File size: 15,848 Bytes
c87881a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 | """Command-line entry points for the isolated rebuild."""
from __future__ import annotations
import argparse
import json
from pathlib import Path
from typing import Any
import pandas as pd
import torch
from .artifacts import build_file_manifest, create_run_directory, sha256_file, write_json_immutable
from .audit import audit_legacy_data
from .baselines import (
aggregate_raw_esm,
cosine_scores,
ensemble_scores,
pfam_jaccard_scores,
select_ensemble_alpha,
weighted_pfam_jaccard_scores,
)
from .checkpoints import load_checkpoint
from .config import LoadedConfig, load_config
from .data import BGCEmbeddingDataset, build_pfam_vocab, load_legacy_labels
from .evaluation import evaluate_retrieval
from .model import ModelConfig, build_model
from .reporting import write_paper_outputs
from .splits import build_group_split, load_split, split_summary
from .training import choose_device, encode_dataset, train_phase1, train_phase2, train_weighted_pfam
def _paths(config: LoadedConfig) -> dict[str, Path]:
return {
"atlas": config.resolve_path("data", "atlas_csv"),
"genes": config.resolve_path("data", "genes_csv"),
"h5": config.resolve_path("data", "embeddings_h5"),
"legacy_root": config.resolve_path("data", "legacy_root"),
"manifest_root": config.resolve_path("project", "manifest_root"),
"run_root": config.resolve_path("project", "run_root"),
}
def command_audit(args: argparse.Namespace) -> None:
config = load_config(args.config)
paths = _paths(config)
legacy_root = paths["legacy_root"]
audit = audit_legacy_data(paths["atlas"], paths["genes"], paths["h5"])
candidates = [
legacy_root / "community_atlas.csv",
legacy_root / "discovery_zone.csv",
legacy_root / "master_gene_annotations.csv",
legacy_root / "training_positive_pairs.csv",
legacy_root / "training_representatives.csv",
legacy_root / "esm2_embeddings.h5",
legacy_root / "paper/main.tex",
]
candidates.extend(sorted((legacy_root / "models").glob("*.pt")))
candidates.extend(sorted((legacy_root / "results/publication").glob("*")))
manifest = build_file_manifest(candidates, legacy_root)
write_json_immutable(paths["manifest_root"] / "legacy_audit.json", audit)
write_json_immutable(paths["manifest_root"] / "legacy_manifest.json", manifest)
print(json.dumps(audit, indent=2, sort_keys=True))
def command_build_splits(args: argparse.Namespace) -> None:
config = load_config(args.config)
paths = _paths(config)
values = config.values
labels = load_legacy_labels(paths["atlas"])
data_config = values["data"]
assignments = build_group_split(
labels,
seed=int(values["project"]["seed"]),
minimum_group_size=int(data_config["minimum_group_size"]),
fractions=(
float(data_config["train_fraction"]),
float(data_config["validation_fraction"]),
float(data_config["test_fraction"]),
),
)
output = paths["manifest_root"] / "silver_split.csv"
summary_path = paths["manifest_root"] / "silver_split_summary.json"
output.parent.mkdir(parents=True, exist_ok=True)
if output.exists() or summary_path.exists():
raise FileExistsError("Refusing to overwrite the frozen split; change its versioned name")
assignments.to_csv(output, index=False)
summary = split_summary(assignments)
summary["file_sha256"] = sha256_file(output)
summary["claim_scope"] = "development-only silver MIBiG-reference split"
write_json_immutable(summary_path, summary)
print(json.dumps(summary, indent=2, sort_keys=True))
def _make_dataset(
config: LoadedConfig,
assignments: pd.DataFrame,
pfam_vocab: dict[str, int],
) -> BGCEmbeddingDataset:
paths = _paths(config)
return BGCEmbeddingDataset(
paths["h5"],
paths["atlas"],
assignments,
esm_dimension=int(config.values["model"]["esm_dimension"]),
pfam_vocab=pfam_vocab,
)
def _build_pfam_vocab(config: LoadedConfig, assignments: pd.DataFrame) -> dict[str, int]:
paths = _paths(config)
train_ids = assignments.loc[assignments["split"].eq("train"), "bgc_id"]
return build_pfam_vocab(paths["atlas"], train_ids)
def command_train(args: argparse.Namespace) -> None:
config = load_config(args.config)
paths = _paths(config)
split_path = Path(args.split).resolve() if args.split else paths["manifest_root"] / "silver_split.csv"
assignments = load_split(split_path)
train_assignments = assignments[assignments["split"] == "train"].copy()
validation_assignments = assignments[assignments["split"] == "validation"].copy()
pfam_vocab = _build_pfam_vocab(config, assignments)
train_dataset = _make_dataset(config, train_assignments, pfam_vocab)
validation_dataset = _make_dataset(config, validation_assignments, pfam_vocab)
run_dir = create_run_directory(paths["run_root"], args.run_id)
model_values = dict(config.values["model"])
model_values["pfam_vocab_size"] = len(pfam_vocab) + 2
config_values = dict(config.values)
config_values["model"] = model_values
write_json_immutable(run_dir / "config.json", config_values)
write_json_immutable(run_dir / "pfam_vocab.json", pfam_vocab)
model_config = ModelConfig.from_dict(model_values)
model = build_model(model_config)
input_paths = [paths["atlas"], paths["h5"]]
if model_config.architecture == "weighted_pfam_jaccard":
if args.stage != "phase2":
raise ValueError("Weighted Pfam Jaccard uses phase2 training directly")
checkpoint = train_weighted_pfam(
model, model_config, train_dataset, validation_dataset, validation_assignments,
split_path, input_paths, run_dir, config.values["training"],
int(config.values["project"]["seed"]),
)
print(checkpoint)
return
if args.stage == "phase1":
checkpoint = train_phase1(
model, model_config, train_dataset, train_dataset, split_path, input_paths,
run_dir, config.values["training"], int(config.values["project"]["seed"]),
)
else:
checkpoint = train_phase2(
model, model_config, train_dataset, validation_dataset, validation_assignments,
split_path, input_paths, run_dir, config.values["training"],
int(config.values["project"]["seed"]), args.phase1_checkpoint,
)
print(checkpoint)
def _pfam_sets(atlas_path: Path) -> dict[str, set[str]]:
atlas = pd.read_csv(atlas_path, usecols=["bgc_id", "pfam_ids"])
return {
str(row.bgc_id): set(str(row.pfam_ids).split(";")) if pd.notna(row.pfam_ids) else set()
for row in atlas.itertuples(index=False)
}
def _raw_esm_embeddings(
dataset: BGCEmbeddingDataset,
) -> tuple[dict[str, torch.Tensor], dict[str, torch.Tensor]]:
mean: dict[str, torch.Tensor] = {}
maximum: dict[str, torch.Tensor] = {}
for index in range(len(dataset)):
item = dataset[index]
mean[item["bgc_id"]] = aggregate_raw_esm(item["gene_embeddings"], "mean")
maximum[item["bgc_id"]] = aggregate_raw_esm(item["gene_embeddings"], "max")
return mean, maximum
def command_evaluate(args: argparse.Namespace) -> None:
config = load_config(args.config)
paths = _paths(config)
split_path = Path(args.split).resolve() if args.split else paths["manifest_root"] / "silver_split.csv"
assignments = load_split(split_path)
pfam_vocab = _build_pfam_vocab(config, assignments)
model_values = dict(config.values["model"])
model_values["pfam_vocab_size"] = len(pfam_vocab) + 2
model_config = ModelConfig.from_dict(model_values)
model = build_model(model_config)
load_checkpoint(args.checkpoint, model, split_path)
device = choose_device()
model.to(device)
dataset = _make_dataset(config, assignments, pfam_vocab)
pfams = _pfam_sets(paths["atlas"])
if model_config.architecture == "weighted_pfam_jaccard":
learned_weights = model.domain_weights().detach().cpu().tolist()
inverse_vocab = {index: token for token, index in pfam_vocab.items()}
weights_by_pfam = {
token: float(learned_weights[index])
for index, token in inverse_vocab.items()
if index < len(learned_weights)
}
unknown_weight = float(learned_weights[1])
def weighted_jaccard(candidates: list[str], references: list[str]) -> dict[str, float]:
return weighted_pfam_jaccard_scores(
candidates, references, pfams, weights_by_pfam, unknown_weight, "max"
)
def plain_jaccard(candidates: list[str], references: list[str]) -> dict[str, float]:
return pfam_jaccard_scores(candidates, references, pfams, "max")
methods = {
"pfam_jaccard_max": plain_jaccard,
"weighted_pfam_jaccard": weighted_jaccard,
}
evaluation_config = config.values["evaluation"]
evaluation_seed = int(evaluation_config.get("seed", config.values["project"]["seed"]))
test_results = evaluate_retrieval(
assignments, "test", methods, int(evaluation_config["reference_size"]),
int(evaluation_config["query_draws"]), evaluation_seed,
evaluation_config["recall_at"], evaluation_config["ndcg_at"],
)
run_dir = create_run_directory(paths["run_root"], args.run_id)
metrics = [evaluation_config["primary_metric"], "mrr", "map", "ndcg@50", "tie_fraction"]
metadata = {
"schema_version": 1,
"selected_alpha": None,
"alpha_selected_on": None,
"split_sha256": sha256_file(split_path),
"checkpoint_sha256": sha256_file(args.checkpoint),
"label_tiers": sorted(assignments["label_tier"].unique()),
"pfam_feature": "BGC-level pfam_ids inventory; vocabulary built from training BGCs only",
"weighted_pfam_feature": "trainable nonnegative domain weights optimized with differentiable weighted Jaccard",
"pfam_vocab_size": len(pfam_vocab) + 2,
"publication_eligible": bool((assignments["label_tier"] == "gold").any()),
"claim_warning": "Silver-only results are development evidence, not the primary biological result.",
}
write_paper_outputs(
run_dir, test_results, metrics, int(evaluation_config["bootstrap_samples"]),
float(evaluation_config["confidence_level"]), evaluation_seed, metadata,
)
pd.DataFrame().to_csv(run_dir / "validation_alpha_search.csv", index=False)
print(json.dumps(metadata, indent=2, sort_keys=True))
return
learned = encode_dataset(model, dataset, device, int(config.values["training"]["num_workers"]))
raw_mean, raw_max = _raw_esm_embeddings(dataset)
pfams = _pfam_sets(paths["atlas"])
evaluation_config = config.values["evaluation"]
evaluation_seed = int(
evaluation_config.get("seed", config.values["project"]["seed"])
)
def jaccard(candidates: list[str], references: list[str]) -> dict[str, float]:
return pfam_jaccard_scores(candidates, references, pfams, "max")
def learned_cosine(candidates: list[str], references: list[str]) -> dict[str, float]:
return cosine_scores(candidates, references, learned, "mean")
validation_methods: dict[str, Any] = {"pfam_jaccard": jaccard, "setnet": learned_cosine}
for alpha in evaluation_config["ensemble_alphas"]:
validation_methods[f"ensemble_a{float(alpha):.1f}"] = (
lambda candidates, references, weight=float(alpha): ensemble_scores(
learned_cosine(candidates, references), jaccard(candidates, references), weight
)
)
validation = evaluate_retrieval(
assignments, "validation", validation_methods, int(evaluation_config["reference_size"]),
int(evaluation_config["query_draws"]), evaluation_seed,
evaluation_config["recall_at"], evaluation_config["ndcg_at"],
)
ensemble_validation = validation[validation["method"].str.startswith("ensemble_a")].copy()
ensemble_validation["alpha"] = ensemble_validation["method"].str.replace(
"ensemble_a", "", regex=False
).astype(float)
alpha = select_ensemble_alpha(ensemble_validation, evaluation_config["primary_metric"])
methods: dict[str, Any] = {
"pfam_jaccard_max": jaccard,
"pfam_jaccard_mean": lambda candidates, references: pfam_jaccard_scores(
candidates, references, pfams, "mean"
),
"raw_esm_mean": lambda candidates, references: cosine_scores(
candidates, references, raw_mean, "mean"
),
"raw_esm_max": lambda candidates, references: cosine_scores(
candidates, references, raw_max, "mean"
),
"setnet": learned_cosine,
f"ensemble_validation_alpha_{alpha:.1f}": lambda candidates, references: ensemble_scores(
learned_cosine(candidates, references), jaccard(candidates, references), alpha
),
}
test_results = evaluate_retrieval(
assignments, "test", methods, int(evaluation_config["reference_size"]),
int(evaluation_config["query_draws"]), evaluation_seed,
evaluation_config["recall_at"], evaluation_config["ndcg_at"],
)
run_dir = create_run_directory(paths["run_root"], args.run_id)
metrics = [evaluation_config["primary_metric"], "mrr", "map", "ndcg@50", "tie_fraction"]
metadata = {
"schema_version": 1,
"selected_alpha": alpha,
"alpha_selected_on": "validation",
"split_sha256": sha256_file(split_path),
"checkpoint_sha256": sha256_file(args.checkpoint),
"label_tiers": sorted(assignments["label_tier"].unique()),
"pfam_feature": "BGC-level pfam_ids inventory; vocabulary built from training BGCs only",
"pfam_vocab_size": len(pfam_vocab) + 2,
"publication_eligible": bool((assignments["label_tier"] == "gold").any()),
"claim_warning": "Silver-only results are development evidence, not the primary biological result.",
}
write_paper_outputs(
run_dir, test_results, metrics, int(evaluation_config["bootstrap_samples"]),
float(evaluation_config["confidence_level"]), evaluation_seed,
metadata,
)
validation.to_csv(run_dir / "validation_alpha_search.csv", index=False)
print(json.dumps(metadata, indent=2, sort_keys=True))
def build_parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser(prog="bgc-rebuild")
subparsers = parser.add_subparsers(dest="command", required=True)
for name, function in (("audit", command_audit), ("build-splits", command_build_splits)):
subparser = subparsers.add_parser(name)
subparser.add_argument("--config", default="configs/main.yaml")
subparser.set_defaults(function=function)
train = subparsers.add_parser("train")
train.add_argument("--config", default="configs/main.yaml")
train.add_argument("--stage", choices=("phase1", "phase2"), required=True)
train.add_argument("--run-id", required=True)
train.add_argument("--split")
train.add_argument("--phase1-checkpoint")
train.set_defaults(function=command_train)
evaluate = subparsers.add_parser("evaluate")
evaluate.add_argument("--config", default="configs/main.yaml")
evaluate.add_argument("--checkpoint", required=True)
evaluate.add_argument("--run-id", required=True)
evaluate.add_argument("--split")
evaluate.set_defaults(function=command_evaluate)
return parser
def main() -> None:
args = build_parser().parse_args()
args.function(args)
if __name__ == "__main__":
main()
|