Consolidate Training/Evaluation Run Outputs into a Single `runs/` Directory

[sadamov]

Problem

A single training run scatters output across five unrelated directories:

Artifact	Current location
Model checkpoints	saved_models/<run-name>/
Lightning CSV logs	lightning_logs/version_N/ (auto-incremented, not named)
W&B offline cache	wandb/offline-run-<timestamp>-<id>/
Eval artifacts (PDFs, CSVs, .pt files)	wandb/ or mlruns/ via self.logger.save_dir
MLFlow artifacts	mlruns/ (hard-coded in CustomMLFlowLogger.save_dir)

Run Name Format

Auto-generated in neural_lam/train_model.py as:
{prefix}{model}-{processor_layers}x{hidden_dim}-{MM_DD_HH}-{random_4digits}
e.g. train-graph_lam-2x64-02_21_12-4571. Override with --logger_run_name.
The run name is applied to saved_models/ and the external logger, but not to lightning_logs/.

Proposed Change

All output should go under runs/<run-name>/:

1. train_model.py: set ModelCheckpoint(dirpath=f"runs/{run_name}/checkpoints") and default_root_dir=f"runs/{run_name}" on pl.Trainer.
2. utils.py: pass save_dir=f"runs/{run_name}" to WandbLogger and CustomMLFlowLogger.
3. custom_loggers.py: make save_dir return the run-scoped path instead of "mlruns".
4. ar_model.py: no changes needed — artifacts already use self.logger.save_dir.
5. README.md / .gitignore: replace saved_models/, lightning_logs/, wandb/ references with runs/.

[PrakshaaleJain]

Hey! I've worked on this issue and would love to open a PR.

[PrakshaaleJain]

Hey! I’ve just opened a PR, please take a look when you can. If you spot any issues, I’d love to hear your feedback!

[techaadii]

Hey I have solved this issue can I open a PR

[PrakshaaleJain]

Hi!! @techaadii, I see a PR (#297) has already been opened for this issue by me. You’re welcome to review it or collaborate if you’d like to contribute further. Just check with the maintainer if another PR would be helpful!

[Shyam-Sunder-saini]

I’ve worked on #293 and consolidated the training and evaluation outputs into a single runs/ directory. Earlier, artifacts were being written to multiple locations (lightning_logs/, wandb/, mlruns/, etc.), which made it difficult to manage experiment outputs.

Changes implemented:

• Updated the trainer configuration to use runs/<run-name>/ as the root directory.
• Configured ModelCheckpoint to save checkpoints under runs/<run-name>/checkpoints.
• Updated logger initialization so WandbLogger and CustomMLFlowLogger write outputs to the same run directory.
• Adjusted the custom logger to return the run-scoped save_dir.
• Updated .gitignore and documentation references to use runs/ instead of scattered output directories.

With this change, all artifacts (checkpoints, logs, experiment outputs) are now grouped under a single run directory, making experiment tracking and cleanup easier.

Please let me know if this approach aligns with what you had in mind for #293 or if you’d like any adjustments before I open the PR.