def main():
setup_default_logging()
args, args_text = _parse_args()
args.pretrained_backbone = not args.no_pretrained_backbone
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info('Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on 1 GPU.')
use_amp = None
if args.amp:
# for backwards compat, `--amp` arg tries apex before native amp
if has_apex:
args.apex_amp = True
elif has_native_amp:
args.native_amp = True
else:
logging.warning("Neither APEX or native Torch AMP is available, using float32. "
"Install NVIDA apex or upgrade to PyTorch 1.6.")
if args.apex_amp:
if has_apex:
use_amp = 'apex'
else:
logging.warning("APEX AMP not available, using float32. Install NVIDA apex")
elif args.native_amp:
if has_native_amp:
use_amp = 'native'
else:
logging.warning("Native AMP not available, using float32. Upgrade to PyTorch 1.6.")
torch.manual_seed(args.seed + args.rank)
model = create_model(
args.model,
bench_task='train',
num_classes=args.num_classes,
pretrained=args.pretrained,
pretrained_backbone=args.pretrained_backbone,
redundant_bias=args.redundant_bias,
label_smoothing=args.smoothing,
new_focal=args.new_focal,
jit_loss=args.jit_loss,
bench_labeler=args.bench_labeler,
checkpoint_path=args.initial_checkpoint,
)
model_config = model.config # grab before we obscure with DP/DDP wrappers
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' % (args.model, sum([m.numel() for m in model.parameters()])))
model.cuda()
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
optimizer = create_optimizer(args, model)
amp_autocast = suppress # do nothing
loss_scaler = None
if use_amp == 'apex':
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
loss_scaler = ApexScaler()
if args.local_rank == 0:
logging.info('Using NVIDIA APEX AMP. Training in mixed precision.')
elif use_amp == 'native':
amp_autocast = torch.cuda.amp.autocast
loss_scaler = NativeScaler()
if args.local_rank == 0:
logging.info('Using native Torch AMP. Training in mixed precision.')
else:
if args.local_rank == 0:
logging.info('AMP not enabled. Training in float32.')
# optionally resume from a checkpoint
resume_epoch = None
if args.resume:
resume_epoch = resume_checkpoint(
unwrap_bench(model), args.resume,
optimizer=None if args.no_resume_opt else optimizer,
loss_scaler=None if args.no_resume_opt else loss_scaler,
log_info=args.local_rank == 0)
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(model, decay=args.model_ema_decay)
if args.resume:
# FIXME bit of a mess with bench, cannot use the load in ModelEma
load_checkpoint(unwrap_bench(model_ema), args.resume, use_ema=True)
if args.distributed:
if args.sync_bn:
try:
if has_apex and use_amp != 'native':
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.')
except Exception as e:
logging.error('Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1')
if has_apex and use_amp != 'native':
if args.local_rank == 0:
logging.info("Using apex DistributedDataParallel.")
model = ApexDDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info("Using torch DistributedDataParallel.")
model = NativeDDP(model, device_ids=[args.device])
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
loader_train, loader_eval, evaluator = create_datasets_and_loaders(args, model_config)
if model_config.num_classes < loader_train.dataset.parser.max_label:
logging.error(
f'Model {model_config.num_classes} has fewer classes than dataset {loader_train.dataset.parser.max_label}.')
exit(1)
if model_config.num_classes > loader_train.dataset.parser.max_label:
logging.warning(
f'Model {model_config.num_classes} has more classes than dataset {loader_train.dataset.parser.max_label}.')
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
args.model
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(
model, optimizer, args=args, model_ema=model_ema, amp_scaler=loss_scaler,
checkpoint_dir=output_dir, decreasing=decreasing, unwrap_fn=unwrap_bench)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(
epoch, model, loader_train, optimizer, args,
lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir,
amp_autocast=amp_autocast, loss_scaler=loss_scaler, model_ema=model_ema)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
logging.info("Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
# the overhead of evaluating with coco style datasets is fairly high, so just ema or non, not both
if model_ema is not None:
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
distribute_bn(model_ema, args.world_size, args.dist_bn == 'reduce')
eval_metrics = validate(model_ema.ema, loader_eval, args, evaluator, log_suffix=' (EMA)')
else:
eval_metrics = validate(model, loader_eval, args, evaluator)
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
if saver is not None:
update_summary(
epoch, train_metrics, eval_metrics, os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
# save proper checkpoint with eval metric
best_metric, best_epoch = saver.save_checkpoint(epoch=epoch, metric=eval_metrics[eval_metric])
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(best_metric, best_epoch))
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
_logger.warning(
'Using more than one GPU per process in distributed mode is not allowed.Setting num_gpu to 1.'
)
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
args.world_size = int(os.environ['WORLD_SIZE'])
args.rank = int(os.environ['RANK'])
torch.distributed.init_process_group(backend='nccl',
init_method=args.init_method,
rank=args.rank,
world_size=args.world_size)
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
_logger.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
_logger.info('Training with a single process on %d GPUs.' %
args.num_gpu)
torch.manual_seed(args.seed + args.rank)
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect, # DEPRECATED, use drop_path
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint)
################## pretrain ############
if args.pretrain_path is not None:
print('Loading:', args.pretrain_path)
state_dict = torch.load(args.pretrain_path)
model.load_state_dict(state_dict, strict=False)
print('Pretrain weights loaded.')
################### flops #################
print(model)
if hasattr(model, 'default_cfg'):
default_cfg = model.default_cfg
input_size = [1] + list(default_cfg['input_size'])
else:
input_size = [1, 3, 224, 224]
input = torch.randn(input_size) #.cuda()
from torchprofile import profile_macs
macs = profile_macs(model, input)
print('model flops:', macs, 'input_size:', input_size)
##########################################
if args.local_rank == 0:
_logger.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model.parameters()])))
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
use_amp = None
if args.amp:
# for backwards compat, `--amp` arg tries apex before native amp
if has_apex:
args.apex_amp = True
elif has_native_amp:
args.native_amp = True
if args.apex_amp and has_apex:
use_amp = 'apex'
elif args.native_amp and has_native_amp:
use_amp = 'native'
elif args.apex_amp or args.native_amp:
_logger.warning(
"Neither APEX or native Torch AMP is available, using float32. "
"Install NVIDA apex or upgrade to PyTorch 1.6")
if args.num_gpu > 1:
if use_amp == 'apex':
_logger.warning(
'Apex AMP does not work well with nn.DataParallel, disabling. Use DDP or Torch AMP.'
)
use_amp = None
model = nn.DataParallel(model,
device_ids=list(range(args.num_gpu))).cuda()
assert not args.channels_last, "Channels last not supported with DP, use DDP."
else:
model.cuda()
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
optimizer = create_optimizer(args, model)
amp_autocast = suppress # do nothing
loss_scaler = None
if use_amp == 'apex':
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
loss_scaler = ApexScaler()
if args.local_rank == 0:
_logger.info('Using NVIDIA APEX AMP. Training in mixed precision.')
elif use_amp == 'native':
amp_autocast = torch.cuda.amp.autocast
loss_scaler = NativeScaler()
if args.local_rank == 0:
_logger.info(
'Using native Torch AMP. Training in mixed precision.')
else:
if args.local_rank == 0:
_logger.info('AMP not enabled. Training in float32.')
# optionally resume from a checkpoint
resume_epoch = None
if args.resume:
resume_epoch = resume_checkpoint(
model,
args.resume,
optimizer=None if args.no_resume_opt else optimizer,
loss_scaler=None if args.no_resume_opt else loss_scaler,
log_info=args.local_rank == 0)
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume=args.resume)
if args.distributed:
if args.sync_bn:
assert not args.split_bn
try:
if has_apex and use_amp != 'native':
# Apex SyncBN preferred unless native amp is activated
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(
model)
if args.local_rank == 0:
_logger.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.'
)
except Exception as e:
_logger.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1'
)
if has_apex and use_amp != 'native':
# Apex DDP preferred unless native amp is activated
if args.local_rank == 0:
_logger.info("Using NVIDIA APEX DistributedDataParallel.")
model = ApexDDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
_logger.info("Using native Torch DistributedDataParallel.")
model = NativeDDP(model, device_ids=[
args.local_rank
]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
_logger.info('Scheduled epochs: {}'.format(num_epochs))
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
_logger.error(
'Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_args = dict(mixup_alpha=args.mixup,
cutmix_alpha=args.cutmix,
cutmix_minmax=args.cutmix_minmax,
prob=args.mixup_prob,
switch_prob=args.mixup_switch_prob,
mode=args.mixup_mode,
label_smoothing=args.smoothing,
num_classes=args.num_classes)
if args.prefetcher:
assert not num_aug_splits # collate conflict (need to support deinterleaving in collate mixup)
collate_fn = FastCollateMixup(**mixup_args)
else:
mixup_fn = Mixup(**mixup_args)
if num_aug_splits > 1:
dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
train_interpolation = args.train_interpolation
if args.no_aug or not train_interpolation:
train_interpolation = data_config['interpolation']
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
no_aug=args.no_aug,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
scale=args.scale,
ratio=args.ratio,
hflip=args.hflip,
vflip=args.vflip,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
use_multi_epochs_loader=args.use_multi_epochs_loader,
repeated_aug=args.repeated_aug)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
_logger.error(
'Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
if args.jsd:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits,
smoothing=args.smoothing).cuda()
elif mixup_active:
# smoothing is handled with mixup target transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
if args.evaluate:
eval_metrics = validate(model,
loader_eval,
validate_loss_fn,
args,
amp_autocast=amp_autocast)
print(eval_metrics)
return
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"), args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(model=model,
optimizer=optimizer,
args=args,
model_ema=model_ema,
amp_scaler=loss_scaler,
checkpoint_dir=output_dir,
recovery_dir=output_dir,
decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
amp_autocast=amp_autocast,
loss_scaler=loss_scaler,
model_ema=model_ema,
mixup_fn=mixup_fn)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
_logger.info(
"Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
eval_metrics = validate(model,
loader_eval,
validate_loss_fn,
args,
amp_autocast=amp_autocast)
if model_ema is not None and not args.model_ema_force_cpu:
if args.distributed and args.dist_bn in ('broadcast',
'reduce'):
distribute_bn(model_ema, args.world_size,
args.dist_bn == 'reduce')
ema_eval_metrics = validate(model_ema.ema,
loader_eval,
validate_loss_fn,
args,
amp_autocast=amp_autocast,
log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
epoch, metric=save_metric)
except KeyboardInterrupt:
pass
if best_metric is not None:
_logger.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
def main():
import os
args, args_text = _parse_args()
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = 'train'
if args.gate_train:
exp_name += '-dynamic'
if args.slim_train:
exp_name += '-slimmable'
exp_name += '-{}'.format(args.model)
exp_info = '-'.join(
[datetime.now().strftime("%Y%m%d-%H%M%S"), args.model])
output_dir = get_outdir(output_base, exp_name, exp_info)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir,
decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
setup_default_logging(outdir=output_dir, local_rank=args.local_rank)
torch.backends.cudnn.benchmark = True
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
logging.warning(
'Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.'
)
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
# torch.distributed.init_process_group(backend='nccl',
# init_method='tcp://127.0.0.1:23334',
# rank=args.local_rank,
# world_size=int(os.environ['WORLD_SIZE']))
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on %d GPUs.' %
args.num_gpu)
# --------- random seed -----------
random.seed(args.seed) # TODO: do we need same seed on all GPU?
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# torch.manual_seed(args.seed + args.rank)
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_path_rate=args.drop_path,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint)
# optionally resume from a checkpoint
resume_state = {}
resume_epoch = None
if args.resume:
resume_state, resume_epoch = resume_checkpoint(model, args.resume)
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model.parameters()])))
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
if args.num_gpu > 1:
if args.amp:
logging.warning(
'AMP does not work well with nn.DataParallel, disabling. Use distributed mode for multi-GPU AMP.'
)
args.amp = False
model = nn.DataParallel(model,
device_ids=list(range(args.num_gpu))).cuda()
else:
model.cuda()
if args.train_mode == 'se':
optimizer = create_optimizer(args, model.get_se())
elif args.train_mode == 'bn':
optimizer = create_optimizer(args, model.get_bn())
elif args.train_mode == 'all':
optimizer = create_optimizer(args, model)
elif args.train_mode == 'gate':
optimizer = create_optimizer(args, model.get_gate())
use_amp = False
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed',
'on' if use_amp else 'off'))
if resume_state and not args.no_resume_opt:
# ----------- Load Optimizer ---------
if 'optimizer' in resume_state:
if args.local_rank == 0:
logging.info('Restoring Optimizer state from checkpoint')
optimizer.load_state_dict(resume_state['optimizer'])
if use_amp and 'amp' in resume_state and 'load_state_dict' in amp.__dict__:
if args.local_rank == 0:
logging.info('Restoring NVIDIA AMP state from checkpoint')
amp.load_state_dict(resume_state['amp'])
del resume_state
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume=args.resume)
if args.distributed:
if args.sync_bn:
assert not args.split_bn
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(
model)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.'
)
except Exception as e:
logging.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1'
)
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
model = DDP(model,
device_ids=[args.local_rank],
find_unused_parameters=True
) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
# ------------- data --------------
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
logging.error(
'Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
if num_aug_splits > 1:
dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=args.train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
)
loader_bn = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=args.train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
logging.error(
'Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
# ------------- loss_fn --------------
if args.jsd:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits,
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
if args.ieb:
distill_loss_fn = SoftTargetCrossEntropy().cuda()
else:
distill_loss_fn = None
if args.local_rank == 0:
model_profiling(model, 224, 224, 1, 3, use_cuda=True, verbose=True)
else:
model_profiling(model, 224, 224, 1, 3, use_cuda=True, verbose=False)
if not args.test_mode:
# start training
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = OrderedDict([('loss', 0.)])
# train
if args.gate_train:
train_metrics = train_epoch_slim_gate(
epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema,
optimizer_step=args.optimizer_step)
else:
train_metrics = train_epoch_slim(
epoch,
model,
loader_train,
optimizer,
loss_fn=train_loss_fn,
distill_loss_fn=distill_loss_fn,
args=args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema,
optimizer_step=args.optimizer_step,
)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
logging.info(
"Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
# eval
if args.gate_train:
eval_sample_list = ['dynamic']
else:
if epoch % 10 == 0 and epoch != 0:
eval_sample_list = ['smallest', 'largest', 'uniform']
else:
eval_sample_list = ['smallest', 'largest']
eval_metrics = [
validate_slim(model,
loader_eval,
validate_loss_fn,
args,
model_mode=model_mode)
for model_mode in eval_sample_list
]
if model_ema is not None and not args.model_ema_force_cpu:
ema_eval_metrics = [
validate_slim(model_ema.ema,
loader_eval,
validate_loss_fn,
args,
model_mode=model_mode)
for model_mode in eval_sample_list
]
eval_metrics = ema_eval_metrics
if isinstance(eval_metrics, list):
eval_metrics = eval_metrics[0]
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
# save
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
model,
optimizer,
args,
epoch=epoch,
model_ema=model_ema,
metric=save_metric,
use_amp=use_amp)
# end training
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
# test
eval_metrics = []
for choice in range(args.num_choice):
# reset bn if not smallest or largest
if choice != 0 and choice != args.num_choice - 1:
for layer in model.modules():
if isinstance(layer, nn.BatchNorm2d) or \
isinstance(layer, nn.SyncBatchNorm) or \
(has_apex and isinstance(layer, apex.parallel.SyncBatchNorm)):
layer.reset_running_stats()
model.train()
with torch.no_grad():
for batch_idx, (input, target) in enumerate(loader_bn):
if args.slim_train:
if hasattr(model, 'module'):
model.module.set_mode('uniform', choice=choice)
else:
model.set_mode('uniform', choice=choice)
model(input)
if batch_idx % 1000 == 0 and batch_idx != 0:
print('Subnet {} : reset bn for {} steps'.format(
choice, batch_idx))
break
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
logging.info(
"Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
eval_metrics.append(
validate_slim(model,
loader_eval,
validate_loss_fn,
args,
model_mode=choice))
if args.local_rank == 0:
print('Test results of the last epoch:\n', eval_metrics)
def main(args):
utils.init_distributed_mode(args)
# disable any harsh augmentation in case of Self-supervise training
if args.training_mode == 'SSL':
print("NOTE: Smoothing, Mixup, CutMix, and AutoAugment will be disabled in case of Self-supervise training")
args.smoothing = args.reprob = args.reprob = args.recount = args.mixup = args.cutmix = 0.0
args.aa = ''
if args.SiT_LinearEvaluation == 1:
print("Warning: Linear Evaluation should be set to 0 during SSL training - changing SiT_LinearEvaluation to 0")
args.SiT_LinearEvaluation = 0
utils.print_args(args)
device = torch.device(args.device)
seed = args.seed + utils.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
cudnn.benchmark = True
print("Loading dataset ....")
dataset_train, args.nb_classes = build_dataset(is_train=True, args=args)
dataset_val, _ = build_dataset(is_train=False, args=args)
num_tasks = utils.get_world_size()
global_rank = utils.get_rank()
if args.repeated_aug:
sampler_train = RASampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True)
else:
sampler_train = torch.utils.data.DistributedSampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
data_loader_train = torch.utils.data.DataLoader(dataset_train, sampler=sampler_train,
batch_size=args.batch_size, num_workers=args.num_workers,
pin_memory=args.pin_mem, drop_last=True, collate_fn=collate_fn)
data_loader_val = torch.utils.data.DataLoader(dataset_val, sampler=sampler_val,
batch_size=int(1.5 * args.batch_size), num_workers=args.num_workers,
pin_memory=args.pin_mem, drop_last=False, collate_fn=collate_fn)
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_fn = Mixup(
mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax,
prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode,
label_smoothing=args.smoothing, num_classes=args.nb_classes)
print(f"Creating model: {args.model}")
model = create_model(
args.model, pretrained=False, num_classes=args.nb_classes,
drop_rate=args.drop, drop_path_rate=args.drop_path, representation_size=args.representation_size,
drop_block_rate=None, training_mode=args.training_mode)
if args.finetune:
checkpoint = torch.load(args.finetune, map_location='cpu')
checkpoint_model = checkpoint['model']
state_dict = model.state_dict()
for k in ['rot_head.weight', 'rot_head.bias', 'contrastive_head.weight', 'contrastive_head.bias']:
if k in checkpoint_model and checkpoint_model[k].shape != state_dict[k].shape:
print(f"Removing key {k} from pretrained checkpoint")
del checkpoint_model[k]
# interpolate position embedding
pos_embed_checkpoint = checkpoint_model['pos_embed']
embedding_size = pos_embed_checkpoint.shape[-1]
num_patches = model.patch_embed.num_patches
num_extra_tokens = model.pos_embed.shape[-2] - num_patches
orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
new_size = int(num_patches ** 0.5)
extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
pos_tokens = torch.nn.functional.interpolate(
pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
checkpoint_model['pos_embed'] = new_pos_embed
model.load_state_dict(checkpoint_model, strict=False)
model.to(device)
# Freeze the backbone in case of linear evaluation
if args.SiT_LinearEvaluation == 1:
requires_grad(model, False)
model.rot_head.weight.requires_grad = True
model.rot_head.bias.requires_grad = True
model.contrastive_head.weight.requires_grad = True
model.contrastive_head.bias.requires_grad = True
if args.representation_size is not None:
model.pre_logits_rot.fc.weight.requires_grad = True
model.pre_logits_rot.fc.bias.requires_grad = True
model.pre_logits_contrastive.fc.weight.requires_grad = True
model.pre_logits_contrastive.fc.bias.requires_grad = True
model_ema = None
if args.model_ema:
model_ema = ModelEma(model, decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '', resume='')
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)
print('number of params:', n_parameters)
linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size() / 512.0
args.lr = linear_scaled_lr
optimizer = create_optimizer(args, model_without_ddp)
loss_scaler = NativeScaler()
lr_scheduler, _ = create_scheduler(args, optimizer)
if args.training_mode == 'SSL':
criterion = MTL_loss(args.device, args.batch_size)
elif args.training_mode == 'finetune' and args.mixup > 0.:
criterion = SoftTargetCrossEntropy()
else:
criterion = torch.nn.CrossEntropyLoss()
output_dir = Path(args.output_dir)
if args.resume:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.model_ema:
utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema'])
if 'scaler' in checkpoint:
loss_scaler.load_state_dict(checkpoint['scaler'])
if args.eval:
test_stats = evaluate_SSL(data_loader_val, model, device)
print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
return
print(f"Start training for {args.epochs} epochs")
start_time = time.time()
max_accuracy = 0.0
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
data_loader_train.sampler.set_epoch(epoch)
if args.training_mode == 'SSL':
train_stats = train_SSL(
model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler,
args.clip_grad, model_ema, mixup_fn)
else:
train_stats = train_finetune(
model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler,
args.clip_grad, model_ema, mixup_fn)
lr_scheduler.step(epoch)
if epoch%args.validate_every == 0:
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
for checkpoint_path in checkpoint_paths:
utils.save_on_master({
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
'model_ema': get_state_dict(model_ema),
'scaler': loss_scaler.state_dict(),
'args': args,
}, checkpoint_path)
if args.training_mode == 'SSL':
test_stats = evaluate_SSL(data_loader_val, model, device, epoch, args.output_dir)
else:
test_stats = evaluate_finetune(data_loader_val, model, device)
print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
max_accuracy = max(max_accuracy, test_stats["acc1"])
print(f'Max accuracy: {max_accuracy:.2f}%')
log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
**{f'test_{k}': v for k, v in test_stats.items()},
'epoch': epoch,
'n_parameters': n_parameters}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def main():
args, args_text = _parse_args()
# dist init
torch.distributed.init_process_group(backend='nccl', init_method='env://')
torch.cuda.set_device(args.local_rank)
args.world_size = torch.distributed.get_world_size()
args.local_rank = torch.distributed.get_rank()
args.save = args.save + args.exp_name
# detectron2 data loader ###########################
# det2_args = default_argument_parser().parse_args()
det2_args = args
det2_args.config_file = args.det2_cfg
cfg = setup(det2_args)
mapper = DatasetMapper(cfg, augmentations=build_sem_seg_train_aug(cfg))
det2_dataset = iter(build_detection_train_loader(cfg, mapper=mapper))
det2_val = build_batch_test_loader(cfg, cfg.DATASETS.TEST[0])
len_det2_train = 20210 // cfg.SOLVER.IMS_PER_BATCH
if args.local_rank == 0:
create_exp_dir(args.save,
scripts_to_save=glob.glob('*.py') + glob.glob('*.sh'))
logger = SummaryWriter(args.save)
log_format = '%(asctime)s %(message)s'
logging.basicConfig(stream=sys.stdout,
level=logging.INFO,
format=log_format,
datefmt='%m/%d %I:%M:%S %p')
fh = logging.FileHandler(os.path.join(args.save, 'log.txt'))
fh.setFormatter(logging.Formatter(log_format))
logging.getLogger().addHandler(fh)
logging.info("args = %s", str(args))
else:
logger = None
# preparation ################
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# config network and criterion ################
gt_down_sampling = 1
min_kept = int(args.batch_size * args.image_height * args.image_width //
(16 * gt_down_sampling**2))
ohem_criterion = ProbOhemCrossEntropy2d(ignore_label=255,
thresh=0.7,
min_kept=min_kept,
use_weight=False)
# data loader ###########################
num_classes = args.num_classes
with open(args.json_file, 'r') as f:
# dict_a = json.loads(f, cls=NpEncoder)
model_dict = json.loads(f.read())
width_mult_list = [
4. / 12,
6. / 12,
8. / 12,
10. / 12,
1.,
]
model = Network(Fch=args.Fch,
num_classes=num_classes,
stem_head_width=(args.stem_head_width,
args.stem_head_width))
last = model_dict["lasts"]
if args.local_rank == 0:
with torch.cuda.device(0):
macs, params = get_model_complexity_info(
model, (3, args.eval_height, args.eval_width),
as_strings=True,
print_per_layer_stat=True,
verbose=True)
logging.info('{:<30} {:<8}'.format('Computational complexity: ',
macs))
logging.info('{:<30} {:<8}'.format('Number of parameters: ',
params))
with open(os.path.join(args.save, 'args.yaml'), 'w') as f:
f.write(args_text)
init_weight(model,
nn.init.kaiming_normal_,
torch.nn.BatchNorm2d,
args.bn_eps,
args.bn_momentum,
mode='fan_in',
nonlinearity='relu')
if args.pretrain:
model.backbone = load_pretrain(model.backbone, args.pretrain)
model = model.cuda()
# if args.sync_bn:
# if has_apex:
# model = apex.parallel.convert_syncbn_model(model)
# else:
# model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
# Optimizer ###################################
base_lr = args.base_lr
if args.opt == "sgd":
optimizer = torch.optim.SGD(model.parameters(),
lr=base_lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
elif args.opt == "adam":
optimizer = torch.optim.Adam(model.parameters(),
lr=base_lr,
betas=(0.9, 0.999),
eps=1e-08)
elif args.opt == "adamw":
optimizer = torch.optim.AdamW(model.parameters(),
lr=base_lr,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=args.weight_decay)
else:
optimizer = create_optimizer(args, model)
if args.sched == "raw":
lr_scheduler = None
else:
max_iteration = args.epochs * len_det2_train
lr_scheduler = Iter_LR_Scheduler(args, max_iteration, len_det2_train)
start_epoch = 0
if os.path.exists(os.path.join(args.save, 'last.pth.tar')):
args.resume = os.path.join(args.save, 'last.pth.tar')
if args.resume:
model_state_file = args.resume
if os.path.isfile(model_state_file):
checkpoint = torch.load(model_state_file,
map_location=torch.device('cpu'))
start_epoch = checkpoint['start_epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
logging.info('Loaded checkpoint (starting from iter {})'.format(
checkpoint['start_epoch']))
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume=None)
if model_ema:
eval_model = model_ema.ema
else:
eval_model = model
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
model = DDP(model, device_ids=[args.local_rank])
best_valid_iou = 0.
best_epoch = 0
temp_iou = 0.
avg_loss = -1
logging.info("rank: {} world_size: {}".format(args.local_rank,
args.world_size))
for epoch in range(start_epoch, args.epochs):
if args.local_rank == 0:
logging.info(args.load_path)
logging.info(args.save)
logging.info("lr: " + str(optimizer.param_groups[0]['lr']))
# training
drop_prob = args.drop_path_prob * epoch / args.epochs
# model.module.drop_path_prob(drop_prob)
train_mIoU = train(len_det2_train, det2_dataset, model, model_ema,
ohem_criterion, num_classes, lr_scheduler,
optimizer, logger, epoch, args, cfg)
# torch.cuda.empty_cache()
# if epoch > args.epochs // 3:
if epoch >= 0:
temp_iou, avg_loss = validation(det2_val, eval_model,
ohem_criterion, num_classes, args,
cfg)
torch.cuda.empty_cache()
if args.local_rank == 0:
logging.info("Epoch: {} train miou: {:.2f}".format(
epoch + 1, 100 * train_mIoU))
if temp_iou > best_valid_iou:
best_valid_iou = temp_iou
best_epoch = epoch
if model_ema is not None:
torch.save(
{
'start_epoch': epoch + 1,
'state_dict': model_ema.ema.state_dict(),
'optimizer': optimizer.state_dict(),
# 'lr_scheduler': lr_scheduler.state_dict(),
},
os.path.join(args.save, 'best_checkpoint.pth.tar'))
else:
torch.save(
{
'start_epoch': epoch + 1,
'state_dict': model.module.state_dict(),
'optimizer': optimizer.state_dict(),
# 'lr_scheduler': lr_scheduler.state_dict(),
},
os.path.join(args.save, 'best_checkpoint.pth.tar'))
logger.add_scalar("mIoU/val", temp_iou, epoch)
logging.info("[Epoch %d/%d] valid mIoU %.4f eval loss %.4f" %
(epoch + 1, args.epochs, temp_iou, avg_loss))
logging.info("Best valid mIoU %.4f Epoch %d" %
(best_valid_iou, best_epoch))
if model_ema is not None:
torch.save(
{
'start_epoch': epoch + 1,
'state_dict': model_ema.ema.state_dict(),
'optimizer': optimizer.state_dict(),
# 'lr_scheduler': lr_scheduler.state_dict(),
},
os.path.join(args.save, 'last.pth.tar'))
else:
torch.save(
{
'start_epoch': epoch + 1,
'state_dict': model.module.state_dict(),
'optimizer': optimizer.state_dict(),
# 'lr_scheduler': lr_scheduler.state_dict(),
},
os.path.join(args.save, 'last.pth.tar'))
def main(args):
utils.init_distributed_mode(args)
update_config_from_file(args.cfg)
print(args)
args_text = yaml.safe_dump(args.__dict__, default_flow_style=False)
device = torch.device(args.device)
# fix the seed for reproducibility
seed = args.seed + utils.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
# random.seed(seed)
cudnn.benchmark = True
dataset_train, args.nb_classes = build_dataset(is_train=True, args=args)
dataset_val, _ = build_dataset(is_train=False, args=args)
if args.distributed:
num_tasks = utils.get_world_size()
global_rank = utils.get_rank()
if args.repeated_aug:
sampler_train = RASampler(dataset_train,
num_replicas=num_tasks,
rank=global_rank,
shuffle=True)
else:
sampler_train = torch.utils.data.DistributedSampler(
dataset_train,
num_replicas=num_tasks,
rank=global_rank,
shuffle=True)
if args.dist_eval:
if len(dataset_val) % num_tasks != 0:
print(
'Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. '
'This will slightly alter validation results as extra duplicate entries are added to achieve '
'equal num of samples per-process.')
sampler_val = torch.utils.data.DistributedSampler(
dataset_val,
num_replicas=num_tasks,
rank=global_rank,
shuffle=False)
else:
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
else:
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
sampler_train = torch.utils.data.RandomSampler(dataset_train)
data_loader_train = torch.utils.data.DataLoader(
dataset_train,
sampler=sampler_train,
batch_size=args.batch_size,
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=True,
)
data_loader_val = torch.utils.data.DataLoader(dataset_val,
batch_size=int(
2 * args.batch_size),
sampler=sampler_val,
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=False)
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_fn = Mixup(mixup_alpha=args.mixup,
cutmix_alpha=args.cutmix,
cutmix_minmax=args.cutmix_minmax,
prob=args.mixup_prob,
switch_prob=args.mixup_switch_prob,
mode=args.mixup_mode,
label_smoothing=args.smoothing,
num_classes=args.nb_classes)
print(f"Creating SuperVisionTransformer")
print(cfg)
model = Vision_TransformerSuper(
img_size=args.input_size,
patch_size=args.patch_size,
embed_dim=cfg.SUPERNET.EMBED_DIM,
depth=cfg.SUPERNET.DEPTH,
num_heads=cfg.SUPERNET.NUM_HEADS,
mlp_ratio=cfg.SUPERNET.MLP_RATIO,
qkv_bias=True,
drop_rate=args.drop,
drop_path_rate=args.drop_path,
gp=args.gp,
num_classes=args.nb_classes,
max_relative_position=args.max_relative_position,
relative_position=args.relative_position,
change_qkv=args.change_qkv,
abs_pos=not args.no_abs_pos)
choices = {
'num_heads': cfg.SEARCH_SPACE.NUM_HEADS,
'mlp_ratio': cfg.SEARCH_SPACE.MLP_RATIO,
'embed_dim': cfg.SEARCH_SPACE.EMBED_DIM,
'depth': cfg.SEARCH_SPACE.DEPTH
}
model.to(device)
if args.teacher_model:
teacher_model = create_model(
args.teacher_model,
pretrained=True,
num_classes=args.nb_classes,
)
teacher_model.to(device)
teacher_loss = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
else:
teacher_model = None
teacher_loss = None
model_ema = None
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu], find_unused_parameters=True)
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size(
) / 512.0
args.lr = linear_scaled_lr
optimizer = create_optimizer(args, model_without_ddp)
loss_scaler = NativeScaler()
lr_scheduler, _ = create_scheduler(args, optimizer)
# criterion = LabelSmoothingCrossEntropy()
if args.mixup > 0.:
# smoothing is handled with mixup label transform
criterion = SoftTargetCrossEntropy()
elif args.smoothing:
criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
else:
criterion = torch.nn.CrossEntropyLoss()
output_dir = Path(args.output_dir)
if not output_dir.exists():
output_dir.mkdir(parents=True)
# save config for later experiments
with open(output_dir / "config.yaml", 'w') as f:
f.write(args_text)
if args.resume:
if args.resume.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.resume,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if 'scaler' in checkpoint:
loss_scaler.load_state_dict(checkpoint['scaler'])
if args.model_ema:
utils._load_checkpoint_for_ema(model_ema,
checkpoint['model_ema'])
retrain_config = None
if args.mode == 'retrain' and "RETRAIN" in cfg:
retrain_config = {
'layer_num': cfg.RETRAIN.DEPTH,
'embed_dim': [cfg.RETRAIN.EMBED_DIM] * cfg.RETRAIN.DEPTH,
'num_heads': cfg.RETRAIN.NUM_HEADS,
'mlp_ratio': cfg.RETRAIN.MLP_RATIO
}
if args.eval:
test_stats = evaluate(data_loader_val,
model,
device,
mode=args.mode,
retrain_config=retrain_config)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
return
print("Start training")
start_time = time.time()
max_accuracy = 0.0
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
data_loader_train.sampler.set_epoch(epoch)
train_stats = train_one_epoch(
model,
criterion,
data_loader_train,
optimizer,
device,
epoch,
loss_scaler,
args.clip_grad,
model_ema,
mixup_fn,
amp=args.amp,
teacher_model=teacher_model,
teach_loss=teacher_loss,
choices=choices,
mode=args.mode,
retrain_config=retrain_config,
)
lr_scheduler.step(epoch)
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
for checkpoint_path in checkpoint_paths:
utils.save_on_master(
{
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
# 'model_ema': get_state_dict(model_ema),
'scaler': loss_scaler.state_dict(),
'args': args,
},
checkpoint_path)
test_stats = evaluate(data_loader_val,
model,
device,
amp=args.amp,
choices=choices,
mode=args.mode,
retrain_config=retrain_config)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
max_accuracy = max(max_accuracy, test_stats["acc1"])
print(f'Max accuracy: {max_accuracy:.2f}%')
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()},
**{f'test_{k}': v
for k, v in test_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
logging.warning('Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.')
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info('Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on %d GPUs.' % args.num_gpu)
torch.manual_seed(args.seed + args.rank)
# my model
use_aux = args.aux_weight > 0.
if args.model_method == 'darts_NAS':
if args.genotype is None:
args.genotype = get_model.get_model(args.model_method, args.model_name)
model = AugmentCNN_ImageNet(224, 3, args.init_channels, args.num_classes, args.layers,
use_aux, args.genotype)
elif args.model_method == 'my_model_collection':
from models.my_searched_model import my_specialized
_ = args.model_name.split(':')
net_config_path = os.path.join(project_path, 'models', 'my_model_collection',
_[0], _[1] + '.json')
model = my_specialized(num_classes=args.num_classes, net_config=net_config_path,
dropout_rate=args.drop)
else:
model_fun = get_model.get_model(args.model_method, args.model_name)
model = model_fun(num_classes=args.num_classes, dropout_rate=args.drop)
# set bn
model.set_bn_param(args.bn_momentum, args.bn_eps)
# model init
model.init_model(model_init=args.model_init)
# pdb.set_trace()
# model = create_model(
# args.model,
# pretrained=args.pretrained,
# num_classes=args.num_classes,
# drop_rate=args.drop,
# drop_connect_rate=args.drop_connect,
# global_pool=args.gp,
# bn_tf=args.bn_tf,
# bn_momentum=args.bn_momentum,
# bn_eps=args.bn_eps,
# checkpoint_path=args.initial_checkpoint)
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel() for m in model.parameters()])))
data_config = resolve_data_config(vars(args), model=model, verbose=args.local_rank == 0)
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
if args.num_gpu > 1:
if args.amp:
logging.warning(
'AMP does not work well with nn.DataParallel, disabling. Use distributed mode for multi-GPU AMP.')
args.amp = False
model = nn.DataParallel(model, device_ids=list(range(args.num_gpu))).cuda()
else:
model.cuda()
optimizer = create_optimizer(args, model)
use_amp = False
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed', 'on' if use_amp else 'off'))
# optionally resume from a checkpoint
resume_state = {}
resume_epoch = None
if args.resume:
resume_state, resume_epoch = resume_checkpoint(model, args.resume)
if resume_state and not args.no_resume_opt:
if 'optimizer' in resume_state:
if args.local_rank == 0:
logging.info('Restoring Optimizer state from checkpoint')
optimizer.load_state_dict(resume_state['optimizer'])
if use_amp and 'amp' in resume_state and 'load_state_dict' in amp.__dict__:
if args.local_rank == 0:
logging.info('Restoring NVIDIA AMP state from checkpoint')
amp.load_state_dict(resume_state['amp'])
del resume_state
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(
model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume=args.resume)
if args.distributed:
if args.sync_bn:
assert not args.split_bn
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.')
except Exception as e:
logging.error('Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1')
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info("Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP.")
model = DDP(model, device_ids=[args.local_rank]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
logging.error('Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
if args.prefetcher and args.mixup > 0:
assert not num_aug_splits # collate conflict (need to support deinterleaving in collate mixup)
collate_fn = FastCollateMixup(args.mixup, args.smoothing, args.num_classes)
if num_aug_splits > 1:
dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=args.train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
logging.error('Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
if args.jsd:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits, smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.mixup > 0.:
# smoothing is handled with mixup label transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
args.model_method,
args.model_name,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir, decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(
epoch, model, loader_train, optimizer, train_loss_fn, args,
lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir,
use_amp=use_amp, model_ema=model_ema)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
logging.info("Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
if model_ema is not None and not args.model_ema_force_cpu:
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
distribute_bn(model_ema, args.world_size, args.dist_bn == 'reduce')
ema_eval_metrics = validate(
model_ema.ema, loader_eval, validate_loss_fn, args, log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
update_summary(
epoch, train_metrics, eval_metrics, os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
model, optimizer, args,
epoch=epoch, model_ema=model_ema, metric=save_metric, use_amp=use_amp)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(best_metric, best_epoch))
def main():
global args
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
logging.warning(
'Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.'
)
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on %d GPUs.' %
args.num_gpu)
torch.manual_seed(args.seed + args.rank)
np.random.seed(args.seed + args.rank)
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint)
if args.binarizable:
Model_binary_patch(model)
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model.parameters()])))
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
if args.num_gpu > 1:
if args.amp:
logging.warning(
'AMP does not work well with nn.DataParallel, disabling. Use distributed mode for multi-GPU AMP.'
)
args.amp = False
model = nn.DataParallel(model,
device_ids=list(range(args.num_gpu))).cuda()
else:
model.cuda()
optimizer = create_optimizer(args, model)
use_amp = False
if has_apex and args.amp:
print('Using amp with --opt-level {}.'.format(args.opt_level))
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.opt_level)
use_amp = True
else:
print('Do NOT use amp.')
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed',
'on' if use_amp else 'off'))
# optionally resume from a checkpoint
resume_state = {}
resume_epoch = None
if args.resume:
resume_state, resume_epoch = resume_checkpoint(model, args.resume)
if resume_state and not args.no_resume_opt:
if 'optimizer' in resume_state:
if args.local_rank == 0:
logging.info('Restoring Optimizer state from checkpoint')
optimizer.load_state_dict(resume_state['optimizer'])
if use_amp and 'amp' in resume_state and 'load_state_dict' in amp.__dict__:
if args.local_rank == 0:
logging.info('Restoring NVIDIA AMP state from checkpoint')
amp.load_state_dict(resume_state['amp'])
resume_state = None
if args.freeze_binary:
Model_freeze_binary(model)
if args.distributed:
if args.sync_bn:
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(
model)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm.')
except Exception as e:
logging.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1'
)
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
model = DDP(model,
device_ids=[args.local_rank
]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
print(num_epochs)
# start_epoch = 0 #
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if args.reset_lr_scheduler is not None:
lr_scheduler.base_values = len(
lr_scheduler.base_values) * [args.reset_lr_scheduler]
lr_scheduler.step(start_epoch)
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
# Using pruner to get sparse weights
if args.prune:
pruner = Pruner_mixed(model, 0, 100, args.pruner)
else:
pruner = None
dataset_train = torchvision.datasets.CIFAR100(root='~/Downloads/CIFAR100',
train=True,
download=True)
collate_fn = None
if args.prefetcher and args.mixup > 0:
collate_fn = FastCollateMixup(args.mixup, args.smoothing,
args.num_classes)
loader_train = create_loader_CIFAR100(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
rand_erase_prob=args.reprob,
rand_erase_mode=args.remode,
rand_erase_count=args.recount,
color_jitter=args.color_jitter,
auto_augment=args.aa,
interpolation='random',
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
is_clean_data=args.clean_train,
)
dataset_eval = torchvision.datasets.CIFAR100(root='~/Downloads/CIFAR100',
train=False,
download=True)
loader_eval = create_loader_CIFAR100(
dataset_eval,
input_size=data_config['input_size'],
batch_size=4 * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
)
if args.mixup > 0.:
# smoothing is handled with mixup label transform
train_loss_fn = SoftTargetCrossEntropy(
multiplier=args.softmax_multiplier).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
saver_last_10_epochs = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"), args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
os.makedirs(output_dir + '/Top')
os.makedirs(output_dir + '/Last')
saver = CheckpointSaver(
checkpoint_dir=output_dir + '/Top',
decreasing=decreasing,
max_history=10) # Save the results of the top 10 epochs
saver_last_10_epochs = CheckpointSaver(
checkpoint_dir=output_dir + '/Last',
decreasing=decreasing,
max_history=10) # Save the results of the last 10 epochs
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
f.write('==============================\n')
f.write(model.__str__())
# if pruner:
# f.write('\n Sparsity \n')
# #f.write(pruner.threshold_dict.__str__())
# f.write('\n pruner.start_epoch={}, pruner.end_epoch={}'.format(pruner.start_epoch, pruner.end_epoch))
tensorboard_writer = SummaryWriter(output_dir)
try:
for epoch in range(start_epoch, num_epochs):
global alpha
alpha = get_alpha(epoch, args)
if args.distributed:
loader_train.sampler.set_epoch(epoch)
if pruner:
pruner.on_epoch_begin(epoch) # pruning
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
tensorboard_writer=tensorboard_writer,
pruner=pruner)
if pruner:
pruner.print_statistics()
eval_metrics = validate(model,
loader_eval,
validate_loss_fn,
args,
tensorboard_writer=tensorboard_writer,
epoch=epoch)
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
model,
optimizer,
args,
epoch=epoch,
metric=save_metric,
use_amp=use_amp)
if saver_last_10_epochs is not None:
# save the checkpoint in the last 5 epochs
_, _ = saver_last_10_epochs.save_checkpoint(model,
optimizer,
args,
epoch=epoch,
metric=epoch,
use_amp=use_amp)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
depth=20,
heads=8,
mlp_dim=2048,
dropout=0.1,
emb_dropout=0.1)
}
BACKBONE = BACKBONE_DICT[BACKBONE_NAME]
print("=" * 60)
print(BACKBONE)
print("{} Backbone Generated".format(BACKBONE_NAME))
print("=" * 60)
LOSS = nn.CrossEntropyLoss()
#embed()
OPTIMIZER = create_optimizer(args, BACKBONE)
print("=" * 60)
print(OPTIMIZER)
print("Optimizer Generated")
print("=" * 60)
lr_scheduler, _ = create_scheduler(args, OPTIMIZER)
# optionally resume from a checkpoint
if BACKBONE_RESUME_ROOT:
print("=" * 60)
print(BACKBONE_RESUME_ROOT)
if os.path.isfile(BACKBONE_RESUME_ROOT):
print("Loading Backbone Checkpoint '{}'".format(
BACKBONE_RESUME_ROOT))
BACKBONE.load_state_dict(torch.load(BACKBONE_RESUME_ROOT))
else:
def main(args):
utils.init_distributed_mode(args)
update_config_from_file(args.cfg)
print(args)
args_text = yaml.safe_dump(args.__dict__, default_flow_style=False)
device = torch.device(args.device)
# fix the seed for reproducibility
seed = args.seed + utils.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
# random.seed(seed)
cudnn.benchmark = True
dataset_train, args.nb_classes = build_dataset(is_train=True, args=args)
dataset_val, _ = build_dataset(is_train=False, args=args)
if args.distributed:
num_tasks = utils.get_world_size()
global_rank = utils.get_rank()
if args.dist_eval:
if len(dataset_val) % num_tasks != 0:
print(
'Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. '
'This will slightly alter validation results as extra duplicate entries are added to achieve '
'equal num of samples per-process.')
sampler_val = torch.utils.data.DistributedSampler(
dataset_val,
num_replicas=num_tasks,
rank=global_rank,
shuffle=False)
else:
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
else:
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
data_loader_val = torch.utils.data.DataLoader(dataset_val,
batch_size=int(
2 * args.batch_size),
sampler=sampler_val,
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=False)
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_fn = Mixup(mixup_alpha=args.mixup,
cutmix_alpha=args.cutmix,
cutmix_minmax=args.cutmix_minmax,
prob=args.mixup_prob,
switch_prob=args.mixup_switch_prob,
mode=args.mixup_mode,
label_smoothing=args.smoothing,
num_classes=args.nb_classes)
print(f"Creating S3-Transformer")
model = SSSTransformer(img_size=args.input_size,
patch_size=args.patch_size,
num_classes=args.nb_classes,
embed_dim=cfg.EMBED_DIM,
depths=cfg.DEPTHS,
num_heads=cfg.NUM_HEADS,
window_size=cfg.WINDOW_SIZE,
mlp_ratio=cfg.MLP_RATIO,
qkv_bias=True,
drop_rate=args.drop,
drop_path_rate=args.drop_path,
patch_norm=True)
model.to(device)
model_ema = None
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu], find_unused_parameters=True)
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size(
) / 512.0
args.lr = linear_scaled_lr
optimizer = create_optimizer(args, model_without_ddp)
loss_scaler = NativeScaler()
lr_scheduler, _ = create_scheduler(args, optimizer)
output_dir = Path(args.output_dir)
if not output_dir.exists():
output_dir.mkdir(parents=True)
# save config for later experiments
with open(output_dir / "config.yaml", 'w') as f:
f.write(args_text)
if args.resume:
if args.resume.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.resume,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
if args.eval:
test_stats = evaluate(data_loader_val, model, device)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
return
def __init__(self, context: PyTorchTrialContext) -> None:
self.context = context
self.hparam = self.context.get_hparam
self.args = DotDict(self.context.get_hparams())
# Create a unique download directory for each rank so they don't overwrite each other.
self.download_directory = f"/tmp/data-rank{self.context.distributed.get_rank()}"
self.num_slots = int(self.context.get_experiment_config()['resources']
['slots_per_trial'])
if self.args.sync_bn and self.num_slots == 1:
print(
'Can not use sync_bn with one slot. Either set sync_bn to False or use distributed training.'
)
sys.exit()
self.args.pretrained_backbone = not self.args.no_pretrained_backbone
self.args.prefetcher = not self.args.no_prefetcher
tmp = []
for arg in self.args.lr_noise.split(' '):
tmp.append(float(arg))
self.args.lr_noise = tmp
self.model = create_model(
self.args.model,
bench_task='train',
num_classes=self.args.num_classes,
pretrained=self.args.pretrained,
pretrained_backbone=self.args.pretrained_backbone,
redundant_bias=self.args.redundant_bias,
label_smoothing=self.args.smoothing,
new_focal=self.args.new_focal,
jit_loss=self.args.jit_loss,
bench_labeler=self.args.bench_labeler,
checkpoint_path=self.args.initial_checkpoint,
)
self.model_config = self.model.config
self.input_config = resolve_input_config(
self.args, model_config=self.model_config)
print('h: ', self.args.model,
sum([m.numel() for m in self.model.parameters()]))
if self.args.sync_bn:
print('creating batch sync model')
if self.args.model_ema:
print('creating batch sync ema model')
self.model_ema = self.context.wrap_model(deepcopy(self.model))
self.model = self.convert_syncbn_model(self.model)
self.model = self.context.wrap_model(self.model)
print('Model created, param count:', self.args.model,
sum([m.numel() for m in self.model.parameters()]))
self.optimizer = self.context.wrap_optimizer(
create_optimizer(self.args, self.model))
print('Created optimizer: ', self.optimizer)
if self.args.amp:
print('using amp')
if self.args.sync_bn and self.args.model_ema:
print('using sync_bn and model_ema when creating apex_amp')
(self.model, self.model_ema
), self.optimizer = self.context.configure_apex_amp(
[self.model, self.model_ema],
self.optimizer,
min_loss_scale=self.hparam("min_loss_scale"))
else:
self.model, self.optimizer = self.context.configure_apex_amp(
self.model,
self.optimizer,
min_loss_scale=self.hparam("min_loss_scale"))
if self.args.model_ema:
print('using model ema')
if self.args.sync_bn:
print('using model ema batch syn')
self.model_ema = ModelEma(self.model_ema,
context=self.context,
decay=self.args.model_ema_decay)
else:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
self.model_ema = ModelEma(self.model,
context=self.context,
decay=self.args.model_ema_decay)
self.lr_scheduler, self.num_epochs = create_scheduler(
self.args, self.optimizer)
self.lr_scheduler = self.context.wrap_lr_scheduler(
self.lr_scheduler, LRScheduler.StepMode.MANUAL_STEP)
self.cur_epoch = 0
self.num_updates = 0 * self.cur_epoch
if self.args.prefetcher:
self.train_mean, self.train_std, self.train_random_erasing = self.calculate_means(
mean=self.input_config['mean'],
std=self.input_config['std'],
re_prob=self.args.reprob,
re_mode=self.args.remode,
re_count=self.args.recount)
self.val_mean, self.val_std, self.val_random_erasing = self.calculate_means(
self.input_config['mean'], self.input_config['std'])
self.val_reducer = self.context.wrap_reducer(self.validation_reducer,
for_training=False)
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
logging.warning(
'Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.'
)
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
DistributedManager.set_args(args)
if args.distributed:
logging.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on %d GPUs.' %
args.num_gpu)
torch.manual_seed(args.seed + args.rank)
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect,
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint)
if args.initial_checkpoint_pruned:
try:
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
model2 = load_module_from_ckpt(
model,
args.initial_checkpoint_pruned,
input_size=data_config['input_size'][1])
logging.info("New pruned model adapted from the checkpoint")
except Exception as e:
raise RuntimeError(e)
else:
model2 = model
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model2.parameters()])))
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
if args.num_gpu > 1:
model2 = nn.DataParallel(model2,
device_ids=list(range(args.num_gpu))).cuda()
else:
model2.cuda()
use_amp = False
if args.distributed:
model2 = nn.parallel.distributed.DistributedDataParallel(
model2,
device_ids=[args.local_rank]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
logging.error(
'Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
if args.prefetcher and args.mixup > 0:
collate_fn = FastCollateMixup(args.mixup, args.smoothing,
args.num_classes)
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=args.train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'test')
if not os.path.isdir(eval_dir):
logging.error(
'Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
test_dir = os.path.join(args.data, 'test')
if not os.path.isdir(test_dir):
test_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(test_dir):
test_dir = os.path.join(args.data, 'val')
if not os.path.isdir(test_dir):
logging.error(
'Test folder does not exist at: {}'.format(test_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
if args.prune_test:
dataset_test = Dataset(test_dir)
else:
dataset_test = Dataset(train_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
len_loader = int(
len(loader_eval) * (4 * args.batch_size) / args.batch_size_prune)
if args.prune_test:
len_loader = None
if args.prune:
loader_p = create_loader(
dataset_test,
input_size=data_config['input_size'],
batch_size=args.batch_size_prune,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
if 'resnet' in model2.__class__.__name__.lower() or (
hasattr(model2, 'module')
and 'resnet' in model2.module.__class__.__name__.lower()):
list_channel_to_prune = compute_num_channels_per_layer_taylor(
model2,
data_config['input_size'],
loader_p,
pruning_ratio=args.pruning_ratio,
taylor_file=args.taylor_file,
local_rank=args.local_rank,
len_data_loader=len_loader,
prune_skip=args.prune_skip,
taylor_abs=args.taylor_abs,
prune_conv1=args.prune_conv1,
use_time=args.use_time,
distributed=args.distributed)
new_net = redesign_module_resnet(
model2,
list_channel_to_prune,
use_amp=use_amp,
distributed=args.distributed,
local_rank=args.local_rank,
input_size=data_config['input_size'][1])
else:
list_channel_to_prune = compute_num_channels_per_layer_taylor(
model2,
data_config['input_size'],
loader_p,
pruning_ratio=args.pruning_ratio,
taylor_file=args.taylor_file,
local_rank=args.local_rank,
len_data_loader=len_loader,
prune_pwl=not args.no_pwl,
taylor_abs=args.taylor_abs,
use_se=not args.use_eca,
use_time=args.use_time,
distributed=args.distributed)
new_net = redesign_module_efnet(
model2,
list_channel_to_prune,
use_amp=use_amp,
distributed=args.distributed,
local_rank=args.local_rank,
input_size=data_config['input_size'][1],
use_se=not args.use_eca)
new_net.train()
model.train()
if isinstance(model, nn.DataParallel) or isinstance(model, DDP):
model = model.module
else:
model = model.cuda()
co_mod = build_co_train_model(
model,
new_net.module.cpu() if hasattr(new_net, 'module') else new_net,
gamma=args.gamma_knowledge,
only_last=args.only_last,
progressive_IKD_factor=args.progressive_IKD_factor)
optimizer = create_optimizer(args, co_mod)
del model
del new_net
gc.collect()
torch.cuda.empty_cache()
if args.num_gpu > 1:
if args.amp:
logging.warning(
'AMP does not work well with nn.DataParallel, disabling. Use distributed mode for multi-GPU AMP.'
)
args.amp = False
co_mod = nn.DataParallel(co_mod,
device_ids=list(range(
args.num_gpu))).cuda()
else:
co_mod = co_mod.cuda()
use_amp = False
if has_apex and args.amp:
co_mod, optimizer = amp.initialize(co_mod,
optimizer,
opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed',
'on' if use_amp else 'off'))
if args.distributed:
if args.sync_bn:
try:
if has_apex and use_amp:
co_mod = convert_syncbn_model(co_mod)
else:
co_mod = torch.nn.SyncBatchNorm.convert_sync_batchnorm(
co_mod)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm.')
except Exception as e:
logging.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1'
)
if has_apex and use_amp:
co_mod = DDP(co_mod, delay_allreduce=False)
else:
if args.local_rank == 0 and use_amp:
logging.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
co_mod = nn.parallel.distributed.DistributedDataParallel(
co_mod,
device_ids=[args.local_rank
]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
co_mod.train()
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.jsd:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits,
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.mixup > 0.:
# smoothing is handled with mixup label transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"), args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir,
decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
if args.local_rank == 0:
logging.info(f'First validation')
co_mod.eval()
eval_metrics = validate(co_mod, loader_eval, validate_loss_fn, args)
if args.local_rank == 0:
logging.info(f'Prec@top1 : {eval_metrics["prec1"]}')
co_mod.train()
for epoch in range(start_epoch, num_epochs):
torch.cuda.empty_cache()
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(epoch,
co_mod,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=None)
torch.cuda.empty_cache()
eval_metrics = validate(co_mod, loader_eval, validate_loss_fn,
args)
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
co_mod,
optimizer,
args,
epoch=epoch,
model_ema=None,
metric=save_metric,
use_amp=use_amp)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
def main(args):
utils.init_distributed_mode(args)
print(args)
if args.distillation_type != 'none' and args.finetune and not args.eval:
raise NotImplementedError(
"Finetuning with distillation not yet supported")
device = torch.device(args.device)
# fix the seed for reproducibility
seed = args.seed + utils.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
# random.seed(seed)
cudnn.benchmark = True
dataset_train, args.nb_classes = build_dataset(is_train=True, args=args)
dataset_val, _ = build_dataset(is_train=False, args=args)
if True: # args.distributed:
num_tasks = utils.get_world_size()
global_rank = utils.get_rank()
if args.repeated_aug:
sampler_train = RASampler(dataset_train,
num_replicas=num_tasks,
rank=global_rank,
shuffle=True)
else:
sampler_train = torch.utils.data.DistributedSampler(
dataset_train,
num_replicas=num_tasks,
rank=global_rank,
shuffle=True)
if args.dist_eval:
if len(dataset_val) % num_tasks != 0:
print(
'Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. '
'This will slightly alter validation results as extra duplicate entries are added to achieve '
'equal num of samples per-process.')
sampler_val = torch.utils.data.DistributedSampler(
dataset_val,
num_replicas=num_tasks,
rank=global_rank,
shuffle=False)
else:
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
data_loader_train = torch.utils.data.DataLoader(
dataset_train,
sampler=sampler_train,
batch_size=args.batch_size,
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=True,
)
data_loader_val = torch.utils.data.DataLoader(dataset_val,
sampler=sampler_val,
batch_size=int(
1.5 * args.batch_size),
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=False)
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_fn = Mixup(mixup_alpha=args.mixup,
cutmix_alpha=args.cutmix,
cutmix_minmax=args.cutmix_minmax,
prob=args.mixup_prob,
switch_prob=args.mixup_switch_prob,
mode=args.mixup_mode,
label_smoothing=args.smoothing,
num_classes=args.nb_classes)
print(f"Creating model: {args.model}")
model = create_model(
args.model,
num_classes=args.nb_classes,
distillation=(args.distillation_type != 'none'),
pretrained=args.eval,
fuse=args.eval,
)
if args.finetune:
if args.finetune.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.finetune,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(args.finetune, map_location='cpu')
checkpoint_model = checkpoint['model']
state_dict = model.state_dict()
for k in [
'head.weight', 'head.bias', 'head_dist.weight',
'head_dist.bias'
]:
if k in checkpoint_model and checkpoint_model[
k].shape != state_dict[k].shape:
print(f"Removing key {k} from pretrained checkpoint")
del checkpoint_model[k]
model.load_state_dict(checkpoint_model, strict=False)
model.to(device)
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but
# before SyncBN and DDP wrapper
model_ema = ModelEma(model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume='')
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size(
) / 512.0
args.lr = linear_scaled_lr
optimizer = create_optimizer(args, model_without_ddp)
loss_scaler = NativeScaler()
lr_scheduler, _ = create_scheduler(args, optimizer)
criterion = LabelSmoothingCrossEntropy()
if args.mixup > 0.:
# smoothing is handled with mixup label transform
criterion = SoftTargetCrossEntropy()
elif args.smoothing:
criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
else:
criterion = torch.nn.CrossEntropyLoss()
teacher_model = None
if args.distillation_type != 'none':
assert args.teacher_path, 'need to specify teacher-path when using distillation'
print(f"Creating teacher model: {args.teacher_model}")
teacher_model = create_model(
args.teacher_model,
pretrained=False,
num_classes=args.nb_classes,
global_pool='avg',
)
if args.teacher_path.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.teacher_path,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(args.teacher_path, map_location='cpu')
teacher_model.load_state_dict(checkpoint['model'])
teacher_model.to(device)
teacher_model.eval()
# wrap the criterion in our custom DistillationLoss, which
# just dispatches to the original criterion if args.distillation_type is
# 'none'
criterion = DistillationLoss(criterion, teacher_model,
args.distillation_type,
args.distillation_alpha,
args.distillation_tau)
output_dir = Path(args.output_dir)
if args.resume:
if args.resume.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.resume,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.model_ema:
utils._load_checkpoint_for_ema(model_ema,
checkpoint['model_ema'])
if 'scaler' in checkpoint:
loss_scaler.load_state_dict(checkpoint['scaler'])
if args.eval:
test_stats = evaluate(data_loader_val, model, device)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
return
print(f"Start training for {args.epochs} epochs")
start_time = time.time()
max_accuracy = 0.0
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
data_loader_train.sampler.set_epoch(epoch)
train_stats = train_one_epoch(
model,
criterion,
data_loader_train,
optimizer,
device,
epoch,
loss_scaler,
args.clip_grad,
args.clip_mode,
model_ema,
mixup_fn,
set_training_mode=args.finetune ==
'' # keep in eval mode during finetuning
)
lr_scheduler.step(epoch)
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
for checkpoint_path in checkpoint_paths:
utils.save_on_master(
{
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
'model_ema': get_state_dict(model_ema),
'scaler': loss_scaler.state_dict(),
'args': args,
}, checkpoint_path)
if epoch % 20 == 19:
test_stats = evaluate(data_loader_val, model, device)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
max_accuracy = max(max_accuracy, test_stats["acc1"])
print(f'Max accuracy: {max_accuracy:.2f}%')
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()},
**{f'test_{k}': v
for k, v in test_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
else:
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def create_optimizer(opt_args, model):
return optim.create_optimizer(opt_args, model)
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
logging.warning(
'Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.'
)
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on %d GPUs.' %
args.num_gpu)
torch.manual_seed(args.seed + args.rank)
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint)
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model.parameters()])))
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
if args.num_gpu > 1:
if args.amp:
logging.warning(
'AMP does not work well with nn.DataParallel, disabling. Use distributed mode for multi-GPU AMP.'
)
args.amp = False
model = nn.DataParallel(model,
device_ids=list(range(args.num_gpu))).cuda()
else:
model.cuda()
optimizer = create_optimizer(args, model)
use_amp = False
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed',
'on' if use_amp else 'off'))
# optionally resume from a checkpoint
resume_state = {}
resume_epoch = None
if args.resume:
resume_state, resume_epoch = resume_checkpoint(model, args.resume)
if resume_state and not args.no_resume_opt:
if 'optimizer' in resume_state:
if args.local_rank == 0:
logging.info('Restoring Optimizer state from checkpoint')
optimizer.load_state_dict(resume_state['optimizer'])
if use_amp and 'amp' in resume_state and 'load_state_dict' in amp.__dict__:
if args.local_rank == 0:
logging.info('Restoring NVIDIA AMP state from checkpoint')
amp.load_state_dict(resume_state['amp'])
resume_state = None
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume=args.resume)
if args.distributed:
if args.sync_bn:
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(
model)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm.')
except Exception as e:
logging.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1'
)
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
model = DDP(model,
device_ids=[args.local_rank
]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
logging.error(
'Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
if args.prefetcher and args.mixup > 0:
collate_fn = FastCollateMixup(args.mixup, args.smoothing,
args.num_classes)
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
rand_erase_prob=args.reprob,
rand_erase_mode=args.remode,
rand_erase_count=args.recount,
color_jitter=args.color_jitter,
auto_augment=args.aa,
interpolation=
'random', # FIXME cleanly resolve this? data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
logging.error(
'Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=4 * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
)
if args.mixup > 0.:
# smoothing is handled with mixup label transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"), args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir,
decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema)
eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
if model_ema is not None and not args.model_ema_force_cpu:
ema_eval_metrics = validate(model_ema.ema,
loader_eval,
validate_loss_fn,
args,
log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
model,
optimizer,
args,
epoch=epoch,
model_ema=model_ema,
metric=save_metric,
use_amp=use_amp)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
_logger.info('Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
_logger.info('Training with a single process on 1 GPUs.')
assert args.rank >= 0
if args.control_amp == 'amp':
args.amp = True
elif args.control_amp == 'apex':
args.apex_amp = True
elif args.control_amp == 'native':
args.native_amp = True
# resolve AMP arguments based on PyTorch / Apex availability
use_amp = None
if args.amp:
# for backwards compat, `--amp` arg tries apex before native amp
if has_apex:
args.apex_amp = True
elif has_native_amp:
args.native_amp = True
if args.apex_amp and has_apex:
use_amp = 'apex'
elif args.native_amp and has_native_amp:
use_amp = 'native'
elif args.apex_amp or args.native_amp:
_logger.warning("Neither APEX or native Torch AMP is available, using float32. "
"Install NVIDA apex or upgrade to PyTorch 1.6")
_logger.info(
'====================\n\n'
'Actfun: {}\n'
'LR: {}\n'
'Epochs: {}\n'
'p: {}\n'
'k: {}\n'
'g: {}\n'
'Extra channel multiplier: {}\n'
'AMP: {}\n'
'Weight Init: {}\n'
'\n===================='.format(args.actfun, args.lr, args.epochs, args.p, args.k, args.g,
args.extra_channel_mult, use_amp, args.weight_init))
torch.manual_seed(args.seed + args.rank)
model = create_model(
args.model,
pretrained=args.pretrained,
actfun=args.actfun,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect, # DEPRECATED, use drop_path
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
scriptable=args.torchscript,
checkpoint_path=args.initial_checkpoint,
p=args.p,
k=args.k,
g=args.g,
extra_channel_mult=args.extra_channel_mult,
weight_init_name=args.weight_init,
partial_ho_actfun=args.partial_ho_actfun
)
if args.tl:
if args.data == 'caltech101' and not os.path.exists('caltech101'):
dir_root = r'101_ObjectCategories'
dir_new = r'caltech101'
dir_new_train = os.path.join(dir_new, 'train')
dir_new_val = os.path.join(dir_new, 'val')
dir_new_test = os.path.join(dir_new, 'test')
if not os.path.exists(dir_new):
os.mkdir(dir_new)
os.mkdir(dir_new_train)
os.mkdir(dir_new_val)
os.mkdir(dir_new_test)
for dir2 in os.listdir(dir_root):
if dir2 != 'BACKGROUND_Google':
curr_path = os.path.join(dir_root, dir2)
new_path_train = os.path.join(dir_new_train, dir2)
new_path_val = os.path.join(dir_new_val, dir2)
new_path_test = os.path.join(dir_new_test, dir2)
if not os.path.exists(new_path_train):
os.mkdir(new_path_train)
if not os.path.exists(new_path_val):
os.mkdir(new_path_val)
if not os.path.exists(new_path_test):
os.mkdir(new_path_test)
train_upper = int(0.8 * len(os.listdir(curr_path)))
val_upper = int(0.9 * len(os.listdir(curr_path)))
curr_files_all = os.listdir(curr_path)
curr_files_train = curr_files_all[:train_upper]
curr_files_val = curr_files_all[train_upper:val_upper]
curr_files_test = curr_files_all[val_upper:]
for file in curr_files_train:
copyfile(os.path.join(curr_path, file),
os.path.join(new_path_train, file))
for file in curr_files_val:
copyfile(os.path.join(curr_path, file),
os.path.join(new_path_val, file))
for file in curr_files_test:
copyfile(os.path.join(curr_path, file),
os.path.join(new_path_test, file))
time.sleep(5)
if args.tl:
pre_model = create_model(
args.model,
pretrained=True,
actfun='swish',
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect, # DEPRECATED, use drop_path
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
scriptable=args.torchscript,
checkpoint_path=args.initial_checkpoint,
p=args.p,
k=args.k,
g=args.g,
extra_channel_mult=args.extra_channel_mult,
weight_init_name=args.weight_init,
partial_ho_actfun=args.partial_ho_actfun
)
model = MLP.MLP(actfun=args.actfun,
input_dim=1280,
output_dim=args.num_classes,
k=args.k,
p=args.p,
g=args.g,
num_params=400_000,
permute_type='shuffle')
pre_model_layers = list(pre_model.children())
pre_model = torch.nn.Sequential(*pre_model_layers[:-1])
else:
pre_model = None
if args.local_rank == 0:
_logger.info('Model %s created, param count: %d' %
(args.model, sum([m.numel() for m in model.parameters()])))
data_config = resolve_data_config(vars(args), model=model, verbose=args.local_rank == 0)
# setup augmentation batch splits for contrastive loss or split bn
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
# enable split bn (separate bn stats per batch-portion)
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
# move model to GPU, enable channels last layout if set
model.cuda()
if args.tl:
pre_model.cuda()
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
# setup synchronized BatchNorm for distributed training
if args.distributed and args.sync_bn:
assert not args.split_bn
if has_apex and use_amp != 'native':
# Apex SyncBN preferred unless native amp is activated
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0:
_logger.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.')
if args.torchscript:
assert not use_amp == 'apex', 'Cannot use APEX AMP with torchscripted model'
assert not args.sync_bn, 'Cannot use SyncBatchNorm with torchscripted model'
model = torch.jit.script(model)
if args.tl:
optimizer = torch.optim.Adam(model.parameters(), weight_decay=1e-5)
else:
optimizer = create_optimizer(args, model)
# setup automatic mixed-precision (AMP) loss scaling and op casting
amp_autocast = suppress # do nothing
loss_scaler = None
if use_amp == 'apex':
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
loss_scaler = ApexScaler()
if args.local_rank == 0:
_logger.info('Using NVIDIA APEX AMP. Training in mixed precision.')
elif use_amp == 'native':
amp_autocast = torch.cuda.amp.autocast
loss_scaler = NativeScaler()
if args.local_rank == 0:
_logger.info('Using native Torch AMP. Training in mixed precision.')
else:
if args.local_rank == 0:
_logger.info('AMP not enabled. Training in float32.')
if args.local_rank == 0:
_logger.info('\n--------------------\nModel:\n' + repr(model) + '--------------------')
# optionally resume from a checkpoint
resume_epoch = None
resume_path = os.path.join(args.resume, 'recover.pth.tar')
if args.resume and os.path.exists(resume_path):
resume_epoch = resume_checkpoint(
model, resume_path,
optimizer=None if args.no_resume_opt else optimizer,
loss_scaler=None if args.no_resume_opt else loss_scaler,
log_info=args.local_rank == 0)
cp_loaded = None
resume_epoch = None
checkname = 'recover'
if args.actfun != 'swish':
checkname = '{}_'.format(args.actfun) + checkname
check_path = os.path.join(args.check_path, checkname) + '.pth'
loader = None
if os.path.isfile(check_path):
loader = check_path
elif args.load_path != '' and os.path.isfile(args.load_path):
loader = args.load_path
if loader is not None:
cp_loaded = torch.load(loader)
model.load_state_dict(cp_loaded['model'])
optimizer.load_state_dict(cp_loaded['optimizer'])
resume_epoch = cp_loaded['epoch']
model.cuda()
loss_scaler.load_state_dict(cp_loaded['amp'])
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
_logger.info('============ LOADED CHECKPOINT: Epoch {}'.format(resume_epoch))
model_raw = model
# setup exponential moving average of model weights, SWA could be used here too
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEmaV2(
model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else None)
if args.resume and os.path.exists(resume_path):
load_checkpoint(model_ema.module, args.resume, use_ema=True)
if cp_loaded is not None:
model_ema.load_state_dict(cp_loaded['model_ema'])
# setup distributed training
if args.distributed:
if has_apex and use_amp != 'native':
# Apex DDP preferred unless native amp is activated
if args.local_rank == 0:
_logger.info("Using NVIDIA APEX DistributedDataParallel.")
model = ApexDDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
_logger.info("Using native Torch DistributedDataParallel.")
model = NativeDDP(model, device_ids=[args.local_rank]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
# setup mixup / cutmix
collate_fn = None
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_args = dict(
mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax,
prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode,
label_smoothing=args.smoothing, num_classes=args.num_classes)
if args.prefetcher:
assert not num_aug_splits # collate conflict (need to support deinterleaving in collate mixup)
collate_fn = FastCollateMixup(**mixup_args)
else:
mixup_fn = Mixup(**mixup_args)
# create the train and eval datasets
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
_logger.error('Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
_logger.error('Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
# wrap dataset in AugMix helper
if num_aug_splits > 1:
dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
# create data loaders w/ augmentation pipeline
train_interpolation = args.train_interpolation
if args.no_aug or not train_interpolation:
train_interpolation = data_config['interpolation']
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
no_aug=args.no_aug,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
scale=args.scale,
ratio=args.ratio,
hflip=args.hflip,
vflip=args.vflip,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
use_multi_epochs_loader=args.use_multi_epochs_loader
)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
# setup learning rate schedule and starting epoch
lr_scheduler, num_epochs = create_scheduler(args, optimizer, dataset_train)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if cp_loaded is not None:
lr_scheduler.load_state_dict(cp_loaded['scheduler'])
if args.local_rank == 0:
_logger.info('Scheduled epochs: {}'.format(num_epochs))
# setup loss function
if args.jsd:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits, smoothing=args.smoothing).cuda()
elif mixup_active:
# smoothing is handled with mixup target transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(smoothing=args.smoothing).cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
# setup checkpoint saver and eval metric tracking
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(
model=model, optimizer=optimizer, args=args, model_ema=model_ema, amp_scaler=loss_scaler,
checkpoint_dir=output_dir, recovery_dir=args.resume, decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
fieldnames = ['seed', 'weight_init', 'actfun', 'epoch', 'max_lr', 'lr', 'train_loss', 'eval_loss', 'eval_acc1', 'eval_acc5', 'ema']
filename = 'output'
if args.actfun != 'swish':
filename = '{}_'.format(args.actfun) + filename
outfile_path = os.path.join(args.output, filename) + '.csv'
if not os.path.exists(outfile_path):
with open(outfile_path, mode='w') as out_file:
writer = csv.DictWriter(out_file, fieldnames=fieldnames, lineterminator='\n')
writer.writeheader()
try:
for epoch in range(start_epoch, num_epochs):
if os.path.exists(args.check_path):
amp_loss = None
if use_amp == 'native':
amp_loss = loss_scaler.state_dict()
elif use_amp == 'apex':
amp_loss = amp.state_dict()
if model_ema is not None:
ema_save = model_ema.state_dict()
else:
ema_save = None
torch.save({'model': model_raw.state_dict(),
'model_ema': ema_save,
'optimizer': optimizer.state_dict(),
'scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
'amp': amp_loss
}, check_path)
_logger.info('============ SAVED CHECKPOINT: Epoch {}'.format(epoch))
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(
epoch, model, loader_train, optimizer, train_loss_fn, args,
lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir,
amp_autocast=amp_autocast, loss_scaler=loss_scaler, model_ema=model_ema, mixup_fn=mixup_fn,
pre_model=pre_model)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
_logger.info("Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
eval_metrics = validate(model, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast,
pre_model=pre_model)
with open(outfile_path, mode='a') as out_file:
writer = csv.DictWriter(out_file, fieldnames=fieldnames, lineterminator='\n')
writer.writerow({'seed': args.seed,
'actfun': args.actfun,
'epoch': epoch,
'lr': train_metrics['lr'],
'train_loss': train_metrics['loss'],
'eval_loss': eval_metrics['loss'],
'eval_acc1': eval_metrics['top1'],
'eval_acc5': eval_metrics['top5'],
'ema': False
})
if model_ema is not None and not args.model_ema_force_cpu:
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
distribute_bn(model_ema, args.world_size, args.dist_bn == 'reduce')
ema_eval_metrics = validate(
model_ema.module, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast, log_suffix=' (EMA)',
pre_model=pre_model)
eval_metrics = ema_eval_metrics
with open(outfile_path, mode='a') as out_file:
writer = csv.DictWriter(out_file, fieldnames=fieldnames, lineterminator='\n')
writer.writerow({'seed': args.seed,
'weight_init': args.weight_init,
'actfun': args.actfun,
'epoch': epoch,
'max_lr': args.lr,
'lr': train_metrics['lr'],
'train_loss': train_metrics['loss'],
'eval_loss': eval_metrics['loss'],
'eval_acc1': eval_metrics['top1'],
'eval_acc5': eval_metrics['top5'],
'ema': True
})
if lr_scheduler is not None and args.sched != 'onecycle':
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
update_summary(
args.seed, epoch, args.lr, args.epochs, args.batch_size, args.actfun,
train_metrics, eval_metrics, os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(epoch, metric=save_metric)
except KeyboardInterrupt:
pass
if best_metric is not None:
_logger.info('*** Best metric: {0} (epoch {1})'.format(best_metric, best_epoch))
def main():
setup_default_logging()
args = parser.parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
logging.warning(
'Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.'
)
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on %d GPUs.' %
args.num_gpu)
torch.manual_seed(args.seed + args.rank)
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
drop_connect_rate=0.2,
checkpoint_path=args.initial_checkpoint,
args=args)
flops, params = get_model_complexity_info(
model, (3, 224, 224),
as_strings=True,
print_per_layer_stat=args.display_info)
print('Flops: ' + flops)
print('Params: ' + params)
if args.KD_train:
teacher_model = create_model("efficientnet_b7_dq",
pretrained=True,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
drop_connect_rate=0.2,
checkpoint_path=args.initial_checkpoint,
args=args)
flops_teacher, params_teacher = get_model_complexity_info(
teacher_model, (3, 224, 224),
as_strings=True,
print_per_layer_stat=False)
print("Using KD training...")
print("FLOPs of teacher model: ", flops_teacher)
print("Params of teacher model: ", params_teacher)
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model.parameters()])))
data_config = resolve_data_config(model,
args,
verbose=args.local_rank == 0)
# optionally resume from a checkpoint
start_epoch = 0
optimizer_state = None
if args.resume:
optimizer_state, start_epoch = resume_checkpoint(
model, args.resume, args.start_epoch)
# import pdb;pdb.set_trace()
if args.num_gpu > 1:
if args.amp:
logging.warning(
'AMP does not work well with nn.DataParallel, disabling. Use distributed mode for multi-GPU AMP.'
)
args.amp = False
model = nn.DataParallel(model,
device_ids=list(range(args.num_gpu))).cuda()
if args.KD_train:
teacher_model = nn.DataParallel(teacher_model,
device_ids=list(range(
args.num_gpu))).cuda()
else:
model.cuda()
if args.KD_train:
teacher_model.cuda()
optimizer = create_optimizer(args, model)
if optimizer_state is not None:
optimizer.load_state_dict(optimizer_state)
use_amp = False
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed',
'on' if use_amp else 'off'))
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
# import pdb; pdb.set_trace()
model_ema = ModelEma(model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume=args.resume)
if args.distributed:
if args.sync_bn:
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(
model)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm.')
except Exception as e:
logging.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1'
)
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
model = DDP(model,
device_ids=[args.local_rank
]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
if start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
logging.error(
'Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
if args.prefetcher and args.mixup > 0:
collate_fn = FastCollateMixup(args.mixup, args.smoothing,
args.num_classes)
if args.auto_augment:
print('using auto data augumentation...')
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
rand_erase_prob=args.reprob,
rand_erase_mode=args.remode,
interpolation=
'bicubic', # FIXME cleanly resolve this? data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
use_auto_aug=args.auto_augment,
use_mixcut=args.mixcut,
)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
logging.error(
'Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=4 * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
)
if args.mixup > 0.:
# smoothing is handled with mixup label transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
if args.KD_train:
train_loss_fn = nn.KLDivLoss(reduction='batchmean').cuda()
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"), args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir,
decreasing=decreasing)
try:
# import pdb;pdb.set_trace()
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
# import pdb; pdb.set_trace()
if args.KD_train:
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema,
teacher_model=teacher_model)
else:
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema)
# def __init__(self, model, bits_activations=8, bits_parameters=8, bits_accum=32,
# overrides=None, mode=LinearQuantMode.SYMMETRIC, clip_acts=ClipMode.NONE,
# per_channel_wts=False, model_activation_stats=None, fp16=False, clip_n_stds=None,
# scale_approx_mult_bits=None):
# import distiller
# import pdb; pdb.set_trace()
# quantizer = quantization.PostTrainLinearQuantizer.from_args(model, args)
# quantizer.prepare_model(distiller.get_dummy_input(input_shape=model.input_shape))
# quantizer = distiller.quantization.PostTrainLinearQuantizer(model, bits_activations=8, bits_parameters=8)
# quantizer.prepare_model()
# distiller.utils.assign_layer_fq_names(model)
# # msglogger.info("Generating quantization calibration stats based on {0} users".format(args.qe_calibration))
# collector = distiller.data_loggers.QuantCalibrationStatsCollector(model)
# with collector_context(collector):
# eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
# # Here call your model evaluation function, making sure to execute only
# # the portion of the dataset specified by the qe_calibration argument
# yaml_path = './dir/quantization_stats.yaml'
# collector.save(yaml_path)
eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
if model_ema is not None and not args.model_ema_force_cpu:
ema_eval_metrics = validate(model_ema.ema,
loader_eval,
validate_loss_fn,
args,
log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
lr_scheduler.step(epoch, eval_metrics[eval_metric])
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
model,
optimizer,
args,
epoch=epoch + 1,
model_ema=model_ema,
metric=save_metric)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
def main():
setup_default_logging()
args, args_text = _parse_args()
args.pretrained_backbone = not args.no_pretrained_backbone
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on 1 GPU.')
torch.manual_seed(args.seed + args.rank)
model = create_model(
args.model,
bench_task='train',
pretrained=args.pretrained,
pretrained_backbone=args.pretrained_backbone,
redundant_bias=args.redundant_bias,
checkpoint_path=args.initial_checkpoint,
)
# FIXME decide which args to keep and overlay on config / pass to backbone
# num_classes=args.num_classes,
# drop_rate=args.drop,
# drop_path_rate=args.drop_path,
# drop_block_rate=args.drop_block,
input_size = model.config.image_size
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model.parameters()])))
model.cuda()
optimizer = create_optimizer(args, model)
use_amp = False
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed',
'on' if use_amp else 'off'))
# optionally resume from a checkpoint
resume_state = {}
resume_epoch = None
if args.resume:
resume_state, resume_epoch = resume_checkpoint(unwrap_bench(model),
args.resume)
if resume_state and not args.no_resume_opt:
if 'optimizer' in resume_state:
if args.local_rank == 0:
logging.info('Restoring Optimizer state from checkpoint')
optimizer.load_state_dict(resume_state['optimizer'])
if use_amp and 'amp' in resume_state and 'load_state_dict' in amp.__dict__:
if args.local_rank == 0:
logging.info('Restoring NVIDIA AMP state from checkpoint')
amp.load_state_dict(resume_state['amp'])
del resume_state
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(model, decay=args.model_ema_decay)
#resume=args.resume) # FIXME bit of a mess with bench
if args.resume:
load_checkpoint(unwrap_bench(model_ema), args.resume, use_ema=True)
if args.distributed:
if args.sync_bn:
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(
model)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.'
)
except Exception as e:
logging.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1'
)
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
model = DDP(model,
device_ids=[args.local_rank
]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
# train_anno_set = 'train2017'
# train_annotation_path = os.path.join(args.data, 'annotations', f'instances_{train_anno_set}.json')
# train_image_dir = train_anno_set
#dataset_train = CocoDetection("/workspace/data/images",
# "/workspace/data/datatrain90n.json")
train_anno_set = 'train'
train_annotation_path = os.path.join(args.data, 'annotations',
f'instances_{train_anno_set}.json')
train_image_dir = train_anno_set
dataset_train = CocoDetection(os.path.join(args.data, train_image_dir),
train_annotation_path)
# FIXME cutmix/mixup worth investigating?
# collate_fn = None
# if args.prefetcher and args.mixup > 0:
# collate_fn = FastCollateMixup(args.mixup, args.smoothing, args.num_classes)
loader_train = create_loader(
dataset_train,
input_size=input_size,
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
#re_prob=args.reprob, # FIXME add back various augmentations
#re_mode=args.remode,
#re_count=args.recount,
#re_split=args.resplit,
#color_jitter=args.color_jitter,
#auto_augment=args.aa,
interpolation=args.train_interpolation,
mean=[0.4533, 0.4744,
0.4722], #[0.4846, 0.5079, 0.5005],#[0.485, 0.456, 0.406],
std=[0.2823, 0.2890,
0.3084], #[0.2687, 0.2705, 0.2869],#[0.485, 0.456, 0.406],
num_workers=args.workers,
distributed=args.distributed,
#collate_fn=collate_fn,
pin_mem=args.pin_mem,
)
train_anno_set = 'val'
train_annotation_path = os.path.join(args.data, 'annotations',
f'instances_{train_anno_set}.json')
train_image_dir = train_anno_set
dataset_eval = CocoDetection(os.path.join(args.data, train_image_dir),
train_annotation_path)
# train_anno_set = 'val'
# train_annotation_path = os.path.join(args.data, 'annotations', f'instances_{train_anno_set}.json')
# train_image_dir = train_anno_set
# dataset_eval = CocoDetection("/workspace/data/val/images",
# "/workspace/data/dataval90n.json")
loader_eval = create_loader(
dataset_eval,
input_size=input_size,
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=args.interpolation,
mean=[0.4535, 0.4744, 0.4724], #[0.4851, 0.5083, 0.5009],
std=[0.2835, 0.2903, 0.3098], #[0.2690, 0.2709, 0.2877],
num_workers=args.workers,
#distributed=args.distributed,
pin_mem=args.pin_mem,
)
# for xx,item in dataset_train :
# print("out",type(xx))
# break
# exit()
array_of_gt = []
if args.local_rank == 0:
for _, item in tqdm(dataset_eval):
# print(item)
for i in range(len(item['cls'])):
array_of_gt.append(
BoundingBox(imageName=str(item["img_id"]),
classId=item["cls"][i],
x=item["bbox"][i][1] * item['img_scale'],
y=item["bbox"][i][0] * item['img_scale'],
w=item["bbox"][i][3] * item['img_scale'],
h=item["bbox"][i][2] * item['img_scale'],
typeCoordinates=CoordinatesType.Absolute,
bbType=BBType.GroundTruth,
format=BBFormat.XYX2Y2,
imgSize=(item['img_size'][0],
item['img_size'][1])))
evaluator = COCOEvaluator(dataset_eval.coco,
distributed=args.distributed,
gtboxes=array_of_gt)
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join(
[datetime.now().strftime("%Y%m%d-%H%M%S"), args.model])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir,
decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
# print(model)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
logging.info(
"Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
# the overhead of evaluating with coco style datasets is fairly high, so just ema or non, not both
if model_ema is not None:
if args.distributed and args.dist_bn in ('broadcast',
'reduce'):
distribute_bn(model_ema, args.world_size,
args.dist_bn == 'reduce')
eval_metrics = validate(model_ema.ema,
loader_eval,
args,
evaluator,
log_suffix=' (EMA)')
else:
eval_metrics = validate(model, loader_eval, args, evaluator)
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
if saver is not None:
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
unwrap_bench(model),
optimizer,
args,
epoch=epoch,
model_ema=unwrap_bench(model_ema),
metric=save_metric,
use_amp=use_amp)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
model = models.init_model(name=args.arch, use_gpu=use_gpu, num_classes=dataset.num_train_pids, loss={'xent', 'htri'}, transformer_num_heads=args.transformer_num_heads, transformer_num_layers=args.transformer_num_layers, attention_flatness=True)
#print(model)
if args.resume:
print("Loading checkpoint from '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
criterion_xent = nn.CrossEntropyLoss()
criterion_flat = FlatnessLoss(reduction='batchmean', use_gpu=use_gpu)
criterion_htri_c = TripletInterCamLoss(margin=args.margin, distance=args.distance, use_gpu=use_gpu)
#criterion_htri_c = TripletWeightedInterCamLoss(margin=args.margin, distance=args.distance, use_gpu=use_gpu, alpha=args.cam_alpha)
#optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
linear_scaled_lr = args.lr * args.train_batch * len(args.gpu_devices.split(',')) / 512.0
args.lr = linear_scaled_lr
optimizer = create_optimizer(args, model)
#scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=args.stepsize, gamma=args.gamma)
scheduler, _ = create_scheduler(args, optimizer)
start_epoch = args.start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
#model = model.cuda()
#model = nn.parallel.DistributedDataParallel(model)
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
def train_imagenet_dq():
setup_default_logging()
args = parser.parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
logging.warning('Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.')
args.num_gpu = 1
args.device = xm.xla_device()
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info('Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on %d GPUs.' % args.num_gpu)
torch.manual_seed(args.seed + args.rank)
device = xm.xla_device()
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
drop_connect_rate=0.2,
checkpoint_path=args.initial_checkpoint,
args = args).to(device)
flops, params = get_model_complexity_info(model, (3, 224, 224), as_strings=True, print_per_layer_stat=args.display_info)
print('Flops: ' + flops)
print('Params: ' + params)
if args.KD_train:
teacher_model = create_model(
"efficientnet_b7_dq",
pretrained=True,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
drop_connect_rate=0.2,
checkpoint_path=args.initial_checkpoint,
args = args)
flops_teacher, params_teacher = get_model_complexity_info(teacher_model, (3, 224, 224), as_strings=True, print_per_layer_stat=False)
print("Using KD training...")
print("FLOPs of teacher model: ", flops_teacher)
print("Params of teacher model: ", params_teacher)
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel() for m in model.parameters()])))
data_config = resolve_data_config(model, args, verbose=args.local_rank == 0)
# optionally resume from a checkpoint
start_epoch = 0
optimizer_state = None
if args.resume:
optimizer_state, start_epoch = resume_checkpoint(model, args.resume, args.start_epoch)
# import pdb;pdb.set_trace()
torch.manual_seed(42)
if args.num_gpu > 1:
if args.amp:
logging.warning(
'AMP does not work well with nn.DataParallel, disabling. Use distributed mode for multi-GPU AMP.')
args.amp = False
# device = xm.xla_device()
# devices = (
# xm.get_xla_supported_devices(
# max_devices=num_cores) if num_cores != 0 else [])
# model = nn.DataParallel(model, device_ids=devices).cuda()
# model = model.to(device)
if args.KD_train:
teacher_model = nn.DataParallel(teacher_model, device_ids=list(range(args.num_gpu))).cuda()
else:
# device = xm.xla_device()
# model = model.to(device)
if args.KD_train:
teacher_model.cuda()
optimizer = create_optimizer(args, model)
if optimizer_state is not None:
optimizer.load_state_dict(optimizer_state)
use_amp = False
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed', 'on' if use_amp else 'off'))
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
# import pdb; pdb.set_trace()
model_e = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
drop_connect_rate=0.2,
checkpoint_path=args.initial_checkpoint,
args = args).to(device)
model_ema = ModelEma(
model_e,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume=args.resume)
if args.distributed:
if args.sync_bn:
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0:
logging.info('Converted model to use Synchronized BatchNorm.')
except Exception as e:
logging.error('Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1')
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info("Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP.")
model = DDP(model, device_ids=[args.local_rank]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
if start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
logging.error('Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
if args.prefetcher and args.mixup > 0:
collate_fn = FastCollateMixup(args.mixup, args.smoothing, args.num_classes)
if args.auto_augment:
print('using auto data augumentation...')
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
rand_erase_prob=args.reprob,
rand_erase_mode=args.remode,
interpolation='bicubic', # FIXME cleanly resolve this? data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
use_auto_aug=args.auto_augment,
use_mixcut=args.mixcut,
)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
logging.error('Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size = args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
)
if args.mixup > 0.:
# smoothing is handled with mixup label transform
train_loss_fn = SoftTargetCrossEntropy()
validate_loss_fn = nn.CrossEntropyLoss()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
validate_loss_fn = nn.CrossEntropyLoss()
else:
train_loss_fn = nn.CrossEntropyLoss()
validate_loss_fn = train_loss_fn
if args.KD_train:
train_loss_fn = nn.KLDivLoss(reduction='batchmean')
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir, decreasing=decreasing)
def train_epoch(
epoch, model, loader, optimizer, loss_fn, args,
lr_scheduler=None, saver=None, output_dir='', use_amp=False, model_ema=None, teacher_model = None, loader_len=0):
if args.prefetcher and args.mixup > 0 and loader.mixup_enabled:
if args.mixup_off_epoch and epoch >= args.mixup_off_epoch:
loader.mixup_enabled = False
batch_time_m = AverageMeter()
data_time_m = AverageMeter()
losses_m = AverageMeter()
model.train()
if args.KD_train:
teacher_model.eval()
end = time.time()
last_idx = loader_len - 1
num_updates = epoch * loader_len
for batch_idx, (input, target) in loader:
last_batch = batch_idx == last_idx
data_time_m.update(time.time() - end)
if not args.prefetcher:
# input = input.cuda()
# target = target.cuda()
if args.mixup > 0.:
lam = 1.
if not args.mixup_off_epoch or epoch < args.mixup_off_epoch:
lam = np.random.beta(args.mixup, args.mixup)
input.mul_(lam).add_(1 - lam, input.flip(0))
target = mixup_target(target, args.num_classes, lam, args.smoothing)
r = np.random.rand(1)
if args.beta > 0 and r < args.cutmix_prob:
# generate mixed sample
lam = np.random.beta(args.beta, args.beta)
rand_index = torch.randperm(input.size()[0])
target_a = target
target_b = target[rand_index]
bbx1, bby1, bbx2, bby2 = rand_bbox(input.size(), lam)
input[:, :, bbx1:bbx2, bby1:bby2] = input[rand_index, :, bbx1:bbx2, bby1:bby2]
# adjust lambda to exactly match pixel ratio
lam = 1 - ((bbx2 - bbx1) * (bby2 - bby1) / (input.size()[-1] * input.size()[-2]))
# compute output
input_var = torch.autograd.Variable(input, requires_grad=True)
target_a_var = torch.autograd.Variable(target_a)
target_b_var = torch.autograd.Variable(target_b)
output = model(input_var)
loss = loss_fn(output, target_a_var) * lam + loss_fn(output, target_b_var) * (1. - lam)
else:
# NOTE KD Train is exclusive with mixcut, FIX it later
output = model(input)
if args.KD_train:
# teacher_model.cuda()
teacher_outputs_tmp = []
assert(input.shape[0]%args.teacher_step == 0)
step_size = int(input.shape[0]//args.teacher_step)
with torch.no_grad():
for k in range(0,int(input.shape[0]),step_size):
input_tmp = input[k:k+step_size,:,:,:]
teacher_outputs_tmp.append(teacher_model(input_tmp))
# torch.cuda.empty_cache()
# import pdb; pdb.set_trace()
teacher_outputs = torch.cat(teacher_outputs_tmp)
alpha = args.KD_alpha
T = args.KD_temperature
loss = loss_fn(F.log_softmax(output/T, dim=1),
F.softmax(teacher_outputs/T, dim=1)) * (alpha * T * T) + \
F.cross_entropy(output, target) * (1. - alpha)
else:
loss = loss_fn(output, target)
if not args.distributed:
losses_m.update(loss.item(), input.size(0))
optimizer.zero_grad()
if use_amp:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
#optimizer.step()
xm.optimizer_step(optimizer)
# torch.cuda.synchronize()
if model_ema is not None:
model_ema.update(model)
num_updates += 1
batch_time_m.update(time.time() - end)
if last_batch or batch_idx % args.log_interval == 0:
lrl = [param_group['lr'] for param_group in optimizer.param_groups]
lr = sum(lrl) / len(lrl)
if args.distributed:
reduced_loss = reduce_tensor(loss.data, args.world_size)
losses_m.update(reduced_loss.item(), input.size(0))
if args.local_rank == 0:
logging.info(
'Train: {} [{:>4d}/{} ({:>3.0f}%)] '
'Loss: {loss.val:>9.6f} ({loss.avg:>6.4f}) '
'Time: {batch_time.val:.3f}s, {rate:>7.2f}/s '
'({batch_time.avg:.3f}s, {rate_avg:>7.2f}/s) '
'LR: {lr:.3e} '
'Data: {data_time.val:.3f} ({data_time.avg:.3f})'.format(
epoch,
batch_idx, loader_len,
100. * batch_idx / last_idx,
loss=losses_m,
batch_time=batch_time_m,
rate=input.size(0) * args.world_size / batch_time_m.val,
rate_avg=input.size(0) * args.world_size / batch_time_m.avg,
lr=lr,
data_time=data_time_m))
if args.save_images and output_dir:
torchvision.utils.save_image(
input,
os.path.join(output_dir, 'train-batch-%d.jpg' % batch_idx),
padding=0,
normalize=True)
if saver is not None and args.recovery_interval and (
last_batch or (batch_idx + 1) % args.recovery_interval == 0):
save_epoch = epoch + 1 if last_batch else epoch
saver.save_recovery(
model, optimizer, args, save_epoch, model_ema=model_ema, batch_idx=batch_idx)
if lr_scheduler is not None:
lr_scheduler.step_update(num_updates=num_updates, metric=losses_m.avg)
end = time.time()
return OrderedDict([('loss', losses_m.avg)])
def validate(model, loader, loss_fn, args, log_suffix='',loader_len=0):
batch_time_m = AverageMeter()
losses_m = AverageMeter()
prec1_m = AverageMeter()
prec5_m = AverageMeter()
model.eval()
end = time.time()
last_idx = loader_len - 1
with torch.no_grad():
for batch_idx, (input, target) in loader:
last_batch = batch_idx == last_idx
# if not args.prefetcher:
# input = input.cuda()
# target = target.cuda()
output = model(input)
if isinstance(output, (tuple, list)):
output = output[0]
# augmentation reduction
reduce_factor = args.tta
if reduce_factor > 1:
output = output.unfold(0, reduce_factor, reduce_factor).mean(dim=2)
target = target[0:target.size(0):reduce_factor]
loss = loss_fn(output, target)
prec1, prec5 = accuracy(output, target, topk=(1, 5))
if args.distributed:
reduced_loss = reduce_tensor(loss.data, args.world_size)
prec1 = reduce_tensor(prec1, args.world_size)
prec5 = reduce_tensor(prec5, args.world_size)
else:
reduced_loss = loss.data
# torch.cuda.synchronize()
losses_m.update(reduced_loss.item(), input.size(0))
prec1_m.update(prec1.item(), output.size(0))
prec5_m.update(prec5.item(), output.size(0))
batch_time_m.update(time.time() - end)
end = time.time()
if args.local_rank == 0 and (last_batch or batch_idx % args.log_interval == 0):
log_name = 'Test' + log_suffix
logging.info(
'{0}: [{1:>4d}/{2}] '
'Time: {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) '
'Prec@1: {top1.val:>7.4f} ({top1.avg:>7.4f}) '
'Prec@5: {top5.val:>7.4f} ({top5.avg:>7.4f})'.format(
log_name, batch_idx, last_idx,
batch_time=batch_time_m, loss=losses_m,
top1=prec1_m, top5=prec5_m))
metrics = OrderedDict([('loss', losses_m.avg), ('prec1', prec1_m.avg), ('prec5', prec5_m.avg)])
return metrics
try:
# import pdb;pdb.set_trace()
for epoch in range(start_epoch, num_epochs):
loader_len=len(loader_train)
if args.distributed:
loader_train.sampler.set_epoch(epoch)
# import pdb; pdb.set_trace()
if args.KD_train:
train_metrics = train_epoch(
epoch, model, loader_train, optimizer, train_loss_fn, args,
lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir,
use_amp=use_amp, model_ema=model_ema, teacher_model = teacher_model)
else:
para_loader = dp.ParallelLoader(loader_train, [device])
train_metrics = train_epoch(
epoch, model, para_loader.per_device_loader(device), optimizer, train_loss_fn, args,
lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir,
use_amp=use_amp, model_ema=model_ema, loader_len=loader_len)
# def __init__(self, model, bits_activations=8, bits_parameters=8, bits_accum=32,
# overrides=None, mode=LinearQuantMode.SYMMETRIC, clip_acts=ClipMode.NONE,
# per_channel_wts=False, model_activation_stats=None, fp16=False, clip_n_stds=None,
# scale_approx_mult_bits=None):
# import distiller
# import pdb; pdb.set_trace()
# quantizer = quantization.PostTrainLinearQuantizer.from_args(model, args)
# quantizer.prepare_model(distiller.get_dummy_input(input_shape=model.input_shape))
# quantizer = distiller.quantization.PostTrainLinearQuantizer(model, bits_activations=8, bits_parameters=8)
# quantizer.prepare_model()
# distiller.utils.assign_layer_fq_names(model)
# # msglogger.info("Generating quantization calibration stats based on {0} users".format(args.qe_calibration))
# collector = distiller.data_loggers.QuantCalibrationStatsCollector(model)
# with collector_context(collector):
# eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
# # Here call your model evaluation function, making sure to execute only
# # the portion of the dataset specified by the qe_calibration argument
# yaml_path = './dir/quantization_stats.yaml'
# collector.save(yaml_path)
loader_len_val = len(loader_eval)
para_loader = dp.ParallelLoader(loader_eval, [device])
eval_metrics = validate(model, para_loader.per_device_loader(device), validate_loss_fn, args, loader_len=loader_len_val)
if model_ema is not None and not args.model_ema_force_cpu:
ema_eval_metrics = validate(model_ema.ema, loader_eval, validate_loss_fn, args, log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
lr_scheduler.step(epoch, eval_metrics[eval_metric])
update_summary(
epoch, train_metrics, eval_metrics, os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
model, optimizer, args,
epoch=epoch + 1,
model_ema=model_ema,
metric=save_metric)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(best_metric, best_epoch))
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on 1 GPU.')
#torch.manual_seed(args.seed + args.rank)
# create model
config = get_efficientdet_config(args.model)
config.redundant_bias = args.redundant_bias # redundant conv + BN bias layers (True to match official models)
model = EfficientDet(config)
if args.initial_checkpoint:
load_checkpoint(model, args.initial_checkpoint)
config.num_classes = 5
model.class_net.predict.conv_pw = create_conv2d(config.fpn_channels,
9 * 5,
1,
padding=config.pad_type,
bias=True)
variance_scaling(model.class_net.predict.conv_pw.weight)
model.class_net.predict.conv_pw.bias.data.fill_(-math.log((1 - 0.01) /
0.01))
model = DetBenchTrain(model, config)
model.cuda()
print(model.model.class_net.predict.conv_pw)
# FIXME create model factory, pretrained zoo
# model = create_model(
# args.model,
# pretrained=args.pretrained,
# num_classes=args.num_classes,
# drop_rate=args.drop,
# drop_connect_rate=args.drop_connect, # DEPRECATED, use drop_path
# drop_path_rate=args.drop_path,
# drop_block_rate=args.drop_block,
# global_pool=args.gp,
# bn_tf=args.bn_tf,
# bn_momentum=args.bn_momentum,
# bn_eps=args.bn_eps,
# checkpoint_path=args.initial_checkpoint)
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model.parameters()])))
optimizer = create_optimizer(args, model)
use_amp = False
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed',
'on' if use_amp else 'off'))
# optionally resume from a checkpoint
resume_state = {}
resume_epoch = None
if args.resume:
resume_state, resume_epoch = resume_checkpoint(_unwrap_bench(model),
args.resume)
if resume_state and not args.no_resume_opt:
if 'optimizer' in resume_state:
if args.local_rank == 0:
logging.info('Restoring Optimizer state from checkpoint')
optimizer.load_state_dict(resume_state['optimizer'])
if use_amp and 'amp' in resume_state and 'load_state_dict' in amp.__dict__:
if args.local_rank == 0:
logging.info('Restoring NVIDIA AMP state from checkpoint')
amp.load_state_dict(resume_state['amp'])
del resume_state
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '')
#resume=args.resume) # FIXME bit of a mess with bench
if args.resume:
load_checkpoint(_unwrap_bench(model_ema),
args.resume,
use_ema=True)
if args.distributed:
if args.sync_bn:
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(
model)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.'
)
except Exception as e:
logging.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1'
)
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
model = DDP(model,
device_ids=[args.local_rank
]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
train_anno_set = 'train_small'
train_annotation_path = os.path.join(args.data, 'annotations_small',
f'train_annotations.json')
train_image_dir = train_anno_set
dataset_train = CocoDetection(os.path.join(args.data, train_image_dir),
train_annotation_path)
# FIXME cutmix/mixup worth investigating?
# collate_fn = None
# if args.prefetcher and args.mixup > 0:
# collate_fn = FastCollateMixup(args.mixup, args.smoothing, args.num_classes)
loader_train = create_loader(
dataset_train,
input_size=config.image_size,
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
#re_prob=args.reprob, # FIXME add back various augmentations
#re_mode=args.remode,
#re_count=args.recount,
#re_split=args.resplit,
#color_jitter=args.color_jitter,
#auto_augment=args.aa,
interpolation=args.train_interpolation,
#mean=data_config['mean'],
#std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
#collate_fn=collate_fn,
pin_mem=args.pin_mem,
)
#train_anno_set = 'valid_small'
#train_annotation_path = os.path.join(args.data, 'annotations_small', f'valid_annotations.json')
#train_image_dir = train_anno_set
dataset_eval = CocoDetection(os.path.join(args.data, train_image_dir),
train_annotation_path)
loader_eval = create_loader(
dataset_eval,
input_size=config.image_size,
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=args.interpolation,
#mean=data_config['mean'],
#std=data_config['std'],
num_workers=args.workers,
#distributed=args.distributed,
pin_mem=args.pin_mem,
)
evaluator = COCOEvaluator(dataset_eval.coco, distributed=args.distributed)
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join(
[datetime.now().strftime("%Y%m%d-%H%M%S"), args.model])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir,
decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
logging.info(
"Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
eval_metrics = validate(model, loader_eval, args, evaluator)
if model_ema is not None and not args.model_ema_force_cpu:
if args.distributed and args.dist_bn in ('broadcast',
'reduce'):
distribute_bn(model_ema, args.world_size,
args.dist_bn == 'reduce')
ema_eval_metrics = validate(model_ema.ema,
loader_eval,
args,
log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
if saver is not None:
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
_unwrap_bench(model),
optimizer,
args,
epoch=epoch,
model_ema=_unwrap_bench(model_ema),
metric=save_metric,
use_amp=use_amp)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
def train_imagenet():
torch.manual_seed(42)
device = xm.xla_device()
# model = get_model_property('model_fn')().to(device)
model = create_model(
FLAGS.model,
pretrained=FLAGS.pretrained,
num_classes=FLAGS.num_classes,
drop_rate=FLAGS.drop,
global_pool=FLAGS.gp,
bn_tf=FLAGS.bn_tf,
bn_momentum=FLAGS.bn_momentum,
bn_eps=FLAGS.bn_eps,
drop_connect_rate=0.2,
checkpoint_path=FLAGS.initial_checkpoint,
args = FLAGS).to(device)
model_ema=None
if FLAGS.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
# import pdb; pdb.set_trace()
model_e = create_model(
FLAGS.model,
pretrained=FLAGS.pretrained,
num_classes=FLAGS.num_classes,
drop_rate=FLAGS.drop,
global_pool=FLAGS.gp,
bn_tf=FLAGS.bn_tf,
bn_momentum=FLAGS.bn_momentum,
bn_eps=FLAGS.bn_eps,
drop_connect_rate=0.2,
checkpoint_path=FLAGS.initial_checkpoint,
args = FLAGS).to(device)
model_ema = ModelEma(
model_e,
decay=FLAGS.model_ema_decay,
device='cpu' if FLAGS.model_ema_force_cpu else '',
resume=FLAGS.resume)
print('==> Preparing data..')
img_dim = 224
if FLAGS.fake_data:
train_dataset_len = 1200000 # Roughly the size of Imagenet dataset.
train_loader = xu.SampleGenerator(
data=(torch.zeros(FLAGS.batch_size, 3, img_dim, img_dim),
torch.zeros(FLAGS.batch_size, dtype=torch.int64)),
sample_count=train_dataset_len // FLAGS.batch_size //
xm.xrt_world_size())
test_loader = xu.SampleGenerator(
data=(torch.zeros(FLAGS.batch_size, 3, img_dim, img_dim),
torch.zeros(FLAGS.batch_size, dtype=torch.int64)),
sample_count=50000 // FLAGS.batch_size // xm.xrt_world_size())
# else:
# normalize = transforms.Normalize(
# mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
# train_dataset = torchvision.datasets.ImageFolder(
# os.path.join(FLAGS.data, 'train'),
# transforms.Compose([
# transforms.RandomResizedCrop(img_dim),
# transforms.RandomHorizontalFlip(),
# transforms.ToTensor(),
# normalize,
# ]))
# train_dataset_len = len(train_dataset.imgs)
# resize_dim = max(img_dim, 256)
# test_dataset = torchvision.datasets.ImageFolder(
# os.path.join(FLAGS.data, 'val'),
# # Matches Torchvision's eval transforms except Torchvision uses size
# # 256 resize for all models both here and in the train loader. Their
# # version crashes during training on 299x299 images, e.g. inception.
# transforms.Compose([
# transforms.Resize(resize_dim),
# transforms.CenterCrop(img_dim),
# transforms.ToTensor(),
# normalize,
# ]))
# train_sampler = None
# if xm.xrt_world_size() > 1:
# train_sampler = torch.utils.data.distributed.DistributedSampler(
# train_dataset,
# num_replicas=xm.xrt_world_size(),
# rank=xm.get_ordinal(),
# shuffle=True)
# train_loader = torch.utils.data.DataLoader(
# train_dataset,
# batch_size=FLAGS.batch_size,
# sampler=train_sampler,
# shuffle=False if train_sampler else True,
# num_workers=FLAGS.workers)
# test_loader = torch.utils.data.DataLoader(
# test_dataset,
# batch_size=FLAGS.batch_size,
# shuffle=False,
# num_workers=FLAGS.workers)
else:
train_dir = os.path.join(FLAGS.data, 'train')
data_config = resolve_data_config(model, FLAGS, verbose=FLAGS.local_rank == 0)
dataset_train = Dataset(train_dir)
collate_fn = None
if not FLAGS.no_prefetcher and FLAGS.mixup > 0:
collate_fn = FastCollateMixup(FLAGS.mixup, FLAGS.smoothing, FLAGS.num_classes)
train_loader = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=FLAGS.batch_size,
is_training=True,
use_prefetcher=not FLAGS.no_prefetcher,
rand_erase_prob=FLAGS.reprob,
rand_erase_mode=FLAGS.remode,
interpolation='bicubic', # FIXME cleanly resolve this? data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=FLAGS.workers,
distributed=FLAGS.distributed,
collate_fn=collate_fn,
use_auto_aug=FLAGS.auto_augment,
use_mixcut=FLAGS.mixcut,
)
eval_dir = os.path.join(FLAGS.data, 'val')
train_dataset_len = len(train_loader)
if not os.path.isdir(eval_dir):
logging.error('Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
test_loader = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size = FLAGS.batch_size,
is_training=False,
use_prefetcher=FLAGS.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=FLAGS.workers,
distributed=FLAGS.distributed,
)
writer = None
start_epoch = 0
if FLAGS.output and xm.is_master_ordinal():
writer = SummaryWriter(log_dir=FLAGS.output)
optimizer = create_optimizer(flags, model)
lr_scheduler, num_epochs = create_scheduler(flags, optimizer)
if start_epoch > 0:
lr_scheduler.step(start_epoch)
# optimizer = optim.SGD(
# model.parameters(),
# lr=FLAGS.lr,
# momentum=FLAGS.momentum,
# weight_decay=5e-4)
num_training_steps_per_epoch = train_dataset_len // (
FLAGS.batch_size * xm.xrt_world_size())
lr_scheduler = schedulers.wrap_optimizer_with_scheduler(
optimizer,
scheduler_type=getattr(FLAGS, 'lr_scheduler_type', None),
scheduler_divisor=getattr(FLAGS, 'lr_scheduler_divisor', None),
scheduler_divide_every_n_epochs=getattr(
FLAGS, 'lr_scheduler_divide_every_n_epochs', None),
num_steps_per_epoch=num_training_steps_per_epoch,
summary_writer=writer)
train_loss_fn = LabelSmoothingCrossEntropy(smoothing=flags.smoothing)
validate_loss_fn = nn.CrossEntropyLoss()
# loss_fn = nn.CrossEntropyLoss()
def train_loop_fn(loader):
tracker = xm.RateTracker()
model.train()
for x, (data, target) in loader:
optimizer.zero_grad()
output = model(data)
loss = train_loss_fn(output, target)
loss.backward()
xm.optimizer_step(optimizer)
tracker.add(FLAGS.batch_size)
if model_ema is not None:
model_ema.update(model)
if lr_scheduler:
lr_scheduler.step()
if x % FLAGS.log_steps == 0:
test_utils.print_training_update(device, x, loss.item(), tracker.rate(),
tracker.global_rate())
def test_loop_fn(loader):
total_samples = 0
correct = 0
model.eval()
for x, (data, target) in loader:
output = model(data)
pred = output.max(1, keepdim=True)[1]
correct += pred.eq(target.view_as(pred)).sum().item()
total_samples += data.size()[0]
accuracy = 100.0 * correct / total_samples
test_utils.print_test_update(device, accuracy)
return accuracy
def test_loop_fn_ema(loader):
total_samples = 0
correct = 0
model_ema.eval()
for x, (data, target) in loader:
output = model_ema(data)
pred = output.max(1, keepdim=True)[1]
correct += pred.eq(target.view_as(pred)).sum().item()
total_samples += data.size()[0]
accuracy = 100.0 * correct / total_samples
test_utils.print_test_update(device, accuracy)
return accuracy
accuracy = 0.0
for epoch in range(1, FLAGS.epochs + 1):
para_loader = dp.ParallelLoader(train_loader, [device])
train_loop_fn(para_loader.per_device_loader(device))
para_loader = dp.ParallelLoader(test_loader, [device])
accuracy = test_loop_fn(para_loader.per_device_loader(device))
print('Epoch: {}, Mean Accuracy: {:.2f}%'.format(epoch, accuracy))
if model_ema is not None:
accuracy = test_loop_fn_ema(para_loader.per_device_loader(device))
print('Epoch: {}, Mean Accuracy: {:.2f}%'.format(epoch, accuracy))
test_utils.add_scalar_to_summary(writer, 'Accuracy/test', accuracy, epoch)
if FLAGS.metrics_debug:
print(torch_xla._XLAC._xla_metrics_report())
return accuracy
def main(args):
utils.my_init_distributed_mode(args)
print(args)
# if args.distillation_type != 'none' and args.finetune and not args.eval:
# raise NotImplementedError("Finetuning with distillation not yet supported")
device = torch.device(args.device)
# fix the seed for reproducibility
seed = args.seed + utils.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
# random.seed(seed)
cudnn.benchmark = True
dataset_train, args.nb_classes = build_dataset(is_train=True, args=args)
dataset_val, _ = build_dataset(is_train=False, args=args)
# if True: # args.distributed:
# num_tasks = utils.get_world_size()
# global_rank = utils.get_rank()
# if args.repeated_aug:
# sampler_train = RASampler(
# dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True
# )
# else:
# sampler_train = torch.utils.data.DistributedSampler(
# dataset_train,
# # num_replicas=num_tasks,
# num_replicas=0,
# rank=global_rank, shuffle=True
# )
# if args.dist_eval:
# if len(dataset_val) % num_tasks != 0:
# print('Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. '
# 'This will slightly alter validation results as extra duplicate entries are added to achieve '
# 'equal num of samples per-process.')
# sampler_val = torch.utils.data.DistributedSampler(
# dataset_val,
# # num_replicas=num_tasks,
# num_replicas=0,
# rank=global_rank, shuffle=False)
# else:
# sampler_val = torch.utils.data.SequentialSampler(dataset_val)
# else:
# sampler_train = torch.utils.data.RandomSampler(dataset_train)
# sampler_val = torch.utils.data.SequentialSampler(dataset_val)
#
# data_loader_train = torch.utils.data.DataLoader(
# dataset_train, sampler=sampler_train,
# batch_size=args.batch_size,
# num_workers=args.num_workers,
# pin_memory=args.pin_mem,
# drop_last=True,
# )
#
# data_loader_val = torch.utils.data.DataLoader(
# dataset_val, sampler=sampler_val,
# batch_size=int(1.5 * args.batch_size),
# num_workers=args.num_workers,
# pin_memory=args.pin_mem,
# drop_last=False
# )
#
if args.distributed:
if args.cache_mode:
sampler_train = samplers.NodeDistributedSampler(dataset_train)
sampler_val = samplers.NodeDistributedSampler(dataset_val,
shuffle=False)
else:
sampler_train = samplers.DistributedSampler(dataset_train)
sampler_val = samplers.DistributedSampler(dataset_val,
shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
num_workers=args.num_workers,
pin_memory=True)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
num_workers=args.num_workers,
pin_memory=True)
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_fn = Mixup(mixup_alpha=args.mixup,
cutmix_alpha=args.cutmix,
cutmix_minmax=args.cutmix_minmax,
prob=args.mixup_prob,
switch_prob=args.mixup_switch_prob,
mode=args.mixup_mode,
label_smoothing=args.smoothing,
num_classes=args.nb_classes)
print(f"Creating model: {args.model}")
model = create_model(
args.model,
pretrained=False,
num_classes=args.nb_classes,
drop_rate=args.drop,
drop_path_rate=args.drop_path,
drop_block_rate=None,
)
model_without_ddp = model
# there are bugs
if args.finetune:
if args.finetune.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.finetune,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(args.finetune, map_location='cpu')
# if 'model' in checkpoint:
# model_without_ddp.load_state_dict(checkpoint['model'])
# else:
# model_without_ddp.load_state_dict(checkpoint)
checkpoint_model = checkpoint['model']
state_dict = model.state_dict()
for k in [
'head.weight', 'head.bias', 'head_dist.weight',
'head_dist.bias'
]:
if k in checkpoint_model and checkpoint_model[
k].shape != state_dict[k].shape:
print(f"Removing key {k} from pretrained checkpoint")
del checkpoint_model[k]
# # interpolate position embedding
# pos_embed_checkpoint = checkpoint_model['pos_embed']
# embedding_size = pos_embed_checkpoint.shape[-1]
# num_patches = model.patch_embed.num_patches
# num_extra_tokens = model.pos_embed.shape[-2] - num_patches
# # height (== width) for the checkpoint position embedding
# orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
# # height (== width) for the new position embedding
# new_size = int(num_patches ** 0.5)
# # class_token and dist_token are kept unchanged
# extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
# # only the position tokens are interpolated
# pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
# pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
# pos_tokens = torch.nn.functional.interpolate(
# pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
# pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
# new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
# checkpoint_model['pos_embed'] = new_pos_embed
_ = model.load_state_dict(checkpoint_model, strict=False)
model.to(device)
model_ema = None
# if args.model_ema:
# # Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
# model_ema = ModelEma(
# model,
# decay=args.model_ema_decay,
# device='cpu' if args.model_ema_force_cpu else '',
# resume='')
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size(
) / 512.0
args.lr = linear_scaled_lr
optimizer = create_optimizer(args, model_without_ddp)
loss_scaler = NativeScaler()
lr_scheduler, _ = create_scheduler(args, optimizer)
criterion = LabelSmoothingCrossEntropy()
if args.mixup > 0.:
# smoothing is handled with mixup label transform
criterion = SoftTargetCrossEntropy()
elif args.smoothing:
criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
else:
criterion = torch.nn.CrossEntropyLoss()
# teacher_model = None
# if args.distillation_type != 'none':
# assert args.teacher_path, 'need to specify teacher-path when using distillation'
# print(f"Creating teacher model: {args.teacher_model}")
# teacher_model = create_model(
# args.teacher_model,
# pretrained=False,
# num_classes=args.nb_classes,
# global_pool='avg',
# )
# if args.teacher_path.startswith('https'):
# checkpoint = torch.hub.load_state_dict_from_url(
# args.teacher_path, map_location='cpu', check_hash=True)
# else:
# checkpoint = torch.load(args.teacher_path, map_location='cpu')
# teacher_model.load_state_dict(checkpoint['model'])
# teacher_model.to(device)
# teacher_model.eval()
# wrap the criterion in our custom DistillationLoss, which
# just dispatches to the original criterion if args.distillation_type is 'none'
# criterion = DistillationLoss(
# criterion, teacher_model, args.distillation_type, args.distillation_alpha, args.distillation_tau
# )
criterion = DistillationLoss(criterion, None, 'none', 0, 0)
output_dir = Path(args.output_dir)
if args.resume:
if args.resume.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.resume,
map_location='cpu',
check_hash=True)
else:
if not os.path.exists(args.resume):
checkpoint = None
print('NOTICE:' + args.resume + ' does not exist!')
else:
checkpoint = torch.load(args.resume, map_location='cpu')
if checkpoint is not None:
if 'model' in checkpoint:
model_without_ddp.load_state_dict(checkpoint['model'])
else:
model_without_ddp.load_state_dict(checkpoint)
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
# if args.model_ema:
# utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema'])
if 'scaler' in checkpoint:
loss_scaler.load_state_dict(checkpoint['scaler'])
print('resume from' + args.resume)
if args.eval:
test_stats = evaluate(data_loader_val, model, device)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
return
print(f"Start training for {args.epochs} epochs")
start_time = time.time()
max_accuracy = 0.0
max_epoch_dp_warm_up = 100
if 'pvt_tiny' in args.model or 'pvt_small' in args.model:
max_epoch_dp_warm_up = 0
if args.start_epoch < max_epoch_dp_warm_up:
model_without_ddp.reset_drop_path(0.0)
for epoch in range(args.start_epoch, args.epochs):
if args.fp32_resume and epoch > args.start_epoch + 1:
args.fp32_resume = False
loss_scaler._scaler = torch.cuda.amp.GradScaler(
enabled=not args.fp32_resume)
if epoch == max_epoch_dp_warm_up:
model_without_ddp.reset_drop_path(args.drop_path)
if args.distributed:
# data_loader_train.sampler.set_epoch(epoch)
sampler_train.set_epoch(epoch)
train_stats = my_train_one_epoch(
model,
criterion,
data_loader_train,
optimizer,
device,
epoch,
loss_scaler,
args.clip_grad,
model_ema,
mixup_fn,
set_training_mode=args.finetune ==
'', # keep in eval mode during finetuning
fp32=args.fp32_resume)
lr_scheduler.step(epoch)
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
for checkpoint_path in checkpoint_paths:
utils.save_on_master(
{
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
# 'model_ema': get_state_dict(model_ema),
'scaler': loss_scaler.state_dict(),
'args': args,
},
checkpoint_path)
test_stats = evaluate(data_loader_val, model, device)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
max_accuracy = max(max_accuracy, test_stats["acc1"])
print(f'Max accuracy: {max_accuracy:.2f}%')
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()},
**{f'test_{k}': v
for k, v in test_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def main():
setup_default_logging()
args, args_text = _parse_args()
use_amp = True
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect, # DEPRECATED, use drop_path
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint)
# model = skip_v3(num_classes=args.num_classes)
optimizer = create_optimizer(args, model)
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
resume_state = {}
resume_epoch = None
if args.resume:
resume_state, resume_epoch = resume_checkpoint(model, args.resume)
for i, param_group in enumerate(optimizer.param_groups):
param_group['lr'] = resume_state['lr']
start_epoch = resume_epoch
if 'amp' in resume_state and 'load_state_dict' in amp.__dict__:
logging.info('Restoring NVIDIA AMP state from checkpoint')
amp.load_state_dict(resume_state['amp'])
del resume_state
model = nn.DataParallel(model).cuda()
print('model to data parallel')
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(
model,
decay=args.model_ema_decay,
device='',
resume=args.resume)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
# if start_epoch < args.warmup_epochs:
# update_lr(args=args, optimizer=optimizer, epoch=start_epoch, per_epoch_update=False) #TODO iterations correction
# else:
# update_lr(args=args, optimizer=optimizer, epoch=start_epoch, per_epoch_update=True)
num_epochs = args.epochs
logging.info('Scheduled epochs: {}'.format(num_epochs))
train_dir = args.data
if not os.path.exists(train_dir):
logging.error('Training folder does not exist at: {}'.format(train_dir))
exit(1)
# Data loading code
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
image_folder = lmdb_loader.ImageFolderLMDB
print( os.path.join(train_dir,'ILSVRC-train.lmdb'))
train_dataset = image_folder(
os.path.join(train_dir,'ILSVRC-train.lmdb'),
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
val_dataset = image_folder(
os.path.join(train_dir, 'ILSVRC-val.lmdb'),
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
]))
loader_train = torch.utils.data.DataLoader(
train_dataset, batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True, sampler=None)
loader_eval = torch.utils.data.DataLoader(
val_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True)
train_loss_fn = nn.CrossEntropyLoss(reduction='none')
validate_loss_fn = train_loss_fn
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
args.model,
str(224)
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir, decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
writer =SummaryWriter()
try:
for epoch in range(start_epoch, num_epochs):
eval_metrics, _ = validate(model, loader_eval, validate_loss_fn, args, epoch=epoch)
train_metrics = train_epoch(
epoch, model, loader_train, optimizer, train_loss_fn, args
, saver=saver, output_dir=output_dir,
use_amp=use_amp, model_ema=model_ema, summary_writer=writer)
eval_metrics, _ = validate(model, loader_eval, validate_loss_fn, args, epoch=epoch)
if model_ema is not None:
ema_eval_metrics, top1 = validate(
model_ema.ema, loader_eval, validate_loss_fn, args,epoch=epoch, log_suffix=' (EMA)', summary_writer=writer)
writer.add_scalar('accuracy top 1', top1, epoch)
# if lr_scheduler is not None and epoch >= args.warmup_epochs:
# # step LR for next epoch
# lr_scheduler.step(epoch + 1)
update_lr(args=args, optimizer=optimizer, epoch=epoch, per_epoch_update=True)
update_summary(
epoch, train_metrics, eval_metrics, ema_eval_metrics, os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
save_metric_ema = ema_eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
model, optimizer, args,
epoch=epoch, model_ema=model_ema, metric=save_metric, metric_ema=save_metric_ema, use_amp=use_amp)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(best_metric, best_epoch))
