6

我正在使用以下代码。我在具有以下配置的 GPU 机器上运行它：

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Sep__1_21:08:03_CDT_2017
Cuda compilation tools, release 9.0, V9.0.176

我正在使用 pytorch 版本 1.0.0.dev20181123。

我面临这个错误：RuntimeError: DataLoader worker (pid 23646) iskilled by signal: Illegal instruction。

认为这是一个共享内存问题，我在这里尝试了解决方案：https ://www.lucidarme.me/increase-shared-memory-limit/

它没有帮助。任何指针？

import argparse
import json
import os
import torch

#=====START: ADDED FOR DISTRIBUTED======
from distributed import init_distributed, apply_gradient_allreduce, reduce_tensor
from torch.utils.data.distributed import DistributedSampler
#=====END:   ADDED FOR DISTRIBUTED======

from torch.utils.data import DataLoader
from glow import WaveGlow, WaveGlowLoss
from mel2samp import Mel2Samp

def load_checkpoint(checkpoint_path, model, optimizer):
    assert os.path.isfile(checkpoint_path)
    checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
    iteration = checkpoint_dict['iteration']
    optimizer.load_state_dict(checkpoint_dict['optimizer'])
    model_for_loading = checkpoint_dict['model']
    model.load_state_dict(model_for_loading.state_dict())
    print("Loaded checkpoint '{}' (iteration {})" .format(
          checkpoint_path, iteration))
    return model, optimizer, iteration

def save_checkpoint(model, optimizer, learning_rate, iteration, filepath):
    print("Saving model and optimizer state at iteration {} to {}".format(
          iteration, filepath))
    model_for_saving = WaveGlow(**waveglow_config).cuda()
    model_for_saving.load_state_dict(model.state_dict())
    torch.save({'model': model_for_saving,
                'iteration': iteration,
                'optimizer': optimizer.state_dict(),
                'learning_rate': learning_rate}, filepath)

def train(num_gpus, rank, group_name, output_directory, epochs, learning_rate,
          sigma, iters_per_checkpoint, batch_size, seed, checkpoint_path):
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    #=====START: ADDED FOR DISTRIBUTED======
    if num_gpus > 1:
        init_distributed(rank, num_gpus, group_name, **dist_config)
    #=====END:   ADDED FOR DISTRIBUTED======

    criterion = WaveGlowLoss(sigma)
    model = WaveGlow(**waveglow_config).cuda()

    #=====START: ADDED FOR DISTRIBUTED======
    if num_gpus > 1:
        model = apply_gradient_allreduce(model)
    #=====END:   ADDED FOR DISTRIBUTED======

    optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

    # Load checkpoint if one exists
    iteration = 0
    if checkpoint_path != "":
        model, optimizer, iteration = load_checkpoint(checkpoint_path, model,
                                                      optimizer)
        iteration += 1  # next iteration is iteration + 1

    trainset = Mel2Samp(**data_config)
    # =====START: ADDED FOR DISTRIBUTED======
    train_sampler = DistributedSampler(trainset) if num_gpus > 1 else None
    # =====END:   ADDED FOR DISTRIBUTED======
    train_loader = DataLoader(trainset, num_workers=1, shuffle=False,
                              sampler=train_sampler,
                              batch_size=batch_size,
                              pin_memory=False,
                              drop_last=True)

    # Get shared output_directory ready
    if rank == 0:
        if not os.path.isdir(output_directory):
            os.makedirs(output_directory)
            os.chmod(output_directory, 0o775)
        print("output directory", output_directory)

    model.train()
    epoch_offset = max(0, int(iteration / len(train_loader)))
    # ================ MAIN TRAINNIG LOOP! ===================
    for epoch in range(epoch_offset, epochs):
        print("Epoch: {}".format(epoch))
        for i, batch in enumerate(train_loader):
            model.zero_grad()

            mel, audio = batch
            mel = torch.autograd.Variable(mel.cuda())
            audio = torch.autograd.Variable(audio.cuda())
            outputs = model((mel, audio))

            loss = criterion(outputs)
            if num_gpus > 1:
                reduced_loss = reduce_tensor(loss.data, num_gpus).item()
            else:
                reduced_loss = loss.item()
            loss.backward()
            optimizer.step()

            print("{}:\t{:.9f}".format(iteration, reduced_loss))

            if (iteration % iters_per_checkpoint == 0):
                if rank == 0:
                    checkpoint_path = "{}/waveglow_{}".format(
                        output_directory, iteration)
                    save_checkpoint(model, optimizer, learning_rate, iteration,
                                    checkpoint_path)

            iteration += 1

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('-c', '--config', type=str,
                        help='JSON file for configuration')
    parser.add_argument('-r', '--rank', type=int, default=0,
                        help='rank of process for distributed')
    parser.add_argument('-g', '--group_name', type=str, default='',
                        help='name of group for distributed')
    args = parser.parse_args()

    # Parse configs.  Globals nicer in this case
    with open(args.config) as f:
        data = f.read()
    config = json.loads(data)
    train_config = config["train_config"]
    global data_config
    data_config = config["data_config"]
    global dist_config
    dist_config = config["dist_config"]
    global waveglow_config
    waveglow_config = config["waveglow_config"]

    num_gpus = torch.cuda.device_count()
    if num_gpus > 1:
        if args.group_name == '':
            print("WARNING: Multiple GPUs detected but no distributed group set")
            print("Only running 1 GPU.  Use distributed.py for multiple GPUs")
            num_gpus = 1

    if num_gpus == 1 and args.rank != 0:
        raise Exception("Doing single GPU training on rank > 0")

    torch.backends.cudnn.enabled = True
    torch.backends.cudnn.benchmark = False
    train(num_gpus, args.rank, args.group_name, **train_config)
4

0 回答 0