Enhance crash recipe with configurable stats directory

[dran-dev]

PhysicsNeMo Pull Request

Description

Make the stats directory (node_stats.json, feature_stats.json,
edge_stats.json) configurable via Hydra in the crash simulation recipe.

Previously, training wrote normalization statistics to a relative ./stats/
path under the Hydra run output directory. During inference, a different Hydra
output directory is used, so the stats files were not found — resulting in:

FileNotFoundError: Node stats file stats/node_stats.json or
feature stats file stats/feature_stats.json not found

This PR adds a datapipe.stats_dir configuration option (defaulting to
null, which preserves the existing ./stats/ behavior) so users can
point inference to the training run's stats directory:

python inference.py --config-name=experiment_bumper_geotransolver \
  ++datapipe.stats_dir=/path/to/training_output/stats

Changes

• Added stats_dir parameter to CrashBaseDataset.__init__ with
  backward-compatible default (None falls back to relative stats/).
• Added stats_dir: null to point_cloud.yaml and graph.yaml
  datapipe configs.
• Added logging of the resolved stats directory path in datapipe.py,
  train.py, and inference.py.
• Updated README to document the new datapipe.stats_dir option.
• Updated CHANGELOG.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• The CHANGELOG.md is up to date with these changes.
• An issue is linked to this pull request.
• If I am implementing a new model or modifying any existing model, I have followed the Models Implementation Coding Standards.

Dependencies

None.

[greptile-apps]

Greptile Summary

This PR successfully adds a configurable stats_dir parameter to the crash simulation recipe, solving the problem where inference couldn't find normalization statistics computed during training.

Key improvements:

• Adds backward-compatible stats_dir parameter to CrashBaseDataset (defaults to ./stats/ under Hydra run directory)
• Enables users to point inference to training stats via ++datapipe.stats_dir=/path/to/training_output/stats
• Properly logs the resolved stats directory in both training and inference
• Updates configuration files and documentation

Implementation approach:
The change leverages Hydra's instantiate() to automatically pass the stats_dir config value to the dataset constructor. The dataset implementation uses stats_dir if stats_dir is not None else STATS_DIRNAME, preserving existing behavior when the parameter is not specified.

The solution is clean and minimal, with good backward compatibility.

Important Files Changed

Filename	Overview
examples/structural_mechanics/crash/datapipe.py	Added stats_dir parameter to CrashBaseDataset.__init__() with backward-compatible default behavior and proper logging
examples/structural_mechanics/crash/train.py	Added logging for stats directory path at startup
examples/structural_mechanics/crash/inference.py	Added logging for stats directory path at startup
examples/structural_mechanics/crash/conf/datapipe/graph.yaml	Added stats_dir: null configuration option with descriptive comment
examples/structural_mechanics/crash/conf/datapipe/point_cloud.yaml	Added stats_dir: null configuration option with descriptive comment

_{Last reviewed commit: 7f6061f}

[greptile-apps]

_{7 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[dran-dev]

/blossom-ci

[dran-dev]

@mnabian @ktangsali PTAL

[ktangsali]

/blossom-ci

[mnabian]

Hi @dran-dev, thanks for the PR. would it make sense to get rid of the hardcoded STATS_DIRNAME and store the default to in the datapipe yaml files?

[dran-dev]

Hi @dran-dev, thanks for the PR. would it make sense to get rid of the hardcoded STATS_DIRNAME and store the default to in the datapipe yaml files?

Makes sense. Fixed.

[mnabian]

/blossom-ci

[mnabian]

LGTM. Left a small comment.

[mnabian]

/blossom-ci