forked from olehb/pytorch_ddp_tutorial
-
Notifications
You must be signed in to change notification settings - Fork 0
/
ddp_tutorial_multi_gpu.py
144 lines (119 loc) · 4.91 KB
/
ddp_tutorial_multi_gpu.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
import argparse
from typing import Tuple
from tqdm import tqdm
import torch
from torch import nn, optim
from torch.distributed import Backend
from torch.nn.parallel.distributed import DistributedDataParallel
from torch.utils.data import DataLoader, DistributedSampler
from torchvision import datasets, transforms
def create_data_loaders(rank: int,
world_size: int,
batch_size: int) -> Tuple[DataLoader, DataLoader]:
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])
dataset_loc = './mnist_data'
train_dataset = datasets.MNIST(dataset_loc,
download=True,
train=True,
transform=transform)
sampler = DistributedSampler(train_dataset,
num_replicas=world_size, # Number of GPUs
rank=rank, # GPU where process is running
shuffle=True, # Shuffling is done by Sampler
seed=42)
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=False, # This is mandatory to set this to False here, shuffling is done by Sampler
num_workers=4,
sampler=sampler,
pin_memory=True)
# This is not necessary to use distributed sampler for the test or validation sets.
test_dataset = datasets.MNIST(dataset_loc,
download=True,
train=False,
transform=transform)
test_loader = DataLoader(test_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4,
pin_memory=True)
return train_loader, test_loader
def create_model():
# create model architecture
model = nn.Sequential(
nn.Linear(28*28, 128), # MNIST images are 28x28 pixels
nn.ReLU(),
nn.Dropout(0.2),
nn.Linear(128, 128),
nn.ReLU(),
nn.Linear(128, 10, bias=False) # 10 classes to predict
)
return model
def main(rank: int,
epochs: int,
model: nn.Module,
train_loader: DataLoader,
test_loader: DataLoader) -> nn.Module:
device = torch.device(f'cuda:{rank}')
model = model.to(device)
model = DistributedDataParallel(model, device_ids=[rank], output_device=rank)
# initialize optimizer and loss function
optimizer = optim.SGD(model.parameters(), lr=0.01)
loss = nn.CrossEntropyLoss()
# train the model
for i in range(epochs):
model.train()
train_loader.sampler.set_epoch(i)
epoch_loss = 0
# train the model for one epoch
pbar = tqdm(train_loader)
for x, y in pbar:
x = x.to(device, non_blocking=True)
y = y.to(device, non_blocking=True)
x = x.view(x.shape[0], -1)
optimizer.zero_grad()
y_hat = model(x)
batch_loss = loss(y_hat, y)
batch_loss.backward()
optimizer.step()
batch_loss_scalar = batch_loss.item()
epoch_loss += batch_loss_scalar / x.shape[0]
pbar.set_description(f'training batch_loss={batch_loss_scalar:.4f}')
# calculate validation loss
with torch.no_grad():
model.eval()
val_loss = 0
pbar = tqdm(test_loader)
for x, y in pbar:
x = x.to(device, non_blocking=True)
y = y.to(device, non_blocking=True)
x = x.view(x.shape[0], -1)
y_hat = model(x)
batch_loss = loss(y_hat, y)
batch_loss_scalar = batch_loss.item()
val_loss += batch_loss_scalar / x.shape[0]
pbar.set_description(f'validation batch_loss={batch_loss_scalar:.4f}')
print(f"Epoch={i}, train_loss={epoch_loss:.4f}, val_loss={val_loss:.4f}")
return model.module
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--local_rank", type=int)
args = parser.parse_args()
batch_size = 128
epochs = 10
rank = args.local_rank
world_size = torch.cuda.device_count()
torch.cuda.set_device(rank)
torch.distributed.init_process_group(backend=Backend.NCCL,
init_method='env://')
train_loader, test_loader = create_data_loaders(rank, world_size, batch_size)
model = main(rank=rank,
epochs=epochs,
model=create_model(),
train_loader=train_loader,
test_loader=test_loader)
if rank == 0:
torch.save(model.state_dict(), 'model.pt')