Skip to content

Plan.md

Latest commit

 

History

History
59 lines (43 loc) · 4.42 KB

Plan.md

File metadata and controls

59 lines (43 loc) · 4.42 KB
Raw

GLN EMULaToR Adaptation Plan

1. Baseline Summary From README

  • Baseline: GLN Conditional Graph Logic Network; retrosynthesis prediction.
  • Docs inspected: README.md, gln/data_process/, gln/training/scripts/run_mf.sh, and gln/test/.
  • Native flow: data-processing steps 0/0.1/1/2/3/4/5, run_mf.sh for training, and main_test.py/test_all.sh for evaluation.
  • Default model settings remain those in run_mf.sh (mean-field, latent 128, embedding 256, 64 negatives, three layers, 3000 iterations per validation, top-k/beam 50, GPU 0).
  • The README describes roughly ten 3000-iteration validation cycles, while the script leaves the cycle limit to the Python entrypoint; the wrapper records this discrepancy and uses the documented script defaults.

2. Expected Input Format

  • GLN raw CSV: id,class,reactants>reagents>production.
  • EMULaToR source: /home/adhil/github/EMULaToR/data/processed/datasets/reaction_outcome_dataset.
  • Required source fields: rxn_smiles, atom_mapped_rxn_smiles, reactants, products.
  • Mapping failures are excluded and written to an audit report; confidence is retained as metadata.

3. Featurization and Preprocessing Path

  • The adapter writes mapped GLN CSV views, then invokes GLN's native steps 0, 0.1, 1, 2, 3, 4, and 5.
  • Schema-5 preprocessing is a content-addressed stage DAG. Each completed stage has a manifest with direct-input and implementation fingerprints, output hashes/sizes/counts, dependency keys, and validation results. The schema-5 root is separate from the old schema-4 namespace, which is quarantined rather than adopted by default.
  • Per-item worker results are immutable microshards. A persistent imap_unordered(..., chunksize=1) pool runs each parallel stage and resumes only missing shards; serial work is limited to deterministic merging, validation, and atomic publication.
  • Raw CSV views retain every successfully atom-mapped input row. Template-extraction failures are written to audits and never cause rows to be removed from train/validation/test views.
  • Static graph/preprocessing artifacts are reused; learned model state is never cached as a feature.

4. Training and Evaluation Entrypoints

  • Training delegates to gln/training/main.py through the documented run_mf.sh argument set.
  • Evaluation delegates to gln/test/main_test.py and preserves native cumulative top-k evaluation.
  • Wrappers additionally emit top-1 and top-10 exact-match summaries and select checkpoints by validation top-1.

5. Dataset and Split Mapping

  • Use random_splits_grouped_rxn_smiles (41,108 rows), uniprot_time_splits (41,108 rows), and drfp_tanimoto_splits/threshold_0.11 (39,732 rows).
  • Do not use the stale random_splits directory.
  • Copy validated split views to /home/adhil/github/EMULaToR/data/processed/baselines/GLN.
  • Preserve atom maps supplied by the user; wrappers never run LocalMapper.

6. Files To Add Under emulator_bench/

  • README.md, utils.py, and dataset_adapter.py for validation, mapping, manifests, and CSV generation.
  • cache_features.py and stage_cache.py for schema-5 durable stage manifests, validation-only runs, narrowly-scoped invalidation, atomic publication, and resumable split-specific caches.
  • train.py, evaluate.py, tune_optuna.py, and launch_parallel.py for training, metrics, safe tuning, and multi-GPU orchestration.
  • torch_compat.py for the bench-local PyTorch segmented-log-softmax fallback required by current PyTorch versions.
  • smoke_test.py and commands.txt for reproducible end-to-end verification.

7. Minimal Edits Outside emulator_bench/

  • Plan.md is the only root-level addition. GLN model, trainer, and preprocessing code remain unchanged.

8. Exact Execution Plan

  1. Validate the refreshed master and selected split parquets, excluding only rows whose atom mapping itself failed.
  2. Generate mapped GLN CSV views and manifests under the EMULaToR GLN baseline directory.
  3. Run schema-5 preprocessing once per split with --legacy-cache-policy fresh; quarantine schema-4 data and validate every stage before publishing preprocess.complete.
  4. Launch native GLN training with the original settings, AMP, checkpoint resume, and Rich progress.
  5. Evaluate native top-k plus top-1/top-10 metrics and write consolidated results.
  6. Run Optuna only over safe training hyperparameters and distribute trials across requested GPUs.
  7. Run the small end-to-end smoke test with CUDA_VISIBLE_DEVICES=3 in conda environment test.