| from __future__ import annotations |
|
|
| import argparse |
| import json |
| import math |
| import os |
| from pathlib import Path |
| import random |
| import sys |
|
|
| sys.path.insert(0, str(Path(__file__).resolve().parents[1])) |
|
|
| import numpy as np |
| import torch |
| from torch.optim import AdamW |
| from torch.utils.data import DataLoader |
| from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup |
|
|
| from train_candidate_verifier import VerifierDataset, build_examples, score |
|
|
|
|
| def load_jsonl(path: Path) -> list[dict]: |
| return [json.loads(line) for line in path.open(encoding="utf-8") if line.strip()] |
|
|
|
|
| def main() -> None: |
| parser = argparse.ArgumentParser(description="Train a full-data phenotype candidate verifier and score deployment candidates.") |
| parser.add_argument("--train-data", type=Path, required=True) |
| parser.add_argument("--train-candidates", type=Path, required=True) |
| parser.add_argument("--predict-data", type=Path, required=True) |
| parser.add_argument("--predict-candidates", type=Path, required=True) |
| parser.add_argument("--model", type=Path, required=True) |
| parser.add_argument("--output", type=Path, required=True) |
| parser.add_argument("--epochs", type=int, default=3) |
| parser.add_argument("--batch-size", type=int, default=16) |
| parser.add_argument("--learning-rate", type=float, default=2e-5) |
| parser.add_argument("--max-length", type=int, default=256) |
| parser.add_argument("--seed", type=int, default=2026) |
| parser.add_argument("--device", default="cuda") |
| args = parser.parse_args() |
|
|
| os.environ.setdefault("HF_HUB_OFFLINE", "1") |
| os.environ.setdefault("TRANSFORMERS_OFFLINE", "1") |
| random.seed(args.seed) |
| np.random.seed(args.seed) |
| torch.manual_seed(args.seed) |
| train_documents = load_jsonl(args.train_data) |
| predict_documents = load_jsonl(args.predict_data) |
| train_candidates = json.loads(args.train_candidates.read_text(encoding="utf-8")) |
| predict_candidates = json.loads(args.predict_candidates.read_text(encoding="utf-8")) |
| train_examples = build_examples(train_documents, train_candidates) |
| predict_examples = build_examples(predict_documents, predict_candidates) |
| tokenizer = AutoTokenizer.from_pretrained(args.model, local_files_only=True, use_fast=True) |
| train_loader = DataLoader( |
| VerifierDataset(train_examples, tokenizer, args.max_length), batch_size=args.batch_size, |
| shuffle=True, generator=torch.Generator().manual_seed(args.seed), |
| num_workers=2, pin_memory=True, |
| ) |
| predict_loader = DataLoader( |
| VerifierDataset(predict_examples, tokenizer, args.max_length), batch_size=args.batch_size, |
| shuffle=False, num_workers=2, pin_memory=True, |
| ) |
| model = AutoModelForSequenceClassification.from_pretrained( |
| args.model, local_files_only=True, num_labels=2, ignore_mismatched_sizes=True |
| ).to(args.device) |
| positives = sum(example.label for example in train_examples) |
| class_weights = torch.tensor( |
| [1.0, (len(train_examples) - positives) / max(1, positives)], device=args.device |
| ) |
| optimizer = AdamW(model.parameters(), lr=args.learning_rate, weight_decay=0.01) |
| updates = len(train_loader) * args.epochs |
| scheduler = get_linear_schedule_with_warmup(optimizer, max(1, updates // 10), updates) |
| history: list[dict[str, float | int]] = [] |
| for epoch in range(1, args.epochs + 1): |
| model.train() |
| total_loss = 0.0 |
| for batch in train_loader: |
| batch.pop("index") |
| labels = batch.pop("labels").to(args.device) |
| inputs = {key: value.to(args.device, non_blocking=True) for key, value in batch.items()} |
| with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=args.device.startswith("cuda")): |
| logits = model(**inputs).logits |
| loss = torch.nn.functional.cross_entropy(logits.float(), labels, weight=class_weights) |
| loss.backward() |
| torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0) |
| optimizer.step() |
| scheduler.step() |
| optimizer.zero_grad(set_to_none=True) |
| total_loss += float(loss.item()) |
| report = {"epoch": epoch, "train_loss": total_loss / max(1, len(train_loader))} |
| history.append(report) |
| print(json.dumps(report), flush=True) |
|
|
| scores = score(model, predict_loader, args.device, len(predict_examples)) |
| predictions = [ |
| { |
| "pmc_id": example.pmc_id, "start": example.start, "end": example.end, |
| "text": example.text, "score": value, **example.metadata, |
| } |
| for example, value in zip(predict_examples, scores, strict=True) |
| ] |
| result = { |
| "configuration": vars(args) | { |
| "train_data": str(args.train_data), "train_candidates": str(args.train_candidates), |
| "predict_data": str(args.predict_data), "predict_candidates": str(args.predict_candidates), |
| "model": str(args.model), "output": str(args.output), |
| }, |
| "train_examples": len(train_examples), "train_positives": positives, |
| "predict_examples": len(predict_examples), "history": history, |
| "predictions": predictions, |
| } |
| args.output.parent.mkdir(parents=True, exist_ok=True) |
| args.output.write_text(json.dumps(result, ensure_ascii=False), encoding="utf-8") |
| print(json.dumps({"train_examples": len(train_examples), "predict_examples": len(predict_examples), "output": str(args.output)})) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
