-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
221 lines (193 loc) · 6.49 KB
/
main.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
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
import argparse
import os
import horovod.torch as hvd
import torch
import torchvision
from torch import nn, optim
from torchvision import transforms
import model
parser = argparse.ArgumentParser(description="PyTorch CIFAR10 Training")
parser.add_argument("--lr", default=0.1, type=float, help="learning rate")
parser.add_argument("--dataset", help="dataset root")
parser.add_argument("--checkpoint_in", help="checkpoint in")
parser.add_argument("--checkpoint_out", help="checkpoint out")
parser.add_argument(
"--epochs",
type=int,
default=200,
metavar="N",
help="number of epochs to train (default: 10)",
)
parser.add_argument(
"--horovod",
action="store_true",
default=False,
help="Enable distributed computing using Horovod",
)
parser.add_argument(
"--no-cuda",
action="store_true",
default=False,
help="Disable CUDA for computing",
)
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
if args.horovod:
hvd.init()
if args.cuda and args.horovod:
torch.cuda.set_device(hvd.local_rank())
device = torch.device("cuda" if args.cuda else "cpu")
start_epoch = 0 # start from epoch 0 or last checkpoint epoch
transform_train = transforms.Compose(
[
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
]
)
transform_test = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
]
)
classes = (
"plane",
"car",
"bird",
"cat",
"deer",
"dog",
"frog",
"horse",
"ship",
"truck",
)
trainset = torchvision.datasets.CIFAR10(
root=args.dataset, train=True, download=False, transform=transform_train
)
if args.horovod:
trainsampler = torch.utils.data.distributed.DistributedSampler(
trainset, num_replicas=hvd.size(), rank=hvd.rank()
)
trainloader = torch.utils.data.DataLoader(
trainset, batch_size=128, sampler=trainsampler
)
else:
trainloader = torch.utils.data.DataLoader(
trainset,
batch_size=128,
)
testset = torchvision.datasets.CIFAR10(
root=args.dataset, train=False, download=False, transform=transform_test
)
if args.horovod:
testsampler = torch.utils.data.distributed.DistributedSampler(
testset, num_replicas=hvd.size(), rank=hvd.rank()
)
testloader = torch.utils.data.DataLoader(
testset, batch_size=100, sampler=testsampler
)
else:
testloader = torch.utils.data.DataLoader(
testset,
batch_size=100,
)
# Model
print("==> Building model...")
torch.set_float32_matmul_precision("high")
net = model.SimpleDLA()
BEST_ACC = 0.0
if not args.horovod:
if args.cuda and torch.cuda.device_count() >= 1:
print("Parallelized on", torch.cuda.device_count(), "GPUs!")
net = torch.nn.DataParallel(net)
else:
print("Parallelized on", torch.get_num_threads(), "threads!")
else:
torch.set_num_threads(1)
net = net.to(device)
net = torch.compile(net)
print("Model is JIT-compiling enabled!")
if args.checkpoint_in:
# Load checkpoint.
print("==> Resuming from checkpoint..")
if os.path.isfile(args.checkpoint_in):
checkpoint = torch.load(args.checkpoint_in)
net.load_state_dict(checkpoint["net"])
BEST_ACC = checkpoint["acc"]
start_epoch = checkpoint["epoch"]
else:
print("no checkpoint found")
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
if args.horovod:
optimizer = hvd.DistributedOptimizer(
optimizer, named_parameters=net.named_parameters()
)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=200)
if args.horovod:
# Broadcast parameters from rank 0 to all other processes.
hvd.broadcast_parameters(net.state_dict(), root_rank=0)
# Training
def train(epoch: int):
print(f"\nEpoch: {epoch}")
net.train()
correct = 0
if args.horovod:
trainsampler.set_epoch(epoch)
for batch_idx, (inputs, targets) in enumerate(trainloader):
inputs, targets = inputs.to(device), targets.to(device)
optimizer.zero_grad()
outputs = net(inputs)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
_, predicted = outputs.max(1)
correct += predicted.eq(targets).sum().item()
if args.horovod:
print(
f"Rank: {hvd.rank()} | Train Epoch: {epoch} | Batch {batch_idx}/{len(trainloader)} | Loss: {loss.item():.6f} | Acc: {100.0 * correct / len(trainsampler):.3f}% ({correct}/{len(trainsampler)})",
)
else:
print(
f"Train Epoch: {epoch} | Batch {batch_idx}/{len(trainloader)} | Loss: {loss.item():.6f} | Acc: {100.0 * correct / len(batch_idx + 1):.3f}% ({correct}/{batch_idx + 1})",
)
def test(epoch: int):
global BEST_ACC
net.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(testloader):
inputs, targets = inputs.to(device), targets.to(device)
outputs = net(inputs)
loss = criterion(outputs, targets)
test_loss += loss.item()
_, predicted = outputs.max(1)
correct += predicted.eq(targets).sum().item()
if args.horovod:
print(
f"Rank: {hvd.rank()} | Epoch: {epoch} | Batch {batch_idx}/{len(testsampler)} | Loss: {test_loss / len(testsampler):.3f} | Acc: {100.0 * correct / len(testsampler):.3f}% ({correct}/{len(testsampler)})",
)
else:
print(
f"Epoch: {epoch} | Batch {batch_idx}/{len(testloader)} | Loss: {test_loss / len(batch_idx + 1):.3f} | Acc: {100.0 * correct / len(batch_idx + 1):.3f}% ({correct}/{batch_idx + 1})",
)
# Save checkpoint.
acc = 100.0 * correct / len(testsampler)
if acc > BEST_ACC and args.checkpoint_out:
if not args.horovod or (args.horovod and hvd.rank() == 0):
print("Saving..")
state = {
"net": net.state_dict(),
"acc": acc,
"epoch": epoch,
}
torch.save(state, args.checkpoint_out)
BEST_ACC = acc
for epoch in range(start_epoch, start_epoch + args.epochs):
train(epoch)
test(epoch)
scheduler.step()