pytorch在Horovod上訓練步驟分為以下幾步：
|
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
|
import torchimport horovod.torch as hvd# Initialize Horovod 初始化horovodhvd.init()# Pin GPU to be used to process local rank (one GPU per process) 分配到每個gpu上torch.cuda.set_device(hvd.local_rank())# Define dataset... 定義datasettrain_dataset = ...# Partition dataset among workers using DistributedSampler 對dataset的采樣器進行調整，使用torch.utils.data.distributed.DistributedSamplertrain_sampler = torch.utils.data.distributed.DistributedSampler( train_dataset, num_replicas=hvd.size(), rank=hvd.rank())train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=..., sampler=train_sampler)# Build model...model = ...model.cuda()optimizer = optim.SGD(model.parameters())# Add Horovod Distributed Optimizer 使用Horovod的分布式優化器函數包裹在原先optimizer上optimizer = hvd.DistributedOptimizer(optimizer, named_parameters=model.named_parameters())# Broadcast parameters from rank 0 to all other processes. 參數廣播到每個gpu上hvd.broadcast_parameters(model.state_dict(), root_rank=0)for epoch in range(100): for batch_idx, (data, target) in enumerate(train_loader): optimizer.zero_grad() output = model(data) loss = F.nll_loss(output, target) loss.backward() optimizer.step() if batch_idx % args.log_interval == 0: print('Train Epoch: {} [{}/{}]\tLoss: {}'.format( epoch, batch_idx * len(data), len(train_sampler), loss.item())) |
完整示例代碼如下，在imagenet上采用resnet50進行訓練
|
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
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
|
from __future__ import print_function import torch import argparse import torch.backends.cudnn as cudnn import torch.nn.functional as F import torch.optim as optim import torch.utils.data.distributed from torchvision import datasets, transforms, modelsimport horovod.torch as hvdimport osimport mathfrom tqdm import tqdmfrom distutils.version import LooseVersion# Training settingsparser = argparse.ArgumentParser(description='PyTorch ImageNet Example', formatter_class=argparse.ArgumentDefaultsHelpFormatter)parser.add_argument('--train-dir', default=os.path.expanduser('~/imagenet/train'), help='path to training data')parser.add_argument('--val-dir', default=os.path.expanduser('~/imagenet/validation'), help='path to validation data')parser.add_argument('--log-dir', default='./logs', help='tensorboard log directory')parser.add_argument('--checkpoint-format', default='./checkpoint-{epoch}.pth.tar', help='checkpoint file format')parser.add_argument('--fp-allreduce', action='store_true', default=False, help='use fp compression during allreduce')parser.add_argument('--batches-per-allreduce', type=int, default=, help='number of batches processed locally before ' 'executing allreduce across workers; it multiplies ' 'total batch size.')parser.add_argument('--use-adasum', action='store_true', default=False, help='use adasum algorithm to do reduction')# Default settings from https://arxiv.org/abs/1706.02677.parser.add_argument('--batch-size', type=int, default=32, help='input batch size for training')parser.add_argument('--val-batch-size', type=int, default=32, help='input batch size for validation')parser.add_argument('--epochs', type=int, default=90, help='number of epochs to train')parser.add_argument('--base-lr', type=float, default=0.0125,44 help='learning rate for a single GPU')45 parser.add_argument('--warmup-epochs', type=float, default=5, help='number of warmup epochs')parser.add_argument('--momentum', type=float, default=0.9, help='SGD momentum')parser.add_argument('--wd', type=float, default=0.00005, help='weight decay')parser.add_argument('--no-cuda', action='store_true', default=False, help='disables CUDA training')parser.add_argument('--seed', type=int, default=42, help='random seed')args = parser.parse_args()args.cuda = not args.no_cuda and torch.cuda.is_available()allreduce_batch_size = args.batch_size * args.batches_per_allreducehvd.init()torch.manual_seed(args.seed)if args.cuda: # Horovod: pin GPU to local rank. torch.cuda.set_device(hvd.local_rank()) torch.cuda.manual_seed(args.seed)cudnn.benchmark = True# If set > 0, will resume training from a given checkpoint.resume_from_epoch = 0for try_epoch in range(args.epochs, 0, -1): if os.path.exists(args.checkpoint_format.format(epoch=try_epoch)): resume_from_epoch = try_epoch break# Horovod: broadcast resume_from_epoch from rank 0 (which will have# checkpoints) to other ranks.resume_from_epoch = hvd.broadcast(torch.tensor(resume_from_epoch), root_rank=0, name='resume_from_epoch').item()# Horovod: print logs on the first worker.verbose = 1 if hvd.rank() == 0 else 0# Horovod: write TensorBoard logs on first worker.try: if LooseVersion(torch.__version__) >= LooseVersion('1.2.0'): from torch.utils.tensorboard import SummaryWriter else: from tensorboardX import SummaryWriter log_writer = SummaryWriter(args.log_dir) if hvd.rank() == 0 else Noneexcept ImportError: log_writer = None# Horovod: limit # of CPU threads to be used per worker.torch.set_num_threads(4)kwargs = {'num_workers': 4, 'pin_memory': True} if args.cuda else {}train_dataset = \ datasets.ImageFolder(args.train_dir, transform=transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[., ., .], std=[0.229, 0.224, 0.225]) ]))# Horovod: use DistributedSampler to partition data among workers. Manually specify# `num_replicas=hvd.size()` and `rank=hvd.rank()`.train_sampler = torch.utils.data.distributed.DistributedSampler( train_dataset, num_replicas=hvd.size(), rank=hvd.rank())train_loader = torch.utils.data.DataLoader( train_dataset, batch_size=allreduce_batch_size, sampler=train_sampler, **kwargs)val_dataset = \ datasets.ImageFolder(args.val_dir, transform=transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]))val_sampler = torch.utils.data.distributed.DistributedSampler( val_dataset, num_replicas=hvd.size(), rank=hvd.rank())val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=args.val_batch_size, sampler=val_sampler, **kwargs)# Set up standard ResNet-50 model.model = models.resnet50()# By default, Adasum doesn't need scaling up learning rate.# For sum/average with gradient Accumulation: scale learning rate by batches_per_allreducelr_scaler = args.batches_per_allreduce * hvd.size() if not args.use_adasum else 1if args.cuda: # Move model to GPU. model.cuda() # If using GPU Adasum allreduce, scale learning rate by local_size. if args.use_adasum and hvd.nccl_built(): lr_scaler = args.batches_per_allreduce * hvd.local_size()# Horovod: scale learning rate by the number of GPUs.optimizer = optim.SGD(model.parameters(), lr=(args.base_lr * lr_scaler), momentum=args.momentum, weight_decay=args.wd)# Horovod: (optional) compression algorithm.compression = hvd.Compression.fp16 if args.fp16_allreduce else hvd.Compression.none# Horovod: wrap optimizer with DistributedOptimizer.optimizer = hvd.DistributedOptimizer( optimizer, named_parameters=model.named_parameters(), compression=compression, backward_passes_per_step=args.batches_per_allreduce, op=hvd.Adasum if args.use_adasum else hvd.Average)# Restore from a previous checkpoint, if initial_epoch is specified.# Horovod: restore on the first worker which will broadcast weights to other workers.if resume_from_epoch > 0 and hvd.rank() == 0: filepath = args.checkpoint_format.format(epoch=resume_from_epoch) checkpoint = torch.load(filepath) model.load_state_dict(checkpoint['model']) optimizer.load_state_dict(checkpoint['optimizer'])# Horovod: broadcast parameters & optimizer state.hvd.broadcast_parameters(model.state_dict(), root_rank=)hvd.broadcast_optimizer_state(optimizer, root_rank=)def train(epoch): model.train() train_sampler.set_epoch(epoch) train_loss = Metric('train_loss') train_accuracy = Metric('train_accuracy') with tqdm(total=len(train_loader), desc='Train Epoch #{}'.format(epoch + 1), disable=not verbose) as t: for batch_idx, (data, target) in enumerate(train_loader): adjust_learning_rate(epoch, batch_idx) if args.cuda: data, target = data.cuda(), target.cuda() optimizer.zero_grad() # Split data into sub-batches of size batch_size for i in range(0, len(data), args.batch_size): data_batch = data[i:i + args.batch_size] target_batch = target[i:i + args.batch_size] output = model(data_batch) train_accuracy.update(accuracy(output, target_batch)) loss = F.cross_entropy(output, target_batch) train_loss.update(loss) # Average gradients among sub-batches loss.div_(math.ceil(float(len(data)) / args.batch_size)) loss.backward() # Gradient is applied across all ranks optimizer.step() t.set_postfix({'loss': train_loss.avg.item(), 'accuracy': 100. * train_accuracy.avg.item()}) t.update(1) if log_writer: log_writer.add_scalar('train/loss', train_loss.avg, epoch) log_writer.add_scalar('train/accuracy', train_accuracy.avg, epoch)def validate(epoch): model.eval() val_loss = Metric('val_loss') val_accuracy = Metric('val_accuracy') with tqdm(total=len(val_loader), desc='Validate Epoch #{}'.format(epoch + ), disable=not verbose) as t: with torch.no_grad(): for data, target in val_loader: if args.cuda: data, target = data.cuda(), target.cuda() output = model(data) val_loss.update(F.cross_entropy(output, target)) val_accuracy.update(accuracy(output, target)) t.set_postfix({'loss': val_loss.avg.item(), 'accuracy': 100. * val_accuracy.avg.item()}) t.update(1) if log_writer: log_writer.add_scalar('val/loss', val_loss.avg, epoch) log_writer.add_scalar('val/accuracy', val_accuracy.avg, epoch)# Horovod: using `lr = base_lr * hvd.size()` from the very beginning leads to worse final# accuracy. Scale the learning rate `lr = base_lr` ---> `lr = base_lr * hvd.size()` during# the first five epochs. See https://arxiv.org/abs/1706.02677 for details.# After the warmup reduce learning rate by 10 on the 30th, 60th and 80th epochs.def adjust_learning_rate(epoch, batch_idx): if epoch < args.warmup_epochs: epoch += float(batch_idx + 1) / len(train_loader) lr_adj = 1. / hvd.size() * (epoch * (hvd.size() - 1) / args.warmup_epochs + 1) elif epoch < 30: lr_adj = 1. elif epoch < 60: lr_adj = 1e-1 elif epoch < 80: lr_adj = 1e-2 else: lr_adj = 1e-3 for param_group in optimizer.param_groups: param_group['lr'] = args.base_lr * hvd.size() * args.batches_per_allreduce * lr_adjdef accuracy(output, target): # get the index of the max log-probability pred = output.max(1, keepdim=True)[1] return pred.eq(target.view_as(pred)).cpu().float().mean()def save_checkpoint(epoch): if hvd.rank() == 0: filepath = args.checkpoint_format.format(epoch=epoch + 1) state = { 'model': model.state_dict(), 'optimizer': optimizer.state_dict(), } torch.save(state, filepath)# Horovod: average metrics from distributed training.class Metric(object): def __init__(self, name): self.name = name self.sum = torch.tensor(0.) self.n = torch.tensor(0.) def update(self, val): self.sum += hvd.allreduce(val.detach().cpu(), name=self.name) self.n += 1 @property def avg(self): return self.sum / self.nfor epoch in range(resume_from_epoch, args.epochs): train(epoch) validate(epoch) save_checkpoint(epoch) |
到此這篇關于pytorch使用horovod多gpu訓練的實現的文章就介紹到這了,更多相關pytorch horovod多gpu訓練內容請搜索服務器之家以前的文章或繼續瀏覽下面的相關文章希望大家以后多多支持服務器之家！
原文鏈接：https://www.cnblogs.com/ywheunji/p/12298518.html
