pytorch单机多GPU训练方式（微调bert）

本来是用Dataparallel来微调，但是发现gpu0上负载太大了，后来上面一看，pytorch官网推荐使用DistributedDataParallel，这个方法可以用于多机多卡或者单机多卡，速度和各方面都比Dataparallel要好很多。
我主要看了下面两个博客：
【分布式训练】单机多卡的正确打开方式（三）：PyTorch
pytorch(分布式)数据并行个人实践总结——DataParallel/DistributedDataParallel

这里说一下Dataparallel和DistributedDataparallel在使用时的区别。比如有三张卡（也就是3张gpu），Batch=30。

1. Dataparallel只开一个进程，DistributedDataparallel有几张卡开几个进程，不过需要在执行程序时手动设置。
2. Dataparallel返回的loss是3张卡的loss，所以需要mean一下，这样才是整个batch的loss，也就是Dataparallel会自动将输入的一个batch的数据切分成3份，每张卡分别计算一部分，也就是10份。
   而DistributedDataparallel是不会自动切分数据的，所以在Dataloader中，需要将sampler设置为DistributedSampler，具体见下面的代码。DistributedDataparallel是将整个数据集切分为3份，然后每一个进程使用其中的一份，但是每个进程的batch还是30。
3. 在保存模型时，Dataparallel和DistributedDataparallel都需要使用model.module.state_dict来保存模型，而不是model.state_dict，不过DistributedDataparallel还需要选择一个进程的模型保存，具体见代码。
4. 执行DistributedDataparallel的程序是要用命令行执行，python -m torch.distributed.launch --nproc_per_node=3 main.py, 其中 nproc_per_node指定了要用为每一张卡开一个进程。

import random
import numpy as np
from transformers import AlbertTokenizer, AlbertForMaskedLM, AdamW, get_linear_schedule_with_warmup, \
    DataCollatorForLanguageModeling
from torch.nn.parallel import DistributedDataParallel
from torch.utils.data import Dataset, DataLoader, RandomSampler
from torch.utils.data.distributed import DistributedSampler
import torch
import time
from utils import format_time
from utils import read_data

torch.distributed.init_process_group(backend='nccl')

tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2', cache_dir='../language_model/albert')
model = AlbertForMaskedLM.from_pretrained('../language_model/albert')

seed_val = 42
random.seed(seed_val)
np.random.seed(seed_val)
torch.manual_seed(seed_val)
torch.cuda.manual_seed_all(seed_val)


def my_collate_fn(batch):
    input_ids = []
    attention_masks = []
    for sent in batch:
        encoded_dict = tokenizer.encode_plus(
            sent[0],  # 输入文本
            add_special_tokens=True,  # 添加 '[CLS]' 和 '[SEP]'
            max_length=128,  # 填充 & 截断长度
            truncation=True,
            pad_to_max_length=True,
            return_attention_mask=True,  # 返回 attn. masks.
        )
        input_ids.append(torch.tensor(encoded_dict['input_ids']))

        attention_masks.append(torch.tensor(encoded_dict['attention_mask']))
    input_ids = torch.stack(input_ids, dim=0)
    attention_masks = torch.stack(attention_masks, dim=0)
    return [input_ids, attention_masks]


class myDataset(Dataset):

    def __init__(self, data):
        self.data = data

    def __getitem__(self, index):
        return self.data[index]

    def __len__(self):
        return len(self.data)


def make_data(batch_size):
    data = read_data('en')
    dataset = myDataset(data)

    data_dataloader = DataLoader(
        dataset,
        sampler=DistributedSampler(dataset),  # 这里也是分布式的sampler
        batch_size=batch_size,
        collate_fn=my_collate_fn,
        num_workers=4,
    )

    return data_dataloader

'''
这个方法也是bert自带的实现了随机mask 15%单词任务的方法。
由于我是要在领域内的数据集上微调bert，所以还是用language model的训练方法。
'''
datacollecter = DataCollatorForLanguageModeling(tokenizer)

def trian(model, train_loader, optimizer, scheduler, local_rank):
    for epoch_i in range(0, epochs):

        # ========================================
        #               Training
        # ========================================

        print("")
        print('======== Epoch {:} / {:} ========'.format(epoch_i + 1, epochs))
        print('Training...')

        total_train_loss = 0
        model.train()
        t0 = time.time()

        for step, batch in enumerate(train_loader):

            if step % 50 == 0 and not step == 0:
                elapsed = format_time(time.time() - t0)
                print('  Batch {:>5,}  of  {:>5,}.    Elapsed: {:}.'.format(step, len(train_loader), elapsed))

            b_input_ids, b_attention_mask = batch[0], batch[1]
            output = datacollecter(b_input_ids)
            b_input_ids, b_labels = output["input_ids"], output["labels"]
            b_input_ids = b_input_ids.to(device)
            b_attention_mask = b_attention_mask.to(device)

            model.zero_grad()
            loss, logits = model(b_input_ids, attention_mask=b_attention_mask, labels=b_labels)

            total_train_loss += loss.item()

            loss.backward()
            optimizer.step()
            scheduler.step()

        avg_train_loss = total_train_loss / len(train_loader)

        training_time = format_time(time.time() - t0)

        print("")
        print("  Average training loss: {0:.2f}".format(avg_train_loss))
        print("  Training epcoh took: {:}".format(training_time))

    ## 选择一个进程保存
    if local_rank == 0:
        model.module.save_pretrained(save_path)  # save_pretrained是bert自带的保存微调模型的方法
        print('Saving model in %s.' % save_path)

    print("")
    print("Training complete!")
    print("Total training took {:} (h:mm:ss)".format(format_time(time.time() - total_t0)))


epochs = 2
batch_size = 128
save_path = './model/'


if __name__ == "__main__":

    # 为每个进程配置GPU
    local_rank = torch.distributed.get_rank()
    torch.cuda.set_device(local_rank)
    device = torch.device("cuda", local_rank)

    # 要先将model放到gpu上
    model = model.to(device)
    model = DistributedDataParallel(model, find_unused_parameters=True, device_ids=[local_rank],
                                    output_device=local_rank)

    training_stats = []
    total_t0 = time.time()

    train_loader = make_data(batch_size)
    total_steps = len(train_loader) * epochs

    optimizer = AdamW(model.parameters(),
                      lr=2e-5,  # args.learning_rate - default is 5e-5
                      eps=1e-8  # args.adam_epsilon  - default is 1e-8
                      )

    scheduler = get_linear_schedule_with_warmup(optimizer,
                                                num_warmup_steps=1000,
                                                num_training_steps=total_steps)

    trian(model, train_loader, optimizer, scheduler, local_rank)

本文地址：https://blog.csdn.net/mch2869253130/article/details/108239063