File size: 4,716 Bytes
a25d5d4 |
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 |
"""
reference
- https://github.com/pytorch/vision/blob/main/references/detection/utils.py
- https://github.com/facebookresearch/detr/blob/master/util/misc.py#L406
by lyuwenyu
"""
import random
import numpy as np
import torch
import torch.nn as nn
import torch.distributed
import torch.distributed as tdist
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import DistributedSampler
from torch.utils.data.dataloader import DataLoader
def init_distributed():
'''
distributed setup
args:
backend (str), ('nccl', 'gloo')
'''
try:
# # https://pytorch.org/docs/stable/elastic/run.html
# LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))
# RANK = int(os.getenv('RANK', -1))
# WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
tdist.init_process_group(init_method='env://', )
torch.distributed.barrier()
rank = get_rank()
device = torch.device(f'cuda:{rank}')
torch.cuda.set_device(device)
setup_print(rank == 0)
print('Initialized distributed mode...')
return True
except:
print('Not init distributed mode.')
return False
def setup_print(is_main):
'''This function disables printing when not in master process
'''
import builtins as __builtin__
builtin_print = __builtin__.print
def print(*args, **kwargs):
force = kwargs.pop('force', False)
if is_main or force:
builtin_print(*args, **kwargs)
__builtin__.print = print
def is_dist_available_and_initialized():
if not tdist.is_available():
return False
if not tdist.is_initialized():
return False
return True
def get_rank():
if not is_dist_available_and_initialized():
return 0
return tdist.get_rank()
def get_world_size():
if not is_dist_available_and_initialized():
return 1
return tdist.get_world_size()
def is_main_process():
return get_rank() == 0
def save_on_master(*args, **kwargs):
if is_main_process():
torch.save(*args, **kwargs)
def warp_model(model, find_unused_parameters=False, sync_bn=False,):
if is_dist_available_and_initialized():
rank = get_rank()
model = nn.SyncBatchNorm.convert_sync_batchnorm(model) if sync_bn else model
model = DDP(model, device_ids=[rank], output_device=rank, find_unused_parameters=find_unused_parameters)
return model
def warp_loader(loader, shuffle=False):
if is_dist_available_and_initialized():
sampler = DistributedSampler(loader.dataset, shuffle=shuffle)
loader = DataLoader(loader.dataset,
loader.batch_size,
sampler=sampler,
drop_last=loader.drop_last,
collate_fn=loader.collate_fn,
pin_memory=loader.pin_memory,
num_workers=loader.num_workers, )
return loader
def is_parallel(model) -> bool:
# Returns True if model is of type DP or DDP
return type(model) in (torch.nn.parallel.DataParallel, torch.nn.parallel.DistributedDataParallel)
def de_parallel(model) -> nn.Module:
# De-parallelize a model: returns single-GPU model if model is of type DP or DDP
return model.module if is_parallel(model) else model
def reduce_dict(data, avg=True):
'''
Args
data dict: input, {k: v, ...}
avg bool: true
'''
world_size = get_world_size()
if world_size < 2:
return data
with torch.no_grad():
keys, values = [], []
for k in sorted(data.keys()):
keys.append(k)
values.append(data[k])
values = torch.stack(values, dim=0)
tdist.all_reduce(values)
if avg is True:
values /= world_size
_data = {k: v for k, v in zip(keys, values)}
return _data
def all_gather(data):
"""
Run all_gather on arbitrary picklable data (not necessarily tensors)
Args:
data: any picklable object
Returns:
list[data]: list of data gathered from each rank
"""
world_size = get_world_size()
if world_size == 1:
return [data]
data_list = [None] * world_size
tdist.all_gather_object(data_list, data)
return data_list
import time
def sync_time():
'''sync_time
'''
if torch.cuda.is_available():
torch.cuda.synchronize()
return time.time()
def set_seed(seed):
# fix the seed for reproducibility
seed = seed + get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
|