Spaces:
Running on Zero
Running on Zero
| """Jubair's hyperparameter grid (email of 2026-10-17) on BRCA, using the Workbench's train() unchanged. | |
| Usage: grid.py <workspace.pkl> <out.jsonl> <shard> <n_shards> [seed] | |
| Resumable: configs already in <out.jsonl> are skipped. | |
| Grid as sent, mapped onto Workbench Hyperparameters: | |
| layer_one -> projection_dim; activation_function_one/two/three -> biological/projection/fusion activation; | |
| l2_reg -> weight_decay; epoch -> epochs (early stopping, patience 12, still applies); | |
| SGD gets momentum 0.03 as instructed. num_ex_neuron is not supported by the Workbench (fixed at 0). | |
| """ | |
| import itertools | |
| import json | |
| import sys | |
| from dataclasses import asdict, replace | |
| from pathlib import Path | |
| import numpy as np | |
| import pickle | |
| 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]) | |
| SEED = int(sys.argv[5]) if len(sys.argv) > 5 else 42 | |
| torch.set_num_threads(2) | |
| ACT = ["relu", "sigmoid", "tanh"] | |
| GRID = list(itertools.product([64, 128], ACT, ACT, ACT, [16, 32], ["SGD", "Adam"])) | |
| workspace = pickle.loads(WORKSPACE.read_bytes()) | |
| names = workspace.label_names | |
| done = set() | |
| if OUT.exists(): | |
| done = {json.dumps(json.loads(line)["key"]) for line in OUT.read_text().splitlines() if line.strip()} | |
| for index, (proj, a1, a2, a3, batch, opt) in enumerate(GRID): | |
| if index % SHARDS != SHARD: | |
| continue | |
| key = {"projection_dim": proj, "act1": a1, "act2": a2, "act3": a3, "batch_size": batch, "optimizer": opt, "seed": SEED} | |
| if json.dumps(key) in done: | |
| continue | |
| hp = replace( | |
| training.Hyperparameters(), | |
| epochs=200, batch_size=batch, learning_rate=0.01, weight_decay=0.01, dropout=0.3, | |
| projection_dim=proj, optimizer=opt, sgd_momentum=0.03 if opt == "SGD" else 0.0, | |
| biological_activation=a1, projection_activation=a2, fusion_activation=a3, seed=SEED, | |
| ) | |
| try: | |
| result = training.train(workspace, hp) | |
| except Exception as exc: # record and keep going | |
| row = {"key": key, "error": repr(exc)} | |
| else: | |
| m = result.bundle.metrics | |
| vp = result.validation_predictions | |
| counts = vp["predicted"].value_counts().reindex(names, fill_value=0).astype(int).tolist() | |
| history = result.bundle.history | |
| best_epoch = int(min(history, key=lambda h: h["validation_loss"])["epoch"]) | |
| row = { | |
| "key": key, "hyperparameters": asdict(hp), | |
| "accuracy": m["accuracy"], "f1_macro": m["f1_macro"], "f1_weighted": m["f1_weighted"], | |
| "epochs_completed": m["epochs_completed"], "best_epoch": best_epoch, | |
| "pred_counts": counts, "classes_predicted": int((np.array(counts) > 0).sum()), | |
| "recall": recall_score(vp["observed"], vp["predicted"], labels=names, average=None, zero_division=0).tolist(), | |
| "final_train_loss": history[-1]["training_loss"], "best_val_loss": m["validation_loss"], | |
| } | |
| with OUT.open("a") as handle: | |
| handle.write(json.dumps(row) + "\n") | |
| print(f"[{SHARD}] {index} {key} -> " + (row.get("error") or f"f1m={row['f1_macro']:.3f} classes={row['classes_predicted']}"), flush=True) | |
| print(f"SHARD {SHARD} DONE", flush=True) | |