train.py

"""Train BERT tiny, evaluating three NanoBEIR datasets every 20%.

Adapted from the multi-vector training skill template and training_contrastive.py.
Run from the repository root with Python and the training dependencies installed.
Use --smoke-test for one step, or --push-to-hub to upload the best checkpoint.
Use --long-run to continue the initial model for 10,000 steps on the full dataset.
Run without --long-run first to create the local initial model.
"""

import argparse
import json
import logging
import shutil
from contextlib import nullcontext
from pathlib import Path

import torch
from datasets import load_dataset
from transformers import BertConfig, BertModel, BertTokenizer, TrainerCallback, set_seed

from sentence_transformers import (
    MultiVectorEncoder,
    MultiVectorEncoderModelCardData,
    MultiVectorEncoderTrainer,
    MultiVectorEncoderTrainingArguments,
)
from sentence_transformers.base.modules import Dense, Normalize, Transformer
from sentence_transformers.base.sampler import BatchSamplers
from sentence_transformers.multi_vector_encoder.evaluation import MultiVectorNanoBEIREvaluator
from sentence_transformers.multi_vector_encoder.losses import MultiVectorMultipleNegativesRankingLoss
from sentence_transformers.multi_vector_encoder.modules import MultiVectorMask

RUN_NAME = "bert-tiny-msmarco"
REPO_ID = "multi-vector-encoder-testing/bert-tiny-multi-vector"


class LogProgress(TrainerCallback):
    def on_log(self, args, state, control, logs=None, **kwargs):
        values = {
            key: value
            for key, value in (logs or {}).items()
            if key in ("loss", "learning_rate", "eval_loss", "eval_NanoBEIR_mean_maxsim_ndcg@10")
        }
        if values:
            logging.info("Step %s/%s: %s", state.global_step, state.max_steps, values)


def autocast_ctx():
    if not torch.cuda.is_available():
        return nullcontext()
    dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
    return torch.autocast("cuda", dtype=dtype)


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--smoke-test", action="store_true")
    parser.add_argument("--push-to-hub", action="store_true")
    parser.add_argument("--long-run", action="store_true")
    cli = parser.parse_args()
    repo_id = REPO_ID
    run_name = RUN_NAME + ("-long" if cli.long_run else "") + ("-smoke" if cli.smoke_test else "")
    output_dir = Path("models") / run_name
    output_dir.mkdir(parents=True, exist_ok=True)
    Path("logs").mkdir(exist_ok=True)
    logging.basicConfig(
        format="%(asctime)s - %(message)s",
        level=logging.INFO,
        handlers=[logging.StreamHandler(), logging.FileHandler(f"logs/{run_name}.log", mode="w")],
        force=True,
    )
    for noisy in ("httpx", "httpcore", "huggingface_hub", "urllib3", "filelock", "fsspec"):
        logging.getLogger(noisy).setLevel(logging.WARNING)
    set_seed(12)
    if torch.cuda.is_available():
        torch.set_float32_matmul_precision("high")

    card = MultiVectorEncoderModelCardData(
        language="en",
        license="mit",
        model_name="BERT tiny multi-vector encoder trained on MS MARCO",
        model_id=repo_id,
    )
    if cli.long_run:
        initial_model = Path("models") / RUN_NAME / "final"
        if not initial_model.is_dir():
            parser.error("Run without --long-run first to create the local initial model.")
        model = MultiVectorEncoder(str(initial_model), model_card_data=card)
        model.model_card_data.set_base_model("prajjwal1/bert-tiny")
    else:
        # The original checkpoint lacks model_type, which recent AutoConfig versions require.
        base_dir = output_dir / "base"
        base_model = BertModel.from_pretrained(
            "prajjwal1/bert-tiny", config=BertConfig.from_pretrained("prajjwal1/bert-tiny")
        )
        base_model.save_pretrained(base_dir)
        BertTokenizer.from_pretrained("prajjwal1/bert-tiny").save_pretrained(base_dir)
        del base_model
        transformer = Transformer(
            str(base_dir),
            query_length=32,
            document_length=256,
            query_expansion={"strategy": "min", "length": 32},
        )
        model = MultiVectorEncoder(
            modules=[
                transformer,
                Dense(128, 128, bias=False, activation_function=None, module_input_name="token_embeddings"),
                MultiVectorMask(),
                Normalize(module_input_name="token_embeddings"),
            ],
            model_card_data=card,
        )
        model.model_card_data.set_base_model("prajjwal1/bert-tiny")
    batch_size = 128 if cli.long_run else 32
    train_size, eval_size = (batch_size * 2, 32) if cli.smoke_test else (16_000, 128)
    split = "train" if cli.long_run and not cli.smoke_test else f"train[:{train_size + eval_size}]"
    dataset = load_dataset("sentence-transformers/msmarco-bm25", "triplet", split=split).select_columns(
        ["query", "positive", "negative"]
    )
    if cli.long_run and not cli.smoke_test:
        eval_size = 1024
    dataset = dataset.train_test_split(test_size=eval_size, seed=12)
    evaluator = MultiVectorNanoBEIREvaluator(dataset_names=["msmarco", "nq", "fiqa2018"], batch_size=64)
    logging.info("Baseline evaluation on three NanoBEIR datasets")
    with autocast_ctx():
        baseline_metrics = evaluator(model, output_path=str(output_dir), steps=0)
    baseline_eval = baseline_metrics[evaluator.primary_metric]
    full_evaluator = None
    full_baseline = None
    if cli.long_run and not cli.smoke_test:
        full_evaluator = MultiVectorNanoBEIREvaluator(batch_size=64)
        full_output = output_dir / "full_eval"
        full_output.mkdir(exist_ok=True)
        logging.info("Baseline evaluation on all 13 NanoBEIR datasets")
        with autocast_ctx():
            full_baseline = full_evaluator(model, output_path=str(full_output), steps=0)
    (output_dir / "baseline.json").write_text(
        json.dumps({"selection": baseline_metrics, "full": full_baseline}, indent=2), encoding="utf-8"
    )

    args = MultiVectorEncoderTrainingArguments(
        output_dir=str(output_dir),
        max_steps=1 if cli.smoke_test else 10_000 if cli.long_run else 500,
        per_device_train_batch_size=batch_size,
        per_device_eval_batch_size=32,
        learning_rate=1e-5 if cli.long_run else 3e-5,
        weight_decay=0.01,
        warmup_steps=0.05,
        bf16=torch.cuda.is_available() and torch.cuda.is_bf16_supported(),
        fp16=torch.cuda.is_available() and not torch.cuda.is_bf16_supported(),
        batch_sampler=BatchSamplers.NO_DUPLICATES,
        eval_strategy="steps",
        eval_steps=0.2,
        save_strategy="steps",
        save_steps=0.2,
        save_total_limit=2,
        logging_steps=0.005 if cli.long_run else 0.02,
        logging_first_step=True,
        disable_tqdm=True,
        load_best_model_at_end=True,
        metric_for_best_model=f"eval_{evaluator.primary_metric}",
        greater_is_better=True,
        report_to="none",
        run_name=run_name,
        seed=12,
    )
    trainer = MultiVectorEncoderTrainer(
        model=model,
        args=args,
        train_dataset=dataset["train"],
        eval_dataset=dataset["test"],
        loss=MultiVectorMultipleNegativesRankingLoss(model, scale=1.0),
        evaluator=evaluator,
        callbacks=[LogProgress()],
    )
    logging.info("Training configuration: %s", args.to_dict())
    trainer.train()
    logging.info("Evaluating the best checkpoint on the same three datasets")
    with autocast_ctx():
        final_metrics = evaluator(model, output_path=str(output_dir / "eval"))
    score = final_metrics[evaluator.primary_metric]
    full_final = None
    if full_evaluator is not None:
        logging.info("Evaluating the best checkpoint on all 13 NanoBEIR datasets")
        with autocast_ctx():
            full_final = full_evaluator(model, output_path=str(full_output))
        baseline_eval = full_baseline[full_evaluator.primary_metric]
        score = full_final[full_evaluator.primary_metric]
    delta = score - baseline_eval
    verdict = "WIN" if delta >= 0.005 else "MARGINAL" if delta >= 0 else "REGRESSION"
    logging.info("VERDICT: %s | score=%.4f | baseline=%.4f | delta=%+.4f", verdict, score, baseline_eval, delta)

    final_dir = output_dir / "final"
    model.save_pretrained(str(final_dir))
    shutil.copy2(__file__, final_dir / "train.py")
    results = {
        "baseline": baseline_metrics,
        "final": final_metrics,
        "best_checkpoint": trainer.state.best_model_checkpoint,
        "history": trainer.state.log_history,
        "verdict": verdict,
        "full_baseline": full_baseline,
        "full_final": full_final,
        "configuration": vars(cli),
        "training_args": args.to_dict(),
    }
    (final_dir / "results.json").write_text(json.dumps(results, indent=2), encoding="utf-8")
    logging.info("Saved model, training script, and metrics to %s", final_dir)
    if cli.push_to_hub and not cli.smoke_test:
        try:
            url = model.push_to_hub(repo_id, local_model_path=str(final_dir))
            logging.info("Uploaded to %s", url)
        except Exception:
            logging.exception("Hub upload failed. The model is saved at %s", final_dir)


if __name__ == "__main__":
    main()