def main():
global best_prec1, args
args.gpu = 0
args.world_size = 1
if args.distributed:
args.gpu = args.local_rank % torch.cuda.device_count()
torch.cuda.set_device(args.gpu)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.total_batch_size = args.world_size * args.batch_size
if not os.path.isdir(args.checkpoint) and args.local_rank == 0:
mkdir_p(args.checkpoint)
if args.fp16:
assert torch.backends.cudnn.enabled, "fp16 mode requires cudnn backend to be enabled."
if args.static_loss_scale != 1.0:
if not args.fp16:
print(
"Warning: if --fp16 is not used, static_loss_scale will be ignored."
)
# create model
if args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
model = models.__dict__[args.arch](pretrained=True)
else:
print("=> creating model '{}'".format(args.arch))
model = Network(backbone=args.arch, num_classes=args.train_class_num)
model = model.cuda()
if args.fp16:
model = network_to_half(model)
if args.distributed:
# shared param/delay all reduce turns off bucketing in DDP, for lower latency runs this can improve perf
# for the older version of APEX please use shared_param, for newer one it is delay_allreduce
model = DDP(model, delay_allreduce=True)
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.fp16:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.static_loss_scale,
dynamic_loss_scale=args.dynamic_loss_scale,
verbose=False)
# optionally resume from a checkpoint
title = 'ImageNet-' + args.arch
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(
args.resume,
map_location=lambda storage, loc: storage.cuda(args.gpu))
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
else:
if args.local_rank == 0:
logger = Logger(os.path.join(args.checkpoint, 'log.txt'),
title=title)
logger.set_names([
'Learning Rate', 'Train Loss', 'Valid Loss', 'Train Acc.',
'Valid Acc.', 'Valid Top5.'
])
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, args.val)
crop_size = 224
val_size = 256
pipe = HybridTrainPipe(batch_size=args.batch_size,
num_threads=args.workers,
device_id=args.local_rank,
data_dir=traindir,
crop=crop_size,
dali_cpu=args.dali_cpu)
pipe.build()
train_loader = DALIClassificationIterator(
pipe, size=int(pipe.epoch_size("Reader") / args.world_size))
pipe = HybridValPipe(batch_size=args.batch_size,
num_threads=args.workers,
device_id=args.local_rank,
data_dir=valdir,
crop=crop_size,
size=val_size)
pipe.build()
val_loader = DALIClassificationIterator(
pipe, size=int(pipe.epoch_size("Reader") / args.world_size))
validate(val_loader, train_loader, model)
def main():
global best_prec1, args
args.gpu = 0
args.world_size = 1
if args.distributed:
args.gpu = args.local_rank % torch.cuda.device_count()
torch.cuda.set_device(args.gpu)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.total_batch_size = args.world_size * args.batch_size
if not os.path.isdir(args.checkpoint) and args.local_rank == 0:
mkdir_p(args.checkpoint)
if args.fp16:
assert torch.backends.cudnn.enabled, "fp16 mode requires cudnn backend to be enabled."
if args.static_loss_scale != 1.0:
if not args.fp16:
print(
"Warning: if --fp16 is not used, static_loss_scale will be ignored."
)
# create model
if args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
model = models.__dict__[args.arch](pretrained=True)
else:
print("=> creating model '{}'".format(args.arch))
model = Network(backbone=args.arch, num_classes=args.train_class_num)
model = model.cuda()
if args.fp16:
model = network_to_half(model)
if args.distributed:
# shared param/delay all reduce turns off bucketing in DDP, for lower latency runs this can improve perf
# for the older version of APEX please use shared_param, for newer one it is delay_allreduce
model = DDP(model, delay_allreduce=True)
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.fp16:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.static_loss_scale,
dynamic_loss_scale=args.dynamic_loss_scale,
verbose=False)
# optionally resume from a checkpoint
title = 'ImageNet-' + args.arch
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(
args.resume,
map_location=lambda storage, loc: storage.cuda(args.gpu))
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
if args.local_rank == 0:
logger = Logger(os.path.join(args.checkpoint, 'log.txt'),
title=title,
resume=True)
else:
print("=> no checkpoint found at '{}'".format(args.resume))
else:
if args.local_rank == 0:
logger = Logger(os.path.join(args.checkpoint, 'log.txt'),
title=title)
logger.set_names([
'Learning Rate', 'Train Loss', 'Valid Loss', 'Train Acc.',
'Valid Acc.', 'Valid Top5.'
])
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val')
if (args.arch == "inception_v3"):
crop_size = 299
val_size = 320 # I chose this value arbitrarily, we can adjust.
else:
crop_size = 224
val_size = 256
pipe = HybridTrainPipe(batch_size=args.batch_size,
num_threads=args.workers,
device_id=args.local_rank,
data_dir=traindir,
crop=crop_size,
dali_cpu=args.dali_cpu)
pipe.build()
train_loader = DALIClassificationIterator(
pipe, size=int(pipe.epoch_size("Reader") / args.world_size))
# pipe = HybridValPipe(batch_size=args.batch_size, num_threads=args.workers, device_id=args.local_rank, data_dir=valdir, crop=crop_size, size=val_size)
# pipe.build()
# val_loader = DALIClassificationIterator(pipe, size=int(pipe.epoch_size("Reader") / args.world_size))
# if args.evaluate:
# validate(val_loader, model, criterion)
# return
total_time = AverageMeter()
for epoch in range(args.start_epoch, args.epochs):
# train for one epoch
adjust_learning_rate(optimizer, epoch, args)
if args.local_rank == 0:
print('\nEpoch: [%d | %d] LR: %f' %
(epoch + 1, args.epochs, optimizer.param_groups[0]['lr']))
[train_loss, train_acc,
avg_train_time] = train(train_loader, model, criterion, optimizer,
epoch)
total_time.update(avg_train_time)
# evaluate on validation set
# [test_loss, prec1, prec5] = validate(val_loader, model, criterion)
# remember best prec@1 and save checkpoint
if args.local_rank == 0:
# append logger file
# logger.append([optimizer.param_groups[0]['lr'], train_loss, test_loss, train_acc, prec1, prec5])
logger.append([
optimizer.param_groups[0]['lr'], train_loss, 0.0, train_acc,
0.0, 0.0
])
# is_best = prec1 > best_prec1
is_best = False
# best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
},
is_best,
checkpoint=args.checkpoint,
filename="checkpoint.pth.tar")
# if epoch == args.epochs - 1:
# print('##Top-1 {0}\n'
# '##Top-5 {1}\n'
# '##Perf {2}'.format(prec1, prec5, args.total_batch_size / total_time.avg))
# reset DALI iterators
train_loader.reset()
# val_loader.reset()
if args.local_rank == 0:
logger.close()