Spaces:
Sleeping
Sleeping
Download diagnostics/2026-10-30-lab-hyperparameter-grid/scripts/seed_check.py from bozdaglab/BioLM-NET-Workbench: direct link, hf CLI and curl.
- Browser
- Download file 3.39 kB
-
https://huggingface.co/spaces/bozdaglab/BioLM-NET-Workbench/resolve/main/diagnostics/2026-10-30-lab-hyperparameter-grid/scripts/seed_check.py
- Command line
-
hf download hf://spaces/bozdaglab/BioLM-NET-Workbench/diagnostics/2026-10-30-lab-hyperparameter-grid/scripts/seed_check.py
-
curl -L -o seed_check.py https://huggingface.co/spaces/bozdaglab/BioLM-NET-Workbench/resolve/main/diagnostics/2026-10-30-lab-hyperparameter-grid/scripts/seed_check.py
3.39 kB
| """Re-run the best grid configurations (plus the 2026-10-16 stable baseline) across 8 seeds. | |
| Usage: seed_check.py <workspace.pkl> <out.jsonl> <shard> <n_shards> | |
| Resumable: (config, seed) pairs already in <out.jsonl> are skipped.""" | |
| import json | |
| import pickle | |
| import sys | |
| from dataclasses import asdict, replace | |
| from pathlib import Path | |
| import numpy as np | |
| import torch | |
| from sklearn.metrics import recall_score | |
| import biolmnet.training as training | |
| WORKSPACE, OUT, SHARD, SHARDS = Path(sys.argv[1]), Path(sys.argv[2]), int(sys.argv[3]), int(sys.argv[4]) | |
| torch.set_num_threads(2) | |
| SEEDS = [0, 1, 2, 3, 7, 42, 123, 2024] | |
| GRID_BASE = dict(epochs=200, learning_rate=0.01, weight_decay=0.01, dropout=0.3, optimizer="Adam") | |
| CONFIGS = { | |
| "A b32 p64 tanh/tanh/relu": dict(GRID_BASE, batch_size=32, projection_dim=64, | |
| biological_activation="tanh", projection_activation="tanh", fusion_activation="relu"), | |
| "B b16 p128 relu/relu/tanh": dict(GRID_BASE, batch_size=16, projection_dim=128, | |
| biological_activation="relu", projection_activation="relu", fusion_activation="tanh"), | |
| "C b32 p128 tanh/relu/relu": dict(GRID_BASE, batch_size=32, projection_dim=128, | |
| biological_activation="tanh", projection_activation="relu", fusion_activation="relu"), | |
| "D b32 p64 tanh/relu/tanh": dict(GRID_BASE, batch_size=32, projection_dim=64, | |
| biological_activation="tanh", projection_activation="relu", fusion_activation="tanh"), | |
| "E b32 p128 tanh/relu/tanh": dict(GRID_BASE, batch_size=32, projection_dim=128, | |
| biological_activation="tanh", projection_activation="relu", fusion_activation="tanh"), | |
| # 2026-10-16 stable setting: Adam lr 1e-3, weight_decay 1e-4, published activations, 100-epoch limit. | |
| "BASE lr1e-3 wd1e-4 relu/sigmoid/tanh": dict(epochs=100, weight_decay=1e-4), | |
| } | |
| workspace = pickle.loads(WORKSPACE.read_bytes()) | |
| names = workspace.label_names | |
| done = set() | |
| if OUT.exists(): | |
| done = {(r["config"], r["seed"]) for r in map(json.loads, OUT.read_text().splitlines()) if r} | |
| jobs = [(c, s) for c in CONFIGS for s in SEEDS] | |
| for index, (config, seed) in enumerate(jobs): | |
| if index % SHARDS != SHARD or (config, seed) in done: | |
| continue | |
| hp = replace(training.Hyperparameters(), seed=seed, **CONFIGS[config]) | |
| result = training.train(workspace, hp) | |
| m = result.bundle.metrics | |
| vp = result.validation_predictions | |
| counts = vp["predicted"].value_counts().reindex(names, fill_value=0).astype(int).tolist() | |
| row = { | |
| "config": config, "seed": seed, "hyperparameters": asdict(hp), | |
| "accuracy": m["accuracy"], "f1_macro": m["f1_macro"], "f1_weighted": m["f1_weighted"], | |
| "epochs_completed": m["epochs_completed"], | |
| "best_epoch": int(min(result.bundle.history, key=lambda h: h["validation_loss"])["epoch"]), | |
| "pred_counts": counts, "collapsed": int((np.array(counts) > 0).sum()) == 1, | |
| "recall": recall_score(vp["observed"], vp["predicted"], labels=names, average=None, zero_division=0).tolist(), | |
| } | |
| with OUT.open("a") as handle: | |
| handle.write(json.dumps(row) + "\n") | |
| print(f"[{SHARD}] {config} seed={seed} f1m={row['f1_macro']:.3f} collapsed={row['collapsed']}", flush=True) | |
| print(f"SHARD {SHARD} DONE", flush=True) | |