def train(args: argparse.Namespace) -> None:
"""Train from RoBERTa, keeping this model's best checkpoint."""
inputs = args.inputs.resolve()
run_dir = (args.out_dir / args.implementation).resolve()
run_dir.mkdir(parents=True, exist_ok=True)
set_seed(SEED)
device = torch.device("cuda")
rel2id = load_json(inputs / "rel2id.json")
id2rel = {value: key for key, value in rel2id.items()}
tokenizer = AutoTokenizer.from_pretrained(ENCODER)
train_documents = load_json(inputs / "train.json")
train_features = prepare_features(train_documents, tokenizer, MAX_LENGTH)
add_relation_targets(train_features, train_documents, rel2id)
if args.implementation == "dreeam":
add_evidence_targets(train_features, train_documents)
dev_documents = load_json(inputs / "dev.json")
dev_features = prepare_features(dev_documents, tokenizer, MAX_LENGTH)
model = build_model(args.implementation, tokenizer, len(rel2id))
model.to(device)
added_names = ("extractor", "bilinear")
encoder_params, head_params = [], []
for name, parameter in model.named_parameters():
(head_params if any(part in name for part in added_names) else encoder_params).append(parameter)
optimizer = torch.optim.AdamW(
[
{"params": encoder_params, "lr": ENCODER_LR},
{"params": head_params, "lr": HEAD_LR},
],
eps=1e-6,
)
train_loader = DataLoader(
train_features,
batch_size=BATCH_SIZE,
shuffle=True,
collate_fn=collate_fn,
drop_last=False,
)
updates_per_epoch = int(np.ceil(len(train_loader) / GRADIENT_ACCUMULATION))
total_updates = updates_per_epoch * args.epochs
scheduler = get_linear_schedule_with_warmup(optimizer, int(total_updates * 0.06), total_updates)
scaler = torch.amp.GradScaler("cuda")
selection = EarlyStopping(patience=None) # every epoch runs; the best dev F1 is kept
dev_gold = label_triples(dev_documents)
optimizer.zero_grad(set_to_none=True)
for _ in range(args.epochs):
train_epoch(model, train_loader, optimizer, scheduler, scaler)
dev_array = infer(model, dev_features, EVAL_BATCH_SIZE)
dev_pred = predicted_triples(dev_array, dev_features, id2rel)
metrics = precision_recall_f1(dev_gold, dev_pred)
selection.update(metrics["f1"], lambda: torch.save(model.state_dict(), run_dir / "best.pt"))