预处理程序

import os
import random
import time
from pathlib import Path
from typing import Optional, Dict, Any

import fire
import numpy as np
import torch
from torch.cuda import amp
from torch.utils.data import DataLoader
from transformers import T5TokenizerFast, T5ForConditionalGeneration, TrainingArguments, T5Config, Adafactor
from transformers.utils.model_parallel_utils import get_device_map
from datasets import load_dataset
from pkot5.args import ARGS, CONFIGS
from pkot5.data import LargeCorpusDatasetFromServerV2, DataCollatorForSeq2Seq, NUM_EXTRA_IDS, \
    LargeCorpusDatasetFromServer, DataCollatorForT5MLM
import jsonlines


def read_jsonl(path):
    """Read jsonlines file into python list

    args:
        path - directory of the jsonlines file
    return:
        jsonlines file content in List
    """
    items = []
    with open(path, 'r', encoding='utf-8') as f:
        for item in jsonlines.Reader(f):
            items.append(item)
    return items


def train(model_size: str = "base", tokenizer_path: str = "paust/pko-t5-base", job_name: str = "pkot5-pretraining",
          resume_checkpoint: Optional[int] = None, version: Optional[str] = "1"):
    local_rank = int(os.getenv("LOCAL_RANK", "-1"))
    model_size = model_size.lower()
    default_args = ARGS[model_size]
    n_pp = default_args.pop("pipeline_parallelism", 1)
    args = TrainingArguments(
        output_dir=f'./models/pko-t5/{model_size}',
        local_rank=local_rank,
        **default_args
    )

    random.seed(args.seed)
    torch.random.manual_seed(args.seed)
    np.random.seed(args.seed)

    tokenizer = T5TokenizerFast.from_pretrained(tokenizer_path, unk_token='<pad>', extra_ids=NUM_EXTRA_IDS)
    config = T5Config(**CONFIGS[model_size])
    config.dropout_rate = 0.0

    ##我加的
    all_data = []
    # data_try = read_jsonl('./simplified-nq-test.jsonl')
    qa_pairs = []
    qaid2dataid = {}
    # qaid = 0
    # input_texts = []
    # for data_dict in data_try:

    datasets = load_dataset('squad_kor_v1')
    for data_split in ['train', 'validation']:
        input_texts, label_texts = [], []
        for data in datasets[data_split]:
            input_texts.append(f"질의: {data['question']} 제목: {data['title']} 본문: {data['context']}".strip())
            label_texts.append(data['answers']['text'][0].strip())

        input_text_ids = tokenizer(input_texts, add_special_tokens=False).input_ids
        label_ids = tokenizer(label_texts, add_special_tokens=False).input_ids

        start = 0
        shengyu_pass = np.array([0] * 564)
        try_pass = np.array([0] * 564)
        for i in range(len(input_text_ids)):
            biaozhi = int(len(input_text_ids[i]) / 564)
            if biaozhi > 0:
                for j in range(biaozhi):
                    try_pass = np.array(input_text_ids[i])[j * 564:(j + 1) * 564]
                    all_data.append({
                        'input_ids': try_pass.copy(),
                    })
            if len(input_text_ids[i]) > biaozhi * 564:
                zhongjian = len(input_text_ids[i]) - biaozhi * 564
                if zhongjian + start < 564:
                    shengyu_pass[start:start + zhongjian] = np.array(input_text_ids[i])[biaozhi * 564:]
                    start = start + zhongjian
                else:
                    shengyu_pass[start:] = np.array(input_text_ids[i])[biaozhi * 564:biaozhi * 564 + 564 - start]
                    all_data.append({

                        'input_ids': shengyu_pass.copy(),
                    })
                    shengyu_pass[:len(input_text_ids[i]) - (biaozhi * 564 + 564 - start)] = np.array(input_text_ids[i])[
                                                                                            biaozhi * 564 + 564 - start:]
                    start = len(input_text_ids[i]) - (biaozhi * 564 + 564 - start)

    if n_pp > 1:
        torch.cuda.set_device(local_rank * n_pp)
        model = T5ForConditionalGeneration(config)
        devices = list(range(local_rank * n_pp, (local_rank + 1) * n_pp))
        model.parallelize(get_device_map(len(model.encoder.block), devices))
        ddp_model = torch.nn.parallel.DistributedDataParallel(model)
    else:
        # torch.cuda.set_device(local_rank)
        # model = T5ForConditionalGeneration(config).cuda()
        model = T5ForConditionalGeneration(config)
        # ddp_model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[local_rank])

    optimizer = Adafactor(model.parameters(), scale_parameter=True, relative_step=True, warmup_init=True)
    scaler = amp.GradScaler()
    step = 0

    if resume_checkpoint is not None:
        step = resume_checkpoint
        ckpt_dir = Path(args.output_dir + f'/checkpoint-{step}')
        model.load_state_dict(torch.load(ckpt_dir / "pytorch_model.bin", map_location='cpu'))
        optimizer.load_state_dict(torch.load(ckpt_dir / "optimizer.pt", map_location='cpu'))

    if version == "1":
        # train_data = LargeCorpusDatasetFromServer(job_name, grpc_endpoint, seed=args.data_seed)
        train_loader = DataLoader(all_data, batch_size=args.per_device_train_batch_size,
                                  collate_fn=DataCollatorForT5MLM(tokenizer, prefix="fill: "))
    # else:
    # train_data = LargeCorpusDatasetFromServerV2(job_name, tokenizer, grpc_endpoint, seed=args.data_seed)
    # train_loader = DataLoader(train_data, batch_size=args.per_device_train_batch_size, collate_fn=DataCollatorForSeq2Seq(tokenizer))
    print(f"Start pretraining of t5-{model_size}")
    epoch=0
    max_epoch=500
    for epoch in range(max_epoch):
        while step < args.max_steps:
            total_loss = 0
            dt = time.time()
            gradient_accumulation_step = 0
            optimizer.zero_grad()
            for data in train_loader:
                if step >= args.max_steps:
                    break

                data = data.convert_to_tensors('pt').to(device='cuda')
                with amp.autocast():
                    loss = ddp_model(**data).loss
                    loss = loss / args.gradient_accumulation_steps

                scaler.scale(loss).backward(retain_graph=True)
                gradient_accumulation_step += 1
                if gradient_accumulation_step == args.gradient_accumulation_steps:
                    scaler.unscale_(optimizer)
                    torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
                    scaler.step(optimizer)
                    scaler.update()
                    optimizer.zero_grad()
                    dt = time.time() - dt

                    step += 1
                    total_loss = loss.detach().item() + total_loss
                    print(f"step={step} loss={total_loss:.4f} time={dt:.4f}s")

                    if step % args.save_steps == 0:
                        ckpt_dir = Path(args.output_dir) / f"checkpoint-{step}"
                        ckpt_dir.mkdir(exist_ok=True, parents=True)
                        config.dropout_rate = 0.1
                        if local_rank == 0:
                            config.save_pretrained(ckpt_dir)
                            torch.save(model.state_dict(), ckpt_dir / "pytorch_model.bin")
                            torch.save(optimizer.state_dict(), ckpt_dir / "optimizer.pt")

                    gradient_accumulation_step = 0
                    total_loss = 0
                    dt = time.time()
                else:
                    total_loss = loss.detach().item() + total_loss

    print(f"End of pretraining t5-{model_size}")


if __name__ == '__main__':
    fire.Fire(train)