mrshenli Sorry for the late reply. Say I want to train the DDP model on 4 gpus and restore it as DDP on 2. I created an mnist example to illustrate my case while following your example. This whole script is borrowed from mnist, modified and split into three scripts:
- mnist_common.py
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
import argparse
from torchvision import datasets, transforms
from torch.utils.data.distributed import DistributedSampler
from torch.utils.data import DataLoader
def cleanup():
dist.destroy_process_group()
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(1, 32, 3, 1)
self.conv2 = nn.Conv2d(32, 64, 3, 1)
self.dropout1 = nn.Dropout2d(0.25)
self.dropout2 = nn.Dropout2d(0.5)
self.fc1 = nn.Linear(9216, 128)
self.fc2 = nn.Linear(128, 10)
def forward(self, x):
x = self.conv1(x)
x = F.relu(x)
x = self.conv2(x)
x = F.relu(x)
x = F.max_pool2d(x, 2)
x = self.dropout1(x)
x = torch.flatten(x, 1)
x = self.fc1(x)
x = F.relu(x)
x = self.dropout2(x)
x = self.fc2(x)
output = F.log_softmax(x, dim=1)
return output
def train(args, model, device, train_loader, optimizer, epoch):
model.train()
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device, non_blocking=True), \
target.to(device, non_blocking=True)
optimizer.zero_grad()
output = model(data)
loss = F.nll_loss(output, target)
loss.backward()
optimizer.step()
def test(args, 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, non_blocking=True), \
target.to(device, non_blocking=True)
output = model(data)
test_loss += F.nll_loss(
output,
target,
reduction='sum').item()
pred = output.argmax(dim=1, keepdim=True)
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
if args.local_rank == 0:
print('Test set: Average loss: {:.4f},'
' Accuracy: {}/{} ({:.2f}%)'.format(
test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size',
type=int,
default=64,
metavar='N',
help='input batch size for training')
parser.add_argument('--test-batch-size',
type=int,
default=1000,
metavar='N',
help='input batch size for testing')
parser.add_argument('--epochs', type=int, default=14, metavar='N',
help='number of epochs to train (default: 14)')
parser.add_argument('--lr', type=float, default=1.0, metavar='LR',
help='learning rate (default: 1.0)')
parser.add_argument('--gamma', type=float, default=0.7, metavar='M',
help='Learning rate step gamma (default: 0.7)')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--local_rank', type=int)
args = parser.parse_args()
train_dataset = datasets.MNIST(
'../data',
train=True,
download=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
]))
train_sampler = DistributedSampler(
train_dataset,
num_replicas=torch.cuda.device_count(),
rank=args.local_rank)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=0,
pin_memory=True,
sampler=train_sampler)
test_loader = DataLoader(
datasets.MNIST(
'../data',
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.test_batch_size,
shuffle=True,
num_workers=0,
pin_memory=True,)
- mnist_train.py
from __future__ import print_function
import torch
import torch.optim as optim
import torch.distributed as dist
import torch.backends.cudnn as cudnn
from torch.optim.lr_scheduler import StepLR
from torch.nn.parallel import DistributedDataParallel as DDP
from mnist_common import args, Net, train_loader, train_sampler,\
test_loader, train, test, cleanup
def main(args):
dist.init_process_group(backend='nccl',
init_method='tcp://localhost:23456',
rank=args.local_rank,
world_size=torch.cuda.device_count())
torch.manual_seed(args.seed)
torch.cuda.set_device(args.local_rank)
cudnn.benchmark = True
model = Net()
model = model.to([args.local_rank][0]) # distribute the model
# Should we set the output_device value in DPP?
model = DDP(model, device_ids=[args.local_rank])
# , output_device=[args.local_rank][0])
optimizer = optim.Adadelta(model.parameters(), lr=args.lr)
scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
for epoch in range(1, args.epochs + 1):
train_sampler.set_epoch(epoch)
train(args, model, args.local_rank,
train_loader, optimizer, epoch)
test(args, model, args.local_rank, test_loader)
scheduler.step(epoch)
# I intend to save the model
# AFTER some training not, not before
if args.local_rank == 0:
torch.save(model, "mnist_cnn.pt")
dist.barrier()
cleanup()
if __name__ == '__main__':
main(args)
Also I intend to test the model, only after training (sometimes up to a few days) has finished, by restoring the saved weights (or model).
2) mnist_test.py
from __future__ import print_function
import torch
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
from mnist_common import args, Net, test_loader, test, cleanup
def main(args):
dist.init_process_group(backend='nccl',
init_method='tcp://localhost:23456',
rank=args.local_rank,
world_size=2)
torch.manual_seed(args.seed)
torch.cuda.set_device(args.local_rank)
model = torch.load("mnist_cnn.pt",
map_location=torch.device(args.local_rank))
model = DDP(model, device_ids=[args.local_rank])
print(f"Rank {args.local_rank} "
f"test on device {list(model.parameters())[0].device}")
test(args, model, args.local_rank, test_loader)
cleanup()
if __name__ == '__main__':
main(args)
The mnist_train.py runs sucessfully using
python -m torch.distributed.launch nproc_per_node=4 (or 2) mnist_train.py.
but when i run the test script using
python -m torch.distributed.launch nproc_per_node=2 mnist_test.py.
I get the following:
Rank 0 test on device cuda:0
Rank 1 test on device cuda:1
Test set: Average loss: 0.0274, Accuracy: 9913/10000 (99.13%)
RuntimeError: Expected tensor for argument #1 input to have
the same device as tensor for argument #2 weight;
but device 0 does not equal 1
(while checking arguments for cudnn_convolution)