Pytorch 多卡并行训练教程 (DDP)

在使用GPU训练大模型时,往往会面临单卡显存不足的情况,这时候就希望通过多卡并行的形式来扩大显存。PyTorch主要提供了两个类来实现多卡并行分别是

  • torch.nn.DataParallel(DP)
  • torch.nn.DistributedDataParallel(DDP)

关于这两者的区别和原理也有许多博客如Pytorch 并行训练(DP, DDP)的原理和应用; DDP系列第一篇:入门教程进行总结,这里就不在赘述了。不过总结来说的话:DP 比较简单,对小白比较友好,一行代码便可以搞定。DDP 每个进程对应一个独立的训练过程,且只对梯度等少量数据进行信息交换。每个进程包含独立的解释器和 GIL。

博主能力有限,很多原理上的东西看得不是特别懂,所以理解起来也比较肤浅,但是编程的时候一直没找到一套合适的蓝本,最终参考了很多网上的博客,吭哧吭哧写了一套不会报错的代码出来,下面把我个人的理解整理出来,不当之处希望大家指出,一起交流学习。后续可能会随着自己的理解的加深持续完善。
主要参考了以下一些博客:

  • PyTorch 并行训练指南:单机多卡并行、混合精度、同步 BN 训练
  • Pytorch 并行训练(DP, DDP)的原理和应用
  • pytorch多gpu并行训练
  • DDP系列第一篇:入门教程
  • 单机多卡训练 踩坑记录

初始化

增加参数local_rank来确定当前进程使用哪块GPU, 用于在每个进程中指定不同的device。

def parse():parser = argparse.ArgumentParser()parser.add_argument('--local_rank', type=int, default=0)args = parser.parse_args()return argsdef main():args = parse()torch.cuda.set_device(args.local_rank)torch.distributed.init_process_group('nccl',init_method='env://')device = torch.device(f'cuda:{args.local_rank}')

其中 torch.distributed.init_process_group 用于初始化GPU通信方式(NCCL)和参数的获取方式(env代表通过环境变量)。

设置随机种子点

假如model中用到了随机数种子来保证可复现性, 那么此时不能再用固定的常数作为seed, 否则会导致DDP中的所有进程都拥有一样的seed, 进而生成同态性的数据, 因此需要在程序中显示地设置随机种子点。

 # 固定随机种子点seed = np.random.randint(1, 10000)np.random.seed(seed)torch.manual_seed(seed)torch.cuda.manual_seed_all(seed)

Dataloader

对于数据加载,在初始化 data loader 的时候需要使用到 torch.utils.data.distributed.DistributedSampler 这个函数:

train_dataset = ...train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset, shuffle=True) # 这个sampler会自动分配数据到各个gpu上train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=opts.batch_size, sampler=train_sampler)

通过以上的函数便可以给每个进程一个不同的 sampler,告诉每个进程自己分别取哪些数据。

在每一个epoch开始的阶段需要为sampler重新设定eopch即:

for ep in range(total_epoch):train_sampler.set_epoch(ep)

这样做的目的是:如果在DistributedSampler设置了shuffle,DistributedSampler使用当前epoch作为随机数种子,从而使得不同epoch下有不同的shuffle结果,但是在DistributedSampler源代码中默认的epoch为0,那么每次dataloader获取的shuffle都是相同的。所以,每次 epoch 开始前都需要要调用 sampler 的 set_epoch 方法,这样才能让数据集随机 shuffle 起来。

模型初始化

对于模型的处理主要包括模型初始化,将模型加载至CUDA;加载预训练权重;或利用主进程的权重 初始化所有的进程;将模型中的BN转换为SyncBN;设置模型并行。

由于 BN 层需要基于传入模型的数据计算均值和方差,造成普通 BN 在多卡模式下实际上就是单卡模式。此时需要使用 SyncBN 利用DDP的分布式计算接口来实现真正的多卡BN。

SyncBN利用分布式通讯接口在各卡间进行通讯,传输各自进程小 batch mean 和小 batch variance,在传输少量数据的基础上利用所有数据进行BN计算。

同时由于 SyncBN 用到 all_gather 这个分布式计算接口,而使用这个接口需要先初始化DDP环境,因此 SyncBN 需要在 DDP 环境初始化后初始化,但是要在 DDP 模型前就准备好。

最后由于 SyncBN 是直接搜索 model 中每个 module,如果这个 module 是 torch.nn.modules.batchnorm._BatchNorm 的子类,就将其替换为 SyncBN。因此如果你的 Normalization 层是自己定义的特殊类,没有继承过 _BatchNorm 类,那么convert_sync_batchnorm 是不支持的,需要你自己实现一个新的SyncBN!

def parse():parser = argparse.ArgumentParser()parser.add_argument('--local_rank', type=int, default=0)parser.add_argument('--device', type=str, default='cuda', help='device id (i.e. 0 or 0,1 or cpu)')parser.add_argument('--resume', type=str, default=None, help='specified the dir of saved models for resume the training')args = parser.parse_args()return args
args = parse()device = torch.device(args.device)model = mymodel().to(device)if args.resume:checkpoint = torch.load(model_save_path, map_location=device)model.load_state_dict(checkpoint['model'])else:save_path = 'initial_weights.pth'if opts.local_rank == 0:torch.save(model.state_dict(), save_path)dist.barrier()# 这里注意,一定要指定map_location参数,否则会导致第一块GPU占用更多资源model.load_state_dict(torch.load(save_path, map_location=device))## 设置同步model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)## 设置模型并行model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True) ## 注意要使用find_unused_parameters=True,因为有时候模型里面定义的一些模块 在forward函数里面没有调用,如果不使用find_unused_parameters=True 会报错

输出日志设置

在每一次需要输出或打印日志时都应该先使用opts.local_rank == 0 来判断,也就是在主进程才执行一些操作,不然日志或者打印的结果会非常混乱。

logger = Noneif opts.local_rank == 0:log_dir = os.path.join(opts.display_dir, 'logger', opts.name)os.makedirs(log_dir, exist_ok=True)log_path = os.path.join(log_dir, 'log.txt')if os.path.exists(log_path):os.remove(log_path)logger = logger_config(log_path=log_path, logging_name='Timer')logger.info('Parameter Space: ABS: {:.1f}, REL: {:.4f}'.format(count_parameters(MPF_model), count_parameters(MPF_model) / 1024 / 1024))logger.info(MPF_model)

模型保存

state = {'model':model.module.state_dict(), 'ep':ep,'total_it':total_it}save_path = os.path.join(self.model_dir, 'model_{:0>5d}.pth'.format(ep))torch.save(state, save_path)

在保存模型是需要注意的是,保存的是{'model':model.module.state_dict()}, 而不是我们之前的{'model':model.state_dict()}, 因为在使用DDP后,原来的model会被封装为新的model的module属性里。

启动方式

PyTorch为提供了一个很方便的启动器 torch.distributed.lunch 用于启动文件,所以可以将运行训练代码的方式调整成下面这样:

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 train.py

最后附上完成了train代码和超参解析代码:

train.py

import torch.optim as optimfrom create_dataset import *from utils import *from MPFNet_Trans_skip import MPFNetfrom options import * from saver import Saver, resumefrom time import timefrom tqdm import tqdmfrom optimizer import Optimizerimport datetimeimport torch.distributed as distdef main():# parse optionsparser = TrainOptions()opts = parser.parse()# define model, optimiser and schedulertorch.cuda.set_device(opts.local_rank)torch.distributed.init_process_group('nccl', init_method='env://')# device = torch.device(f'cuda:{opts.local_rank}') #device 这样的设置可能会有问题device = torch.device(opts.gpu)# device = torch.device("cuda:{}".format(opts.gpu) if torch.cuda.is_available() else "cpu")# 固定随机种子seed = np.random.randint(1, 10000)np.random.seed(seed)torch.manual_seed(seed)torch.cuda.manual_seed_all(seed) # define datasettrain_dataset = MSRSData(opts, is_train=True)train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset, shuffle=True)train_loader = torch.utils.data.DataLoader(dataset=train_dataset,batch_size=opts.batch_size,num_workers = opts.nThreads,sampler=train_sampler,pin_memory=False,)test_dataset = MSRSData(opts, is_train=False)test_sampler = torch.utils.data.distributed.DistributedSampler(test_dataset)test_loader = torch.utils.data.DataLoader(dataset=test_dataset,batch_size=12,sampler=test_sampler,num_workers = opts.nThreads,)## 先加载dataloader 计算每个epoch的的迭代步数 然后计算总的迭代步数ep_iter = len(train_loader)max_iter = opts.n_ep * ep_iterif opts.local_rank == 0:print('Training iter: {}'.format(max_iter))print(opts.local_rank)## 初始化模型MPF_model = MPFNet(opts.class_nb).to(device)momentum = 0.9weight_decay = 5e-4lr_start = 1e-3# max_iter = 150000power = 0.9warmup_steps = 1000warmup_start_lr = 1e-5optimizer = Optimizer(model = MPF_model,lr0 = lr_start,momentum = momentum,wd = weight_decay,warmup_steps = warmup_steps,warmup_start_lr = warmup_start_lr,max_iter = max_iter,power = power)if opts.resume:if opts.local_rank == 0:MPF_model, ep, total_it = resume(MPF_model, opts.resume, device)optimizer = Optimizer(model = MPF_model,lr0 = lr_start,momentum = momentum,wd = weight_decay,warmup_steps = warmup_steps,warmup_start_lr = warmup_start_lr,max_iter = max_iter,power = power, it=total_it)lr = optimizer.get_lr()print('lr:{}'.format(lr))else: model_dir = os.path.join(opts.result_dir, opts.name)os.makedirs(model_dir, exist_ok=True)save_path = os.path.join(model_dir, 'initial_weights.pth')if opts.local_rank == 0:torch.save(MPF_model.state_dict(), save_path)dist.barrier()# 这里注意,一定要指定map_location参数,否则会导致第一块GPU占用更多资源MPF_model.load_state_dict(torch.load(save_path, map_location=device))ep = -1total_it = 0ep += 1MPF_model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(MPF_model)MPF_model = torch.nn.parallel.DistributedDataParallel(MPF_model, device_ids=[opts.local_rank], output_device=opts.local_rank, find_unused_parameters=True)# optimizer = optim.Adam(MPF_model.parameters(), lr=opts.lr)# scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=50, gamma=0.9)logger = Noneif opts.local_rank == 0:log_dir = os.path.join(opts.display_dir, 'logger', opts.name)os.makedirs(log_dir, exist_ok=True)log_path = os.path.join(log_dir, 'log.txt')if os.path.exists(log_path):os.remove(log_path)logger = logger_config(log_path=log_path, logging_name='Timer')logger.info('Parameter Space: ABS: {:.1f}, REL: {:.4f}'.format(count_parameters(MPF_model), count_parameters(MPF_model) / 1024 / 1024))logger.info(MPF_model) # Train and evaluate multi-task networkmulti_task_trainer(train_loader,train_sampler,test_loader,MPF_model,device,optimizer,opts,logger,ep,total_it)def multi_task_trainer(train_loader, train_sampler, test_loader, multi_task_model, device, optimizer, opt, logger=None, start_ep=0, total_it=0):total_epoch = opt.n_epsaver = Saver(opt)## 计算分割损失相关的设计score_thres = 0.7ignore_idx = 255n_min = 8 * 256 * 256 // 8criteria = OhemCELoss(thresh=score_thres, n_min=n_min, device=device, ignore_lb=ignore_idx)binary_class_weight = np.array([1.4548, 19.8962])binary_class_weight = torch.tensor(binary_class_weight).float().to(device)binary_class_weight = binary_class_weight.unsqueeze(0)binary_class_weight = binary_class_weight.unsqueeze(2)binary_class_weight = binary_class_weight.unsqueeze(2)lb_ignore = [255]if opt.resume:best_mIou = multi_task_tester(test_loader, multi_task_model, device, opt)else:best_mIou = 0.0if opt.local_rank == 0:print('best mIoU: {:.4f}'.format(best_mIou))start = glob_st = time()for ep in range(start_ep, total_epoch): ## 每一个epoch 计算一次动态权重train_sampler.set_epoch(ep)multi_task_model.train()seg_metric = SegmentationMetric(opt.class_nb, device=device) ## 这里可能会有问题 for it, (img_ir, img_vi, label, bi, bd, mask) in enumerate(train_loader):total_it += 1img_ir = img_ir.to(device)img_vi = img_vi.to(device)label = label.to(device)bi = bi.to(device).squeeze(1)bd = bd.to(device).squeeze(1)vi_Y, vi_Cb, vi_Cr = RGB2YCrCb(img_vi)vi_Y = vi_Y.to(device)vi_Cb = vi_Cb.to(device)vi_Cr = vi_Cr.to(device)mask = mask.to(device)seg_pred, bi_pred, bd_pred, fused_img, re_vi, re_ir = multi_task_model(img_vi, img_ir)# seg_pred = F.softmax(seg_pred, dim=1) # seg_pred = multi_task_model(img_vi, img_ir)optimizer.zero_grad()seg_loss = Seg_loss(seg_pred, label, device, criteria)bd = F.one_hot(bd,num_classes=2)bd=bd.permute(0,3,1,2).float()bi = F.one_hot(bi,num_classes=2)bi= bi.permute(0,3,1,2).float()bd_loss = F.binary_cross_entropy_with_logits(bd_pred, bd) bi_loss = F.binary_cross_entropy_with_logits(bi_pred, bi, pos_weight=binary_class_weight)seg_results = torch.argmax(seg_pred, dim=1, keepdim=True) ## print(seg_result.shape())train_seg_loss = 10 * seg_loss + 5 * bi_loss + 5 * bd_loss## reconstruction-related lossfusion_loss, ssim_loss, grad_loss, int_loss = Fusion_loss(img_ir, vi_Y, fused_img, mask)vi_re_loss, vi_int_loss, vi_grad_loss = Re_loss(re_vi, vi_Y, mask=mask, ir_flag=False)ir_re_loss, ir_int_loss, ir_grad_loss = Re_loss(re_ir, img_ir, mask=mask, ir_flag=True)train_loss = 1 * train_seg_loss + 1 * fusion_loss + 0.5 * vi_re_loss + 0.5 * ir_re_losstrain_loss.backward()optimizer.step()seg_metric.addBatch(seg_results, label, lb_ignore)# dist.destroy_process_group()if opt.local_rank == 0:lr = optimizer.get_lr()mIoU = np.array(seg_metric.meanIntersectionOverUnion().item())Acc = np.array(seg_metric.pixelAccuracy().item())end = time()training_time, glob_t_intv = end - start, end - glob_stnow_it = total_it+1eta = int((total_epoch * len(train_loader) - now_it) * (glob_t_intv / (now_it)))eta = str(datetime.timedelta(seconds=eta))logger.info('ep: [{}/{}], learning rate: {:.6f}, time consuming: {:.2f}s, segmentation loss: {:.4f}, fusion loss: {:.4f}, vi rec loss: {:.4f}, ir rec loss: {:.4f}'.format(ep+1, total_epoch, lr, training_time, seg_loss.item(), fusion_loss.item(), vi_re_loss.item(), ir_re_loss.item()))logger.info('ssim loss: [{:.4f}], grad loss: [{:.4f}], int loss: [{:.4f}], segmentation loss: {:.4f}, mIou: {:.4f}, Acc: {:.4f}, Eta: {}\n'.format(ssim_loss.item(), grad_loss.item(), int_loss.item(), seg_loss.item(), mIoU, Acc, eta))start = time()## save Visualization resultsif (ep + 1) % opt.img_save_freq == 0 and opt.local_rank == 0:input = [img_ir, img_vi, fused_img, label]fused_rgb = YCbCr2RGB(fused_img, vi_Cb, vi_Cr)vi_rgb = YCbCr2RGB(re_vi, vi_Cb, vi_Cr)output = [re_ir, vi_rgb, fused_rgb, seg_results]saver.write_img(ep, input, output)## save modelif (ep + 1) % opt.model_save_freq == 0 and opt.local_rank == 0:test_mIoU = multi_task_tester(test_loader, multi_task_model, device, opt)logger.info('test mIoU: {:.4f}, best mIoU:{:.4f}'.format(test_mIoU, best_mIou))if test_mIoU > best_mIou:best_mIou = test_mIoUsaver.write_model(ep, total_it, multi_task_model, optimizer.optim, best_mIou, device)def multi_task_tester(test_loader, multi_task_model, device, opts):multi_task_model.eval()test_bar= tqdm(test_loader)seg_metric = SegmentationMetric(opts.class_nb, device=device)lb_ignore = [255]## define save dirwith torch.no_grad():# operations inside don't track historyfor it, (img_ir, img_vi, label, img_names) in enumerate(test_bar):img_ir = img_ir.to(device)img_vi = img_vi.to(device)label = label.to(device) Seg_pred, _, _, fused_img, re_vi, re_ir = multi_task_model(img_vi, img_ir)seg_result = torch.argmax(Seg_pred, dim=1, keepdim=True) ## print(seg_result.shape())seg_metric.addBatch(seg_result, label, lb_ignore)mIoU = np.array(seg_metric.meanIntersectionOverUnion().item())return mIoUif __name__ == '__main__':main()

options.py

import argparseclass TrainOptions():def __init__(self):self.parser = argparse.ArgumentParser()# data loader relatedself.parser.add_argument('--dataroot', type=str, default='/data/timer/Idea/mtan/dataset/MSRS', help='path of data')self.parser.add_argument('--phase', type=str, default='train', help='phase for dataloading')self.parser.add_argument('--batch_size', type=int, default=12, help='batch size')self.parser.add_argument('--nThreads', type=int, default=16, help='# of threads for data loader')# training relatedself.parser.add_argument('--lr', default=1e-3, type=int, help='Initial learning rate for training model')self.parser.add_argument('--weight', default='dwa', type=str, help='multi-task weighting: equal, uncert, dwa')self.parser.add_argument('--n_ep', type=int, default=1500, help='number of epochs') # 400 * d_iterself.parser.add_argument('--n_ep_decay', type=int, default=1000, help='epoch start decay learning rate, set -1 if no decay') # 200 * d_iterself.parser.add_argument('--resume', type=str, default=None, help='specified the dir of saved models for resume the training') # 不要改该参数,系统会自动分配self.parser.add_argument('--gpu', type=str, default='cuda', help='device id (i.e. 0 or 0,1 or cpu)')self.parser.add_argument('--temp', default=2.0, type=float, help='temperature for DWA (must be positive)')# ouptput relatedself.parser.add_argument('--name', type=str, default='MPF-Trans-skip_DDP', help='folder name to save outputs')self.parser.add_argument('--class_nb', type=int, default=9, help='class number for segmentation model')self.parser.add_argument('--display_dir', type=str, default='/data/timer/Idea/mtan/logs', help='path for saving display results')self.parser.add_argument('--result_dir', type=str, default='/data/timer/Idea/mtan/results', help='path for saving result images and models')self.parser.add_argument('--display_freq', type=int, default=10, help='freq (iteration) of display')self.parser.add_argument('--img_save_freq', type=int, default=10, help='freq (epoch) of saving images')self.parser.add_argument('--model_save_freq', type=int, default=10, help='freq (epoch) of saving models')# DDP relatedself.parser.add_argument('--local_rank', type=int, default=0, help='Specifying the default GPU')def parse(self):self.opt = self.parser.parse_args()args = vars(self.opt)print('\n--- load options ---')for name, value in sorted(args.items()):print('%s: %s' % (str(name), str(value)))return self.optclass TestOptions():def __init__(self):self.parser = argparse.ArgumentParser()# data loader relatedself.parser.add_argument('--dataroot', type=str, default='/data/timer/Idea/mtan/dataset/MSRS', help='path of data')self.parser.add_argument('--phase', type=str, default='test', help='phase for dataloading')self.parser.add_argument('--batch_size', type=int, default=16, help='batch size')self.parser.add_argument('--nThreads', type=int, default=16, help='# of threads for data loader')## mode relatedself.parser.add_argument('--class_nb', type=int, default=9, help='class number for segmentation model')self.parser.add_argument('--resume', type=str, default='/data/timer/Idea/mtan/results/MPF-skip/best_model.pth', help='specified the dir of saved models for resume the training')self.parser.add_argument('--gpu', type=int, default=0, help='GPU id')# results relatedself.parser.add_argument('--name', type=str, default='MPF_skip', help='folder name to save outputs')self.parser.add_argument('--result_dir', type=str, default='/data/timer/Idea/mtan/test', help='path for saving result images and models')def parse(self):self.opt = self.parser.parse_args()args = vars(self.opt)print('\n--- load options ---')for name, value in sorted(args.items()):print('%s: %s' % (str(name), str(value)))return self.opt

一些主要的操作都在train.py文件里有所涉及,因为是第一次系统的使用DDP,还有很多地方理解的不够透彻,不当之处希望大家指出一起交流。