PR ♻️ Modernize the Hugging Face bridge #240 , still time variable samples to implement.
• Problem_definition: indicate in and out split concerned by the regression task and name the score function for the moment (maybe later we should find a way to define an implementation), see next comment.

OPTIONAL

• Clarify the status of the time in feature_identifiers. time=all means all time steps are considered by the feature ? Or no mention for trying to access all existing time steps for this feature ? What about sample.get_all_features_identifiers() ?
• Enforce the use of feature_identifiers in problem_definitions (in the yaml as well) and replace node feature by 'mesh' (the complete support, with nodes, elements and tags, at a particular base and zone). Proposal: flattened keys (already implemented in the Huggingface bridge) for feature ids ?
• [ARCHITECTURE UPDATE] update data organisation #241 should also be addressed for the plaid part - with the mandatory part above, it will be implemented totally for the HF format

Important note

We just tneed to converge on a representation on HuggingFace to start converting other bases.

[casenave]

Here what a problem_definition can look on file, for Tensile2d:

input_features:
- Base_2_2/Zone
- Base_2_2/Zone/Elements_TRI_3/ElementConnectivity
- Base_2_2/Zone/Elements_TRI_3/ElementRange
- Base_2_2/Zone/GridCoordinates/CoordinateX
- Base_2_2/Zone/GridCoordinates/CoordinateY
- Base_2_2/Zone/ZoneBC/Bottom/PointList
- Base_2_2/Zone/ZoneBC/BottomLeft/PointList
- Base_2_2/Zone/ZoneBC/Top/PointList
- Global/P
- Global/p1
- Global/p2
- Global/p3
- Global/p4
- Global/p5
output_features:
- Base_2_2/Zone/PointData/U1
- Base_2_2/Zone/PointData/U2
- Base_2_2/Zone/PointData/q
- Base_2_2/Zone/PointData/sig11
- Base_2_2/Zone/PointData/sig12
- Base_2_2/Zone/PointData/sig22
- Global/max_U2_top
- Global/max_q
- Global/max_sig22_top
- Global/max_von_mises
task: regression
score_function: RRMSE
splits:
  - train_splits:
      - train
  - test_splits:
      - test
      - OOD

where all the features are taken form the variable_features entry of key_mappings. Notice that this file is automatically produced by the Hugging Face bridge, which scans the complete datasets for infering the constant and variable features from the keys of the flatten CGNS trees.

Subsplits can be specified:

splits:
  - train_splits:
      - train: subset_150 
  - test_splits:
      - test

and these subsets can be provided in json file.

[xroynard]

Nice proposition !

Just input_splits and output_splits seems non-sense, there is a proposition to handle splits here #250

[casenave]

Nice proposition !

Just input_splits and output_splits seems non-sense, there is a proposition to handle splits here #250

I agree, I renamed input_splits and output_splits to train_splits and test_splits

[casenave]

#274 introduced many improvement towards this goal. Maybe #276 would be nice before mass conversion.