sentence-transformers/finetuning.py

# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved

import pdb
import os
import time
import sys
import torch
from torch.utils.tensorboard import SummaryWriter
import logging
import json
import numpy as np
import torch.distributed as dist
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler

from src.options import Options
from src import data, beir_utils, slurm, dist_utils, utils, contriever, finetuning_data, inbatch

import train

os.environ["TOKENIZERS_PARALLELISM"] = "false"

logger = logging.getLogger(__name__)


def finetuning(opt, model, optimizer, scheduler, tokenizer, step):

    run_stats = utils.WeightedAvgStats()

    tb_logger = utils.init_tb_logger(opt.output_dir)

    if hasattr(model, "module"):
        eval_model = model.module
    else:
        eval_model = model
    eval_model = eval_model.get_encoder()

    train_dataset = finetuning_data.Dataset(
        datapaths=opt.train_data,
        negative_ctxs=opt.negative_ctxs,
        negative_hard_ratio=opt.negative_hard_ratio,
        negative_hard_min_idx=opt.negative_hard_min_idx,
        normalize=opt.eval_normalize_text,
        global_rank=dist_utils.get_rank(),
        world_size=dist_utils.get_world_size(),
        maxload=opt.maxload,
        training=True,
    )
    collator = finetuning_data.Collator(tokenizer, passage_maxlength=opt.chunk_length)
    train_sampler = RandomSampler(train_dataset)
    train_dataloader = DataLoader(
        train_dataset,
        sampler=train_sampler,
        batch_size=opt.per_gpu_batch_size,
        drop_last=True,
        num_workers=opt.num_workers,
        collate_fn=collator,
    )

    train.eval_model(opt, eval_model, None, tokenizer, tb_logger, step)
    evaluate(opt, eval_model, tokenizer, tb_logger, step)

    epoch = 1

    model.train()
    prev_ids, prev_mask = None, None
    while step < opt.total_steps:
        logger.info(f"Start epoch {epoch}, number of batches: {len(train_dataloader)}")
        for i, batch in enumerate(train_dataloader):
            batch = {key: value.cuda() if isinstance(value, torch.Tensor) else value for key, value in batch.items()}
            step += 1

            train_loss, iter_stats = model(**batch, stats_prefix="train")
            train_loss.backward()

            if opt.optim == "sam" or opt.optim == "asam":
                optimizer.first_step(zero_grad=True)

                sam_loss, _ = model(**batch, stats_prefix="train/sam_opt")
                sam_loss.backward()
                optimizer.second_step(zero_grad=True)
            else:
                optimizer.step()
            scheduler.step()
            optimizer.zero_grad()

            run_stats.update(iter_stats)

            if step % opt.log_freq == 0:
                log = f"{step} / {opt.total_steps}"
                for k, v in sorted(run_stats.average_stats.items()):
                    log += f" | {k}: {v:.3f}"
                    if tb_logger:
                        tb_logger.add_scalar(k, v, step)
                log += f" | lr: {scheduler.get_last_lr()[0]:0.3g}"
                log += f" | Memory: {torch.cuda.max_memory_allocated()//1e9} GiB"

                logger.info(log)
                run_stats.reset()

            if step % opt.eval_freq == 0:

                train.eval_model(opt, eval_model, None, tokenizer, tb_logger, step)
                evaluate(opt, eval_model, tokenizer, tb_logger, step)

                if step % opt.save_freq == 0 and dist_utils.get_rank() == 0:
                    utils.save(
                        eval_model,
                        optimizer,
                        scheduler,
                        step,
                        opt,
                        opt.output_dir,
                        f"step-{step}",
                    )
                model.train()

            if step >= opt.total_steps:
                break

        epoch += 1


def evaluate(opt, model, tokenizer, tb_logger, step):
    dataset = finetuning_data.Dataset(
        datapaths=opt.eval_data,
        normalize=opt.eval_normalize_text,
        global_rank=dist_utils.get_rank(),
        world_size=dist_utils.get_world_size(),
        maxload=opt.maxload,
        training=False,
    )
    collator = finetuning_data.Collator(tokenizer, passage_maxlength=opt.chunk_length)
    sampler = SequentialSampler(dataset)
    dataloader = DataLoader(
        dataset,
        sampler=sampler,
        batch_size=opt.per_gpu_batch_size,
        drop_last=False,
        num_workers=opt.num_workers,
        collate_fn=collator,
    )

    model.eval()
    if hasattr(model, "module"):
        model = model.module
    correct_samples, total_samples, total_step = 0, 0, 0
    all_q, all_g, all_n = [], [], []
    with torch.no_grad():
        for i, batch in enumerate(dataloader):
            batch = {key: value.cuda() if isinstance(value, torch.Tensor) else value for key, value in batch.items()}

            all_tokens = torch.cat([batch["g_tokens"], batch["n_tokens"]], dim=0)
            all_mask = torch.cat([batch["g_mask"], batch["n_mask"]], dim=0)

            q_emb = model(input_ids=batch["q_tokens"], attention_mask=batch["q_mask"], normalize=opt.norm_query)
            all_emb = model(input_ids=all_tokens, attention_mask=all_mask, normalize=opt.norm_doc)

            g_emb, n_emb = torch.split(all_emb, [len(batch["g_tokens"]), len(batch["n_tokens"])])

            all_q.append(q_emb)
            all_g.append(g_emb)
            all_n.append(n_emb)

        all_q = torch.cat(all_q, dim=0)
        all_g = torch.cat(all_g, dim=0)
        all_n = torch.cat(all_n, dim=0)

        labels = torch.arange(0, len(all_q), device=all_q.device, dtype=torch.long)

        all_sizes = dist_utils.get_varsize(all_g)
        all_g = dist_utils.varsize_gather_nograd(all_g)
        all_n = dist_utils.varsize_gather_nograd(all_n)
        labels = labels + sum(all_sizes[: dist_utils.get_rank()])

        scores_pos = torch.einsum("id, jd->ij", all_q, all_g)
        scores_neg = torch.einsum("id, jd->ij", all_q, all_n)
        scores = torch.cat([scores_pos, scores_neg], dim=-1)

        argmax_idx = torch.argmax(scores, dim=1)
        sorted_scores, indices = torch.sort(scores, descending=True)
        isrelevant = indices == labels[:, None]
        rs = [r.cpu().numpy().nonzero()[0] for r in isrelevant]
        mrr = np.mean([1.0 / (r[0] + 1) if r.size else 0.0 for r in rs])

        acc = (argmax_idx == labels).sum() / all_q.size(0)
        acc, total = dist_utils.weighted_average(acc, all_q.size(0))
        mrr, _ = dist_utils.weighted_average(mrr, all_q.size(0))
        acc = 100 * acc

        message = []
        if dist_utils.is_main():
            message = [f"eval acc: {acc:.2f}%", f"eval mrr: {mrr:.3f}"]
            logger.info(" | ".join(message))
            if tb_logger is not None:
                tb_logger.add_scalar(f"eval_acc", acc, step)
                tb_logger.add_scalar(f"mrr", mrr, step)


def main():
    logger.info("Start")

    options = Options()
    opt = options.parse()

    torch.manual_seed(opt.seed)
    slurm.init_distributed_mode(opt)
    slurm.init_signal_handler()

    directory_exists = os.path.isdir(opt.output_dir)
    if dist.is_initialized():
        dist.barrier()
    os.makedirs(opt.output_dir, exist_ok=True)
    if not directory_exists and dist_utils.is_main():
        options.print_options(opt)
    if dist.is_initialized():
        dist.barrier()
    utils.init_logger(opt)

    step = 0

    retriever, tokenizer, retriever_model_id = contriever.load_retriever(opt.model_path, opt.pooling, opt.random_init)
    opt.retriever_model_id = retriever_model_id
    model = inbatch.InBatch(opt, retriever, tokenizer)

    model = model.cuda()

    optimizer, scheduler = utils.set_optim(opt, model)
    # if dist_utils.is_main():
    #    utils.save(model, optimizer, scheduler, global_step, 0., opt, opt.output_dir, f"step-{0}")
    logger.info(utils.get_parameters(model))

    for name, module in model.named_modules():
        if isinstance(module, torch.nn.Dropout):
            module.p = opt.dropout

    if torch.distributed.is_initialized():
        model = torch.nn.parallel.DistributedDataParallel(
            model,
            device_ids=[opt.local_rank],
            output_device=opt.local_rank,
            find_unused_parameters=False,
        )

    logger.info("Start training")
    finetuning(opt, model, optimizer, scheduler, tokenizer, step)


if __name__ == "__main__":
    main()