mnist_distributed.py

"""Usage
pip install -r requirements.txt

# run locally
python mnist.py

# run remotely
# set your AWS_DEFAULT_REGION/AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY
python mnist.py --remote
"""

from __future__ import print_function

import argparse
import os

import torch
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 20, 5, 1)
        self.conv2 = nn.Conv2d(20, 50, 5, 1)
        self.fc1 = nn.Linear(4*4*50, 500)
        self.fc2 = nn.Linear(500, 10)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        x = F.max_pool2d(x, 2, 2)
        x = F.relu(self.conv2(x))
        x = F.max_pool2d(x, 2, 2)
        x = x.view(-1, 4*4*50)
        x = F.relu(self.fc1(x))
        x = self.fc2(x)
        return F.log_softmax(x, dim=1)


def train(model, device, train_loader, optimizer, epoch):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()
        output = model(data)
        loss = F.nll_loss(output, target)
        loss.backward()
        optimizer.step()
        if batch_idx % args.log_interval == 0 and args.local_rank == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, batch_idx * len(data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.item()))


def test(model, device, test_loader):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            output = model(data)
            test_loss += F.nll_loss(output, target, reduction='sum').item()  # sum up batch loss
            pred = output.argmax(dim=1, keepdim=True)  # get the index of the max log-probability
            correct += pred.eq(target.view_as(pred)).sum().item()

    test_loss /= len(test_loader.dataset)

    if args.local_rank == 0:
        print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
            test_loss, correct, len(test_loader.dataset),
            100. * correct / len(test_loader.dataset)))


def worker(ddp=True):
    """
    Main training script.

    Args:
        ddp: True if this script runs as part of DistributedDataParallel ensemble
    """
    use_cuda = not args.no_cuda and torch.cuda.is_available()

    torch.manual_seed(args.seed)

    device = torch.device("cuda" if use_cuda else "cpu")

    world_size = int(os.environ.get('WORLD_SIZE', 1))
    rank = int(os.environ.get('RANK', 0))

    if ddp:
        dist.init_process_group(backend='gloo', init_method='env://')
     
    kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}

    # only download dataset once per machine, sync workers
    if args.local_rank == 0:
        datasets.MNIST('/tmp/data', download=True)
        print(f"DDP: process {rank}/{world_size}")
    if ddp:
        dist.barrier()
        
    train_loader = torch.utils.data.DataLoader(
        datasets.MNIST('/tmp/data', train=True,
                       transform=transforms.Compose([
                           transforms.ToTensor(),
                           transforms.Normalize((0.1307,), (0.3081,))
                       ])),
        batch_size=args.batch_size, shuffle=True, **kwargs)
    test_loader = torch.utils.data.DataLoader(
        datasets.MNIST('/tmp/data', train=False, transform=transforms.Compose([
                           transforms.ToTensor(),
                           transforms.Normalize((0.1307,), (0.3081,))
                       ])),
        batch_size=args.test_batch_size, shuffle=True, **kwargs)

    model = Net().to(device)
    if ddp:
        model = nn.parallel.DistributedDataParallel(model)
    else:
        model = nn.DataParallel(model, dim=1)

    optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)

    for epoch in range(1, args.epochs + 1):
        train(model, device, train_loader, optimizer, epoch)
        test(model, device, test_loader)

    if args.save_model:
        torch.save(model.state_dict(), "mnist_cnn.pt")


def format_args(dict_: dict):
    """Helper to format dict into "--key1=val2 --key2=val2" string"""
    def item_to_arg(item: tuple):
        k, v = item
        if v is False or v is None:
            return ''
        if v is True:
            return f'--{k}'
        return f'--{k} {v}'

    return ' '.join([item_to_arg(item) for item in dict_.items()])+' '


def remote_launcher():
    import ncluster

    job = ncluster.make_job(name='mnist_distributed',
                            image_name='Deep Learning AMI (Ubuntu) Version 23.0',
                            instance_type='c5.xlarge',
                            num_tasks=args.nnodes)

    job.upload('mnist_distributed.py')
    job.run('rm -Rf /tmp/data')
    job.run('source activate pytorch_p36')
    task0 = job.tasks[0]
    for i, task in enumerate(job.tasks):
        launcher_args = {'nproc_per_node': args.proc_per_node,
                         'nnodes': args.nnodes,
                         'node_rank': i,
                         'master_addr': task0.ip,
                         'master_port': 6016}
        task.run((f'python -m torch.distributed.launch {format_args(launcher_args)} '
                  f'mnist_distributed.py --mode=worker'), non_blocking=True, stream_output=(i == 0))


def local_launcher():
    os.system(f'python -m torch.distributed.launch --nproc_per_node={args.proc_per_node} mnist_distributed.py --mode=worker ')


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
    parser.add_argument('--batch-size', type=int, default=64, metavar='N',
                        help='input batch size for training (default: 64)')
    parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
                        help='input batch size for testing (default: 1000)')
    parser.add_argument('--epochs', type=int, default=10, metavar='N',
                        help='number of epochs to train (default: 10)')
    parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
                        help='learning rate (default: 0.01)')
    parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
                        help='SGD momentum (default: 0.5)')
    parser.add_argument('--no-cuda', action='store_true', default=False,
                        help='disables CUDA training')
    parser.add_argument('--seed', type=int, default=1, metavar='S',
                        help='random seed (default: 1)')
    parser.add_argument('--log-interval', type=int, default=10, metavar='N',
                        help='how many batches to wait before logging training status')
    parser.add_argument('--save-model', action='store_true', default=False,
                        help='For Saving the current Model')

    parser.add_argument('--remote', action='store_true', default=False,
                        help='run training remotely')

    parser.add_argument('--proc_per_node', default=2, help='number of processes per machine')
    parser.add_argument('--nnodes', default=2, type=int, help='number of nodes (machines)')

    parser.add_argument('--mode', default='localworker', choices=['remote', 'local', 'worker', 'localworker'], help="local: spawn multiple processes locally, remote: launch multiple machines/processes on AWS, worker: DDP aware single process process version, localworker: standalone single process version")

    # worker args
    parser.add_argument('--local_rank', default=0, type=int)

    args = parser.parse_args()
    if args.mode == 'remote':
        remote_launcher()
    elif args.mode == 'local':
        local_launcher()
    elif args.mode == 'worker':
        worker()
    elif args.mode == 'localworker':
        worker(ddp=False)