def main():
#wandb.init(project=args.experiment, config=args)
setup_default_logging()
args, args_text = _parse_args()
if args.log_wandb:
if has_wandb:
wandb.init(project=args.experiment, config=args)
else:
_logger.warning(
"You've requested to log metrics to wandb but package not found. "
"Metrics not being logged to wandb, try `pip install wandb`")
args.prefetcher = not args.no_prefetcher
args.distributed = False
if torch.cuda.is_available():
device = torch.device("cuda")
_logger.info(f'GPU: {torch.cuda.get_device_name(0)}')
else:
device = torch.device("cpu")
args.world_size = 1
args.rank = 0 # global rank
assert args.rank >= 0
# resolve AMP arguments based on PyTorch / Apex availability
use_amp = None
random_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,
scriptable=args.torchscript,
checkpoint_path=args.initial_checkpoint)
if args.num_classes is None:
assert hasattr(
model, 'num_classes'
), 'Model must have `num_classes` attr if not set on cmd line/config.'
args.num_classes = model.num_classes # FIXME handle model default vs config num_classes more elegantly
if args.local_rank == 0:
_logger.info(
f'Model {safe_model_name(args.model)} created, param count:{sum([m.numel() for m in model.parameters()])}'
)
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
model = determine_layer(model, args.finetune)
# setup augmentation batch splits for contrastive loss or split bn
num_aug_splits = 0
model.to(device)
optimizer = create_optimizer_v2(model, **optimizer_kwargs(cfg=args))
# setup automatic mixed-precision (AMP) loss scaling and op casting
# amp_autocast = suppress # do nothing
loss_scaler = None
# setup learning rate schedule and starting epoch
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.local_rank == 0:
_logger.info('Scheduled epochs: {}'.format(num_epochs))
# create the train and eval datasets
# dataset_train = create_dataset(
# args.dataset,
# root=args.data_dir, split=args.train_split, is_training=True,
# batch_size=args.batch_size, repeats=args.epoch_repeats)
# dataset_eval = create_dataset(
# args.dataset, root=args.data_dir, split=args.val_split, is_training=False, batch_size=args.batch_size)
# # setup mixup / cutmix
# collate_fn = None
# mixup_fn = None
# create data loaders w/ augmentation pipeiine
if 'skin' in args.experiment:
_logger.info('Loading Dataset')
img_df = pd.read_csv(args.csv_path) # csv directory
img_names, labels = list(img_df['image_name']), list(
img_df['diagnosis'])
img_index = list(range(len(img_names)))
train_valid_index, _, train_valid_labels, _ = train_test_split(
img_index,
labels,
test_size=0.2,
shuffle=True,
stratify=labels,
random_state=args.seed)
kf = StratifiedKFold(n_splits=5, shuffle=True, random_state=args.seed)
for fold_index, (train_index, valid_index) in enumerate(
kf.split(train_valid_index, train_valid_labels)):
train_index = train_index
valid_index = valid_index
break
_logger.info(f'augmentation : {args.augment}')
train_df = img_df[img_df.index.isin(train_index)].reset_index(
drop=True)
train_dataset = SkinDataset(
data_dir=args.data_dir,
df=train_df,
transform=get_skin_transforms(augment=args.augment,
args=args)) # file directory
valid_df = img_df[img_df.index.isin(valid_index)].reset_index(
drop=True)
valid_dataset = SkinDataset(
data_dir=args.data_dir,
df=valid_df,
transform=get_skin_transforms(augment='none',
args=args)) # file directory
_logger.info('Load Sampler & Loader')
print(len(train_index), len(valid_index))
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
pin_memory=False)
valid_loader = DataLoader(valid_dataset,
batch_size=args.batch_size,
shuffle=True,
pin_memory=False)
elif 'lung' in args.experiment:
_logger.info('Loading Dataset')
img_df = pd.read_csv(args.csv_path) # csv directory
img_names, labels = list(img_df['image_link']), list(img_df['label'])
img_index = list(range(len(img_names)))
_logger.info(f'augmentation : {args.augment}')
train_df = img_df[img_df['tvt'] == 'train'].reset_index(drop=True)
train_dataset = LungDataset(
df=train_df,
transform=get_lung_transforms(augment=args.augment,
args=args)) # file directory
train_loader = DataLoader(train_dataset, batch_size=args.batch_size)
valid_df = img_df[img_df['tvt'] == 'valid'].reset_index(drop=True)
valid_dataset = LungDataset(
df=valid_df,
transform=get_lung_transforms(augment='none',
args=args)) # file directory
valid_loader = DataLoader(valid_dataset, batch_size=args.batch_size)
_logger.info('Load Sampler & Loader')
_logger.info(len(train_dataset), len(valid_dataset))
train_loader = DataLoader(train_dataset,
shuffle=True,
batch_size=args.batch_size)
valid_loader = DataLoader(valid_dataset,
shuffle=True,
batch_size=args.batch_size)
# setup loss function
if args.loss == 'focal':
train_loss_fn = FocalLoss().to(device)
else:
if args.smoothing:
_logger.info('default loss is LabelSmoothing')
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).to(device)
else:
train_loss_fn = nn.CrossEntropyLoss().to(device)
validate_loss_fn = nn.CrossEntropyLoss().to(device)
# setup checkpoint saver and eval metric tracking
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = None
early_stopping = EarlyStopping(patience=args.early_patience,
delta=args.early_value,
verbose=True)
if args.rank == 0:
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
safe_model_name(args.model),
str(data_config['input_size'][-1])
])
output_dir = get_outdir(
args.output if args.output else './output/train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(model=model,
optimizer=optimizer,
args=args,
checkpoint_dir=output_dir,
recovery_dir=output_dir,
decreasing=decreasing,
max_history=args.checkpoint_hist)
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):
train_metrics = train_one_epoch(epoch,
model,
train_loader,
optimizer,
train_loss_fn,
device,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir)
eval_metrics = validate(model, valid_loader, validate_loss_fn,
device, args)
if lr_scheduler is not None:
# step LR for next epoch
if args.sched == 'step':
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
else:
lr_scheduler.step(epoch)
early_stopping(eval_metric, eval_metrics[eval_metric], model)
if early_stopping.early_stop:
_logger.info('Early Stop')
break
if output_dir is not None:
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None,
log_wandb=args.log_wandb and has_wandb)
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():
args, cfg = parse_config_args('child net training')
# resolve logging
output_dir = os.path.join(
cfg.SAVE_PATH, "{}-{}".format(datetime.date.today().strftime('%m%d'),
cfg.MODEL))
if args.local_rank == 0:
logger = get_logger(os.path.join(output_dir, 'retrain.log'))
writer = SummaryWriter(os.path.join(output_dir, 'runs'))
else:
writer, logger = None, None
# retrain model selection
if cfg.NET.SELECTION == 481:
arch_list = [[0], [3, 4, 3, 1], [3, 2, 3, 0], [3, 3, 3, 1, 1],
[3, 3, 3, 3], [3, 3, 3, 3], [0]]
cfg.DATASET.IMAGE_SIZE = 224
elif cfg.NET.SELECTION == 43:
arch_list = [[0], [3], [3, 1], [3, 1], [3, 3, 3], [3, 3], [0]]
cfg.DATASET.IMAGE_SIZE = 96
elif cfg.NET.SELECTION == 14:
arch_list = [[0], [3], [3, 3], [3, 3], [3], [3], [0]]
cfg.DATASET.IMAGE_SIZE = 64
elif cfg.NET.SELECTION == 114:
arch_list = [[0], [3], [3, 3], [3, 3], [3, 3, 3], [3, 3], [0]]
cfg.DATASET.IMAGE_SIZE = 160
elif cfg.NET.SELECTION == 287:
arch_list = [[0], [3], [3, 3], [3, 1, 3], [3, 3, 3, 3], [3, 3, 3], [0]]
cfg.DATASET.IMAGE_SIZE = 224
elif cfg.NET.SELECTION == 604:
arch_list = [[0], [3, 3, 2, 3, 3], [3, 2, 3, 2, 3], [3, 2, 3, 2, 3],
[3, 3, 2, 2, 3, 3], [3, 3, 2, 3, 3, 3], [0]]
cfg.DATASET.IMAGE_SIZE = 224
else:
raise ValueError("Model Retrain Selection is not Supported!")
# define childnet architecture from arch_list
stem = ['ds_r1_k3_s1_e1_c16_se0.25', 'cn_r1_k1_s1_c320_se0.25']
choice_block_pool = [
'ir_r1_k3_s2_e4_c24_se0.25', 'ir_r1_k5_s2_e4_c40_se0.25',
'ir_r1_k3_s2_e6_c80_se0.25', 'ir_r1_k3_s1_e6_c96_se0.25',
'ir_r1_k5_s2_e6_c192_se0.25'
]
arch_def = [[stem[0]]] + [[
choice_block_pool[idx]
for repeat_times in range(len(arch_list[idx + 1]))
] for idx in range(len(choice_block_pool))] + [[stem[1]]]
# generate childnet
model = gen_childnet(arch_list,
arch_def,
num_classes=cfg.DATASET.NUM_CLASSES,
drop_rate=cfg.NET.DROPOUT_RATE,
global_pool=cfg.NET.GP)
# initialize training parameters
eval_metric = cfg.EVAL_METRICS
best_metric, best_epoch, saver = None, None, None
# initialize distributed parameters
distributed = cfg.NUM_GPU > 1
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
if args.local_rank == 0:
logger.info('Training on Process {} with {} GPUs.'.format(
args.local_rank, cfg.NUM_GPU))
# fix random seeds
torch.manual_seed(cfg.SEED)
torch.cuda.manual_seed_all(cfg.SEED)
np.random.seed(cfg.SEED)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# get parameters and FLOPs of model
if args.local_rank == 0:
macs, params = get_model_flops_params(
model,
input_size=(1, 3, cfg.DATASET.IMAGE_SIZE, cfg.DATASET.IMAGE_SIZE))
logger.info('[Model-{}] Flops: {} Params: {}'.format(
cfg.NET.SELECTION, macs, params))
# create optimizer
model = model.cuda()
optimizer = create_optimizer(cfg, model)
# optionally resume from a checkpoint
resume_state, resume_epoch = {}, None
if cfg.AUTO_RESUME:
resume_state, resume_epoch = resume_checkpoint(model, cfg.RESUME_PATH)
optimizer.load_state_dict(resume_state['optimizer'])
del resume_state
model_ema = None
if cfg.NET.EMA.USE:
model_ema = ModelEma(
model,
decay=cfg.NET.EMA.DECAY,
device='cpu' if cfg.NET.EMA.FORCE_CPU else '',
resume=cfg.RESUME_PATH if cfg.AUTO_RESUME else None)
if distributed:
if cfg.BATCHNORM.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:
logger.info(
'Converted model to use Synchronized BatchNorm.')
except Exception as e:
if args.local_rank == 0:
logger.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1 with exception {}'
.format(e))
if HAS_APEX:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logger.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
# can use device str in Torch >= 1.1
model = DDP(model, device_ids=[args.local_rank])
# imagenet train dataset
train_dir = os.path.join(cfg.DATA_DIR, 'train')
if not os.path.exists(train_dir) and args.local_rank == 0:
logger.error('Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
loader_train = create_loader(dataset_train,
input_size=(3, cfg.DATASET.IMAGE_SIZE,
cfg.DATASET.IMAGE_SIZE),
batch_size=cfg.DATASET.BATCH_SIZE,
is_training=True,
color_jitter=cfg.AUGMENTATION.COLOR_JITTER,
auto_augment=cfg.AUGMENTATION.AA,
num_aug_splits=0,
crop_pct=DEFAULT_CROP_PCT,
mean=IMAGENET_DEFAULT_MEAN,
std=IMAGENET_DEFAULT_STD,
num_workers=cfg.WORKERS,
distributed=distributed,
collate_fn=None,
pin_memory=cfg.DATASET.PIN_MEM,
interpolation='random',
re_mode=cfg.AUGMENTATION.RE_MODE,
re_prob=cfg.AUGMENTATION.RE_PROB)
# imagenet validation dataset
eval_dir = os.path.join(cfg.DATA_DIR, 'val')
if not os.path.exists(eval_dir) and args.local_rank == 0:
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=(3, cfg.DATASET.IMAGE_SIZE, cfg.DATASET.IMAGE_SIZE),
batch_size=cfg.DATASET.VAL_BATCH_MUL * cfg.DATASET.BATCH_SIZE,
is_training=False,
interpolation='bicubic',
crop_pct=DEFAULT_CROP_PCT,
mean=IMAGENET_DEFAULT_MEAN,
std=IMAGENET_DEFAULT_STD,
num_workers=cfg.WORKERS,
distributed=distributed,
pin_memory=cfg.DATASET.PIN_MEM)
# whether to use label smoothing
if cfg.AUGMENTATION.SMOOTHING > 0.:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=cfg.AUGMENTATION.SMOOTHING).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
# create learning rate scheduler
lr_scheduler, num_epochs = create_scheduler(cfg, optimizer)
start_epoch = resume_epoch if resume_epoch is not None else 0
if start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logger.info('Scheduled epochs: {}'.format(num_epochs))
try:
best_record, best_ep = 0, 0
for epoch in range(start_epoch, num_epochs):
if distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
cfg,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
model_ema=model_ema,
logger=logger,
writer=writer,
local_rank=args.local_rank)
eval_metrics = validate(epoch,
model,
loader_eval,
validate_loss_fn,
cfg,
logger=logger,
writer=writer,
local_rank=args.local_rank)
if model_ema is not None and not cfg.NET.EMA.FORCE_CPU:
ema_eval_metrics = validate(epoch,
model_ema.ema,
loader_eval,
validate_loss_fn,
cfg,
log_suffix='_EMA',
logger=logger,
writer=writer,
local_rank=args.local_rank)
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
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,
cfg,
epoch=epoch,
model_ema=model_ema,
metric=save_metric)
if best_record < eval_metrics[eval_metric]:
best_record = eval_metrics[eval_metric]
best_ep = epoch
if args.local_rank == 0:
logger.info('*** Best metric: {0} (epoch {1})'.format(
best_record, best_ep))
except KeyboardInterrupt:
pass
if best_metric is not None:
logger.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,
pretrained=False,
num_classes=args.nb_classes,
drop_rate=args.drop,
drop_path_rate=args.drop_path,
drop_block_rate=None,
)
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]
# 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
)
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, 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)
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()
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(args):
utils.init_distributed_mode(args)
print(args)
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)
else:
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(
1.5 * args.batch_size),
shuffle=False,
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,
pretrained=False,
num_classes=args.nb_classes,
drop_rate=args.drop,
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
)
# TODO: finetuning
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)
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 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 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("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)
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),
'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():
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(config, writer):
dataset_train, dataset_val, data_loader_train, data_loader_val, mixup_fn = build_loader(
config)
logger.info(f"Creating model:{config.MODEL.TYPE}/{config.MODEL.NAME}")
model = build_model(config)
model.cuda()
logger.info(str(model))
optimizer = build_optimizer(config, model)
if config.AMP_OPT_LEVEL != "O0":
model, optimizer = amp.initialize(model,
optimizer,
opt_level=config.AMP_OPT_LEVEL)
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
logger.info(f"number of params: {n_parameters}")
if hasattr(model, 'flops'):
flops = model.flops()
logger.info(f"number of GFLOPs: {flops / 1e9}")
lr_scheduler = build_scheduler(config, optimizer, len(data_loader_train))
if config.AUG.MIXUP > 0.:
# smoothing is handled with mixup label transform
criterion = SoftTargetCrossEntropy()
elif config.MODEL.LABEL_SMOOTHING > 0.:
criterion = LabelSmoothingCrossEntropy(
smoothing=config.MODEL.LABEL_SMOOTHING)
else:
criterion = torch.nn.CrossEntropyLoss()
max_accuracy = 0.0
if config.TRAIN.AUTO_RESUME:
resume_file = auto_resume_helper(config.OUTPUT)
if resume_file:
if config.MODEL.RESUME:
logger.warning(
f"auto-resume changing resume file from {config.MODEL.RESUME} to {resume_file}"
)
config.defrost()
config.MODEL.RESUME = resume_file
config.freeze()
logger.info(f'auto resuming from {resume_file}')
else:
logger.info(
f'no checkpoint found in {config.OUTPUT}, ignoring auto resume'
)
if config.MODEL.RESUME:
max_accuracy = load_checkpoint(config, model, optimizer, lr_scheduler,
logger)
acc1, acc5, loss = validate(config, data_loader_val, model)
logger.info(
f"Accuracy of the network on the {len(dataset_val)} test images: {acc1:.1f}%"
)
if config.EVAL_MODE:
return
if config.THROUGHPUT_MODE:
throughput(data_loader_val, model, logger)
return
logger.info("Start training")
start_time = time.time()
for epoch in range(config.TRAIN.START_EPOCH, config.TRAIN.EPOCHS):
train_one_epoch(config, model, criterion, data_loader_train, optimizer,
epoch, mixup_fn, lr_scheduler, writer)
if (epoch % config.SAVE_FREQ == 0
or epoch == (config.TRAIN.EPOCHS - 1)):
save_checkpoint(config, epoch, model, max_accuracy, optimizer,
lr_scheduler, logger)
acc1, acc5, loss = validate(config, data_loader_val, model)
writer.add_scalar("validate/acc1",
scalar_value=acc1,
global_step=epoch)
writer.add_scalar("validate/acc5",
scalar_value=acc5,
global_step=epoch)
writer.add_scalar("validate/loss",
scalar_value=loss,
global_step=epoch)
logger.info(
f"Accuracy of the network on the {len(dataset_val)} test images: {acc1:.1f}%"
)
max_accuracy = max(max_accuracy, acc1)
logger.info(f'Max accuracy: {max_accuracy:.2f}%')
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
logger.info('Training time {}'.format(total_time_str))
def main():
args, cfg = parse_config_args('super net training')
# resolve logging
output_dir = os.path.join(
cfg.SAVE_PATH, "{}-{}".format(datetime.date.today().strftime('%m%d'),
cfg.MODEL))
if args.local_rank == 0:
logger = get_logger(os.path.join(output_dir, "train.log"))
else:
logger = None
# initialize distributed parameters
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
if args.local_rank == 0:
logger.info('Training on Process %d with %d GPUs.', args.local_rank,
cfg.NUM_GPU)
# fix random seeds
torch.manual_seed(cfg.SEED)
torch.cuda.manual_seed_all(cfg.SEED)
np.random.seed(cfg.SEED)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# generate supernet
model, sta_num, resolution = gen_supernet(
flops_minimum=cfg.SUPERNET.FLOPS_MINIMUM,
flops_maximum=cfg.SUPERNET.FLOPS_MAXIMUM,
num_classes=cfg.DATASET.NUM_CLASSES,
drop_rate=cfg.NET.DROPOUT_RATE,
global_pool=cfg.NET.GP,
resunit=cfg.SUPERNET.RESUNIT,
dil_conv=cfg.SUPERNET.DIL_CONV,
slice=cfg.SUPERNET.SLICE,
verbose=cfg.VERBOSE,
logger=logger)
# initialize meta matching networks
MetaMN = MetaMatchingNetwork(cfg)
# number of choice blocks in supernet
choice_num = len(model.blocks[1][0])
if args.local_rank == 0:
logger.info('Supernet created, param count: %d',
(sum([m.numel() for m in model.parameters()])))
logger.info('resolution: %d', (resolution))
logger.info('choice number: %d', (choice_num))
#initialize prioritized board
prioritized_board = PrioritizedBoard(cfg,
CHOICE_NUM=choice_num,
sta_num=sta_num)
# initialize flops look-up table
model_est = FlopsEst(model)
# optionally resume from a checkpoint
optimizer_state = None
resume_epoch = None
if cfg.AUTO_RESUME:
optimizer_state, resume_epoch = resume_checkpoint(
model, cfg.RESUME_PATH)
# create optimizer and resume from checkpoint
optimizer = create_optimizer_supernet(cfg, model, USE_APEX)
if optimizer_state is not None:
optimizer.load_state_dict(optimizer_state['optimizer'])
model = model.cuda()
# convert model to distributed mode
if cfg.BATCHNORM.SYNC_BN:
try:
if USE_APEX:
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.')
except Exception as exception:
logger.info(
'Failed to enable Synchronized BatchNorm. '
'Install Apex or Torch >= 1.1 with Exception %s', exception)
if USE_APEX:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logger.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
# can use device str in Torch >= 1.1
model = DDP(model, device_ids=[args.local_rank])
# create learning rate scheduler
lr_scheduler, num_epochs = create_supernet_scheduler(cfg, optimizer)
start_epoch = resume_epoch if resume_epoch is not None else 0
if start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logger.info('Scheduled epochs: %d', num_epochs)
# imagenet train dataset
train_dir = os.path.join(cfg.DATA_DIR, 'train')
if not os.path.exists(train_dir):
logger.info('Training folder does not exist at: %s', train_dir)
sys.exit()
dataset_train = Dataset(train_dir)
loader_train = create_loader(dataset_train,
input_size=(3, cfg.DATASET.IMAGE_SIZE,
cfg.DATASET.IMAGE_SIZE),
batch_size=cfg.DATASET.BATCH_SIZE,
is_training=True,
use_prefetcher=True,
re_prob=cfg.AUGMENTATION.RE_PROB,
re_mode=cfg.AUGMENTATION.RE_MODE,
color_jitter=cfg.AUGMENTATION.COLOR_JITTER,
interpolation='random',
num_workers=cfg.WORKERS,
distributed=True,
collate_fn=None,
crop_pct=DEFAULT_CROP_PCT,
mean=IMAGENET_DEFAULT_MEAN,
std=IMAGENET_DEFAULT_STD)
# imagenet validation dataset
eval_dir = os.path.join(cfg.DATA_DIR, 'val')
if not os.path.isdir(eval_dir):
logger.info('Validation folder does not exist at: %s', eval_dir)
sys.exit()
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(dataset_eval,
input_size=(3, cfg.DATASET.IMAGE_SIZE,
cfg.DATASET.IMAGE_SIZE),
batch_size=4 * cfg.DATASET.BATCH_SIZE,
is_training=False,
use_prefetcher=True,
num_workers=cfg.WORKERS,
distributed=True,
crop_pct=DEFAULT_CROP_PCT,
mean=IMAGENET_DEFAULT_MEAN,
std=IMAGENET_DEFAULT_STD,
interpolation=cfg.DATASET.INTERPOLATION)
# whether to use label smoothing
if cfg.AUGMENTATION.SMOOTHING > 0.:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=cfg.AUGMENTATION.SMOOTHING).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
# initialize training parameters
eval_metric = cfg.EVAL_METRICS
best_metric, best_epoch, saver, best_children_pool = None, None, None, []
if args.local_rank == 0:
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir,
decreasing=decreasing)
# training scheme
try:
for epoch in range(start_epoch, num_epochs):
loader_train.sampler.set_epoch(epoch)
# train one epoch
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
prioritized_board,
MetaMN,
cfg,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
logger=logger,
est=model_est,
local_rank=args.local_rank)
# evaluate one epoch
eval_metrics = validate(model,
loader_eval,
validate_loss_fn,
prioritized_board,
cfg,
local_rank=args.local_rank,
logger=logger)
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, cfg, epoch=epoch, metric=save_metric)
except KeyboardInterrupt:
pass
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))
def main(args, ds_init):
utils.init_distributed_mode(args)
if ds_init is not None:
utils.create_ds_config(args)
print(args)
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)
if args.disable_eval_during_finetuning:
dataset_val = None
else:
dataset_val, _ = build_dataset(is_train=False, args=args)
if True: # args.distributed:
num_tasks = utils.get_world_size()
global_rank = utils.get_rank()
sampler_train = torch.utils.data.DistributedSampler(
dataset_train,
num_replicas=num_tasks,
rank=global_rank,
shuffle=True)
print("Sampler_train = %s" % str(sampler_train))
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)
if global_rank == 0 and args.log_dir is not None:
os.makedirs(args.log_dir, exist_ok=True)
log_writer = utils.TensorboardLogger(log_dir=args.log_dir)
else:
log_writer = None
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,
)
if dataset_val is not None:
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)
else:
data_loader_val = None
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
print("Mixup is activated!")
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)
model = create_model(
args.model,
pretrained=False,
num_classes=args.nb_classes,
drop_rate=args.drop,
drop_path_rate=args.drop_path,
attn_drop_rate=args.attn_drop_rate,
drop_block_rate=None,
use_mean_pooling=args.use_mean_pooling,
init_scale=args.init_scale,
use_rel_pos_bias=args.rel_pos_bias,
use_abs_pos_emb=args.abs_pos_emb,
init_values=args.layer_scale_init_value,
)
patch_size = model.patch_embed.patch_size
print("Patch size = %s" % str(patch_size))
args.window_size = (args.input_size // patch_size[0],
args.input_size // patch_size[1])
args.patch_size = patch_size
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')
print("Load ckpt from %s" % args.finetune)
checkpoint_model = None
for model_key in args.model_key.split('|'):
if model_key in checkpoint:
checkpoint_model = checkpoint[model_key]
print("Load state_dict by model_key = %s" % model_key)
break
if checkpoint_model is None:
checkpoint_model = checkpoint
state_dict = model.state_dict()
for k in ['head.weight', '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]
if model.use_rel_pos_bias and "rel_pos_bias.relative_position_bias_table" in checkpoint_model:
print(
"Expand the shared relative position embedding to each transformer block. "
)
num_layers = model.get_num_layers()
rel_pos_bias = checkpoint_model[
"rel_pos_bias.relative_position_bias_table"]
for i in range(num_layers):
checkpoint_model["blocks.%d.attn.relative_position_bias_table"
% i] = rel_pos_bias.clone()
checkpoint_model.pop("rel_pos_bias.relative_position_bias_table")
all_keys = list(checkpoint_model.keys())
for key in all_keys:
if "relative_position_index" in key:
checkpoint_model.pop(key)
if "relative_position_bias_table" in key:
rel_pos_bias = checkpoint_model[key]
src_num_pos, num_attn_heads = rel_pos_bias.size()
dst_num_pos, _ = model.state_dict()[key].size()
dst_patch_shape = model.patch_embed.patch_shape
if dst_patch_shape[0] != dst_patch_shape[1]:
raise NotImplementedError()
num_extra_tokens = dst_num_pos - (
dst_patch_shape[0] * 2 - 1) * (dst_patch_shape[1] * 2 - 1)
src_size = int((src_num_pos - num_extra_tokens)**0.5)
dst_size = int((dst_num_pos - num_extra_tokens)**0.5)
if src_size != dst_size:
print("Position interpolate for %s from %dx%d to %dx%d" %
(key, src_size, src_size, dst_size, dst_size))
extra_tokens = rel_pos_bias[-num_extra_tokens:, :]
rel_pos_bias = rel_pos_bias[:-num_extra_tokens, :]
def geometric_progression(a, r, n):
return a * (1.0 - r**n) / (1.0 - r)
left, right = 1.01, 1.5
while right - left > 1e-6:
q = (left + right) / 2.0
gp = geometric_progression(1, q, src_size // 2)
if gp > dst_size // 2:
right = q
else:
left = q
# if q > 1.090307:
# q = 1.090307
dis = []
cur = 1
for i in range(src_size // 2):
dis.append(cur)
cur += q**(i + 1)
r_ids = [-_ for _ in reversed(dis)]
x = r_ids + [0] + dis
y = r_ids + [0] + dis
t = dst_size // 2.0
dx = np.arange(-t, t + 0.1, 1.0)
dy = np.arange(-t, t + 0.1, 1.0)
print("Original positions = %s" % str(x))
print("Target positions = %s" % str(dx))
all_rel_pos_bias = []
for i in range(num_attn_heads):
z = rel_pos_bias[:, i].view(src_size,
src_size).float().numpy()
f = interpolate.interp2d(x, y, z, kind='cubic')
all_rel_pos_bias.append(
torch.Tensor(f(dx, dy)).contiguous().view(
-1, 1).to(rel_pos_bias.device))
rel_pos_bias = torch.cat(all_rel_pos_bias, dim=-1)
new_rel_pos_bias = torch.cat((rel_pos_bias, extra_tokens),
dim=0)
checkpoint_model[key] = new_rel_pos_bias
# interpolate position embedding
if 'pos_embed' in checkpoint_model:
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
if orig_size != new_size:
print("Position interpolate from %dx%d to %dx%d" %
(orig_size, orig_size, new_size, new_size))
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
utils.load_state_dict(model,
checkpoint_model,
prefix=args.model_prefix)
# 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='')
print("Using EMA with decay = %.8f" % args.model_ema_decay)
model_without_ddp = model
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print("Model = %s" % str(model_without_ddp))
print('number of params:', n_parameters)
total_batch_size = args.batch_size * args.update_freq * utils.get_world_size(
)
num_training_steps_per_epoch = len(dataset_train) // total_batch_size
print("LR = %.8f" % args.lr)
print("Batch size = %d" % total_batch_size)
print("Update frequent = %d" % args.update_freq)
print("Number of training examples = %d" % len(dataset_train))
print("Number of training training per epoch = %d" %
num_training_steps_per_epoch)
num_layers = model_without_ddp.get_num_layers()
if args.layer_decay < 1.0:
assigner = LayerDecayValueAssigner(
list(args.layer_decay**(num_layers + 1 - i)
for i in range(num_layers + 2)))
else:
assigner = None
if assigner is not None:
print("Assigned values = %s" % str(assigner.values))
skip_weight_decay_list = model.no_weight_decay()
if args.disable_weight_decay_on_rel_pos_bias:
for i in range(num_layers):
skip_weight_decay_list.add(
"blocks.%d.attn.relative_position_bias_table" % i)
if args.enable_deepspeed:
loss_scaler = None
optimizer_params = get_parameter_groups(
model, args.weight_decay, skip_weight_decay_list,
assigner.get_layer_id if assigner is not None else None,
assigner.get_scale if assigner is not None else None)
model, optimizer, _, _ = ds_init(
args=args,
model=model,
model_parameters=optimizer_params,
dist_init_required=not args.distributed,
)
print("model.gradient_accumulation_steps() = %d" %
model.gradient_accumulation_steps())
assert model.gradient_accumulation_steps() == args.update_freq
else:
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu], find_unused_parameters=True)
model_without_ddp = model.module
optimizer = create_optimizer(args,
model_without_ddp,
skip_list=skip_weight_decay_list,
get_num_layer=assigner.get_layer_id
if assigner is not None else None,
get_layer_scale=assigner.get_scale
if assigner is not None else None)
loss_scaler = NativeScaler()
print("Use step level LR scheduler!")
lr_schedule_values = utils.cosine_scheduler(
args.lr,
args.min_lr,
args.epochs,
num_training_steps_per_epoch,
warmup_epochs=args.warmup_epochs,
warmup_steps=args.warmup_steps,
)
if args.weight_decay_end is None:
args.weight_decay_end = args.weight_decay
wd_schedule_values = utils.cosine_scheduler(args.weight_decay,
args.weight_decay_end,
args.epochs,
num_training_steps_per_epoch)
print("Max WD = %.7f, Min WD = %.7f" %
(max(wd_schedule_values), min(wd_schedule_values)))
if mixup_fn is not None:
# smoothing is handled with mixup label transform
criterion = SoftTargetCrossEntropy()
elif args.smoothing > 0.:
criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
else:
criterion = torch.nn.CrossEntropyLoss()
print("criterion = %s" % str(criterion))
utils.auto_load_model(args=args,
model=model,
model_without_ddp=model_without_ddp,
optimizer=optimizer,
loss_scaler=loss_scaler,
model_ema=model_ema)
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}%"
)
exit(0)
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 log_writer is not None:
log_writer.set_step(epoch * num_training_steps_per_epoch *
args.update_freq)
train_stats = train_one_epoch(
model,
criterion,
data_loader_train,
optimizer,
device,
epoch,
loss_scaler,
args.clip_grad,
model_ema,
mixup_fn,
log_writer=log_writer,
start_steps=epoch * num_training_steps_per_epoch,
lr_schedule_values=lr_schedule_values,
wd_schedule_values=wd_schedule_values,
num_training_steps_per_epoch=num_training_steps_per_epoch,
update_freq=args.update_freq,
)
if args.output_dir and args.save_ckpt:
if (epoch +
1) % args.save_ckpt_freq == 0 or epoch + 1 == args.epochs:
utils.save_model(args=args,
model=model,
model_without_ddp=model_without_ddp,
optimizer=optimizer,
loss_scaler=loss_scaler,
epoch=epoch,
model_ema=model_ema)
if data_loader_val is not None:
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}%"
)
if max_accuracy < test_stats["acc1"]:
max_accuracy = test_stats["acc1"]
if args.output_dir and args.save_ckpt:
utils.save_model(args=args,
model=model,
model_without_ddp=model_without_ddp,
optimizer=optimizer,
loss_scaler=loss_scaler,
epoch="best",
model_ema=model_ema)
print(f'Max accuracy: {max_accuracy:.2f}%')
if log_writer is not None:
log_writer.update(test_acc1=test_stats['acc1'],
head="perf",
step=epoch)
log_writer.update(test_acc5=test_stats['acc5'],
head="perf",
step=epoch)
log_writer.update(test_loss=test_stats['loss'],
head="perf",
step=epoch)
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()},
# **{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():
if log_writer is not None:
log_writer.flush()
with open(os.path.join(args.output_dir, "log.txt"),
mode="a",
encoding="utf-8") 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
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, cfg = parse_config_args('nni.cream.supernet')
# resolve logging
output_dir = os.path.join(
cfg.SAVE_PATH, "{}-{}".format(datetime.date.today().strftime('%m%d'),
cfg.MODEL))
if not os.path.exists(output_dir):
os.mkdir(output_dir)
if args.local_rank == 0:
logger = get_logger(os.path.join(output_dir, "train.log"))
else:
logger = None
# initialize distributed parameters
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
if args.local_rank == 0:
logger.info('Training on Process %d with %d GPUs.', args.local_rank,
cfg.NUM_GPU)
# fix random seeds
torch.manual_seed(cfg.SEED)
torch.cuda.manual_seed_all(cfg.SEED)
np.random.seed(cfg.SEED)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# generate supernet
model, sta_num, resolution = gen_supernet(
flops_minimum=cfg.SUPERNET.FLOPS_MINIMUM,
flops_maximum=cfg.SUPERNET.FLOPS_MAXIMUM,
num_classes=cfg.DATASET.NUM_CLASSES,
drop_rate=cfg.NET.DROPOUT_RATE,
global_pool=cfg.NET.GP,
resunit=cfg.SUPERNET.RESUNIT,
dil_conv=cfg.SUPERNET.DIL_CONV,
slice=cfg.SUPERNET.SLICE,
verbose=cfg.VERBOSE,
logger=logger)
# number of choice blocks in supernet
choice_num = len(model.blocks[7])
if args.local_rank == 0:
logger.info('Supernet created, param count: %d',
(sum([m.numel() for m in model.parameters()])))
logger.info('resolution: %d', (resolution))
logger.info('choice number: %d', (choice_num))
# initialize flops look-up table
model_est = FlopsEst(model)
flops_dict, flops_fixed = model_est.flops_dict, model_est.flops_fixed
# optionally resume from a checkpoint
optimizer_state = None
resume_epoch = None
if cfg.AUTO_RESUME:
optimizer_state, resume_epoch = resume_checkpoint(
model, cfg.RESUME_PATH)
# create optimizer and resume from checkpoint
optimizer = create_optimizer_supernet(cfg, model, USE_APEX)
if optimizer_state is not None:
optimizer.load_state_dict(optimizer_state['optimizer'])
model = model.cuda()
# convert model to distributed mode
if cfg.BATCHNORM.SYNC_BN:
try:
if USE_APEX:
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.')
except Exception as exception:
logger.info(
'Failed to enable Synchronized BatchNorm. '
'Install Apex or Torch >= 1.1 with Exception %s', exception)
if USE_APEX:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logger.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
# can use device str in Torch >= 1.1
model = DDP(model, device_ids=[args.local_rank])
# create learning rate scheduler
lr_scheduler, num_epochs = create_supernet_scheduler(cfg, optimizer)
start_epoch = resume_epoch if resume_epoch is not None else 0
if start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logger.info('Scheduled epochs: %d', num_epochs)
# imagenet train dataset
train_dir = os.path.join(cfg.DATA_DIR, 'train')
if not os.path.exists(train_dir):
logger.info('Training folder does not exist at: %s', train_dir)
sys.exit()
dataset_train = Dataset(train_dir)
loader_train = create_loader(dataset_train,
input_size=(3, cfg.DATASET.IMAGE_SIZE,
cfg.DATASET.IMAGE_SIZE),
batch_size=cfg.DATASET.BATCH_SIZE,
is_training=True,
use_prefetcher=True,
re_prob=cfg.AUGMENTATION.RE_PROB,
re_mode=cfg.AUGMENTATION.RE_MODE,
color_jitter=cfg.AUGMENTATION.COLOR_JITTER,
interpolation='random',
num_workers=cfg.WORKERS,
distributed=True,
collate_fn=None,
crop_pct=DEFAULT_CROP_PCT,
mean=IMAGENET_DEFAULT_MEAN,
std=IMAGENET_DEFAULT_STD)
# imagenet validation dataset
eval_dir = os.path.join(cfg.DATA_DIR, 'val')
if not os.path.isdir(eval_dir):
logger.info('Validation folder does not exist at: %s', eval_dir)
sys.exit()
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(dataset_eval,
input_size=(3, cfg.DATASET.IMAGE_SIZE,
cfg.DATASET.IMAGE_SIZE),
batch_size=4 * cfg.DATASET.BATCH_SIZE,
is_training=False,
use_prefetcher=True,
num_workers=cfg.WORKERS,
distributed=True,
crop_pct=DEFAULT_CROP_PCT,
mean=IMAGENET_DEFAULT_MEAN,
std=IMAGENET_DEFAULT_STD,
interpolation=cfg.DATASET.INTERPOLATION)
# whether to use label smoothing
if cfg.AUGMENTATION.SMOOTHING > 0.:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=cfg.AUGMENTATION.SMOOTHING).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
mutator = RandomMutator(model)
trainer = CreamSupernetTrainer(model,
train_loss_fn,
validate_loss_fn,
optimizer,
num_epochs,
loader_train,
loader_eval,
mutator=mutator,
batch_size=cfg.DATASET.BATCH_SIZE,
log_frequency=cfg.LOG_INTERVAL,
meta_sta_epoch=cfg.SUPERNET.META_STA_EPOCH,
update_iter=cfg.SUPERNET.UPDATE_ITER,
slices=cfg.SUPERNET.SLICE,
pool_size=cfg.SUPERNET.POOL_SIZE,
pick_method=cfg.SUPERNET.PICK_METHOD,
choice_num=choice_num,
sta_num=sta_num,
acc_gap=cfg.ACC_GAP,
flops_dict=flops_dict,
flops_fixed=flops_fixed,
local_rank=args.local_rank,
callbacks=[
LRSchedulerCallback(lr_scheduler),
ModelCheckpoint(output_dir)
])
trainer.train()
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)
net_config = json.load(open(args.model_config))
model = NSGANetV2.build_from_config(net_config,
drop_connect_rate=args.drop_path)
init = torch.load(args.initial_checkpoint,
map_location='cpu')['state_dict']
model.load_state_dict(init)
if args.reset_classifier:
NSGANetV2.reset_classifier(model,
last_channel=model.classifier.in_features,
n_classes=args.num_classes,
dropout_rate=args.drop)
# create a dummy model to get the model cfg just for running this timm training script
dummy_model = create_model('efficientnet_b0')
# add a teacher model
if args.teacher:
# using the supernet at full scale to supervise the training of the subnets
# this is taken from https://github.com/mit-han-lab/once-for-all/blob/
# 4decacf9d85dbc948a902c6dc34ea032d711e0a9/train_ofa_net.py#L125
from evaluator import OFAEvaluator
supernet = OFAEvaluator(n_classes=1000, model_path=args.teacher)
teacher, _ = supernet.sample({
'ks': [7] * 20,
'e': [6] * 20,
'd': [4] * 5
})
# as alternative, you could simply use a pretrained model from timm, e.g. "tf_efficientnet_b1",
# "mobilenetv3_large_100", etc. See https://rwightman.github.io/pytorch-image-models/results/ for full list.
# import timm
# teacher = timm.create_model(args.teacher, pretrained=True)
args.kd_ratio = 1.0
print("teacher model loaded")
else:
args.kd_ratio = 0.0
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=dummy_model,
verbose=args.local_rank == 0)
if args.img_size is not None:
data_config['input_size'] = (3, args.img_size, args.img_size
) # override the input image resolution
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()
if args.teacher:
teacher = nn.DataParallel(teacher,
device_ids=list(range(
args.num_gpu))).cuda()
else:
model.cuda()
if args.teacher: teacher.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)
if args.teacher: teacher = DDP(teacher, 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
if teacher: teacher = DDP(teacher, device_ids=[args.local_rank])
# 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,
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,
teacher=teacher)
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(args):
utils.init_distributed_mode(args)
print(args)
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()
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(
1.0 * 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,
# pretrained=False,
# num_classes=args.nb_classes,
# drop_rate=args.drop,
# drop_path_rate=args.drop_path,
# drop_block_rate=None,
# )
model = getattr(SwinTransformer, args.model)(num_classes=args.nb_classes,
drop_path_rate=args.drop_path)
model.to(device)
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
linear_scaled_warmup_lr = args.warmup_lr * args.batch_size * utils.get_world_size(
) / 512.0
args.warmup_lr = linear_scaled_warmup_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 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.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)
lr_scheduler.step(epoch + 1)
train_stats = train_one_epoch(
model,
criterion,
data_loader_train,
optimizer,
device,
epoch,
loss_scaler,
args.clip_grad,
mixup_fn,
set_training_mode=True # keep in eval mode during finetuning
)
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,
'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 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, config):
dataset_train, dataset_val, data_loader_train, data_loader_val, mixup_fn = build_loader(
config)
logger.info(f"Creating model:{config.MODEL.TYPE}/{config.MODEL.NAME}")
model = build_model(config)
model.cuda()
if args.use_sync_bn:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
logger.info(str(model))
optimizer = build_optimizer(config, model)
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[config.LOCAL_RANK], broadcast_buffers=False)
loss_scaler = NativeScalerWithGradNormCount()
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
logger.info(f"number of params: {n_parameters}")
if hasattr(model_without_ddp, 'flops'):
flops = model_without_ddp.flops()
logger.info(f"number of GFLOPs: {flops / 1e9}")
lr_scheduler = build_scheduler(
config, optimizer,
len(data_loader_train) // config.TRAIN.ACCUMULATION_STEPS)
if config.DISTILL.ENABLED:
# we disable MIXUP and CUTMIX when knowledge distillation
assert len(config.DISTILL.TEACHER_LOGITS_PATH
) > 0, "Please fill in DISTILL.TEACHER_LOGITS_PATH"
criterion = torch.nn.CrossEntropyLoss(reduction='mean')
else:
if config.AUG.MIXUP > 0.:
# smoothing is handled with mixup label transform
criterion = SoftTargetCrossEntropy()
elif config.MODEL.LABEL_SMOOTHING > 0.:
criterion = LabelSmoothingCrossEntropy(
smoothing=config.MODEL.LABEL_SMOOTHING)
else:
criterion = torch.nn.CrossEntropyLoss()
max_accuracy = 0.0
if config.TRAIN.AUTO_RESUME:
resume_file = auto_resume_helper(config.OUTPUT)
if resume_file:
if config.MODEL.RESUME:
logger.warning(
f"auto-resume changing resume file from {config.MODEL.RESUME} to {resume_file}"
)
config.defrost()
config.MODEL.RESUME = resume_file
config.freeze()
logger.info(f'auto resuming from {resume_file}')
else:
logger.info(
f'no checkpoint found in {config.OUTPUT}, ignoring auto resume'
)
if config.MODEL.RESUME:
max_accuracy = load_checkpoint(config, model_without_ddp, optimizer,
lr_scheduler, loss_scaler, logger)
acc1, acc5, loss = validate(args, config, data_loader_val, model)
logger.info(
f"Accuracy of the network on the {len(dataset_val)} test images: {acc1:.1f}%"
)
if config.EVAL_MODE:
return
if config.MODEL.PRETRAINED and (not config.MODEL.RESUME):
load_pretrained(config, model_without_ddp, logger)
acc1, acc5, loss = validate(args, config, data_loader_val, model)
logger.info(
f"Accuracy of the network on the {len(dataset_val)} test images: {acc1:.1f}%"
)
if config.THROUGHPUT_MODE:
throughput(data_loader_val, model, logger)
return
logger.info("Start training")
start_time = time.time()
for epoch in range(config.TRAIN.START_EPOCH, config.TRAIN.EPOCHS):
# set_epoch for dataset_train when distillation
if hasattr(dataset_train, 'set_epoch'):
dataset_train.set_epoch(epoch)
data_loader_train.sampler.set_epoch(epoch)
if config.DISTILL.ENABLED:
train_one_epoch_distill_using_saved_logits(
args, config, model, criterion, data_loader_train, optimizer,
epoch, mixup_fn, lr_scheduler, loss_scaler)
else:
train_one_epoch(args, config, model, criterion, data_loader_train,
optimizer, epoch, mixup_fn, lr_scheduler,
loss_scaler)
if dist.get_rank() == 0 and (epoch % config.SAVE_FREQ == 0
or epoch == (config.TRAIN.EPOCHS - 1)):
save_checkpoint(config, epoch, model_without_ddp, max_accuracy,
optimizer, lr_scheduler, loss_scaler, logger)
acc1, acc5, loss = validate(args, config, data_loader_val, model)
logger.info(
f"Accuracy of the network on the {len(dataset_val)} test images: {acc1:.1f}%"
)
max_accuracy = max(max_accuracy, acc1)
logger.info(f'Max accuracy: {max_accuracy:.2f}%')
if is_main_process() and args.use_wandb:
wandb.log({
f"val/acc@1": acc1,
f"val/acc@5": acc5,
f"val/loss": loss,
"epoch": epoch,
})
wandb.run.summary['epoch'] = epoch
wandb.run.summary['best_acc@1'] = max_accuracy
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
logger.info('Training time {}'.format(total_time_str))
def main():
setup_default_logging()
args, args_text = _parse_args()
if args.log_wandb:
if has_wandb:
wandb.init(project=args.experiment, config=args)
else:
_logger.warning(
"You've requested to log metrics to wandb but package not found. "
"Metrics not being logged to wandb, try `pip install wandb`")
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
# resolve AMP arguments based on PyTorch / Apex availability
use_amp = None
if args.amp:
# `--amp` chooses native amp before apex (APEX ver not actively maintained)
if has_native_amp:
args.native_amp = True
elif has_apex:
args.apex_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")
random_seed(args.seed, args.rank)
if args.fuser:
set_jit_fuser(args.fuser)
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_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
scriptable=args.torchscript,
checkpoint_path=args.initial_checkpoint)
if args.num_classes is None:
assert hasattr(
model, 'num_classes'
), 'Model must have `num_classes` attr if not set on cmd line/config.'
args.num_classes = model.num_classes # FIXME handle model default vs config num_classes more elegantly
if args.grad_checkpointing:
model.set_grad_checkpointing(enable=True)
if args.local_rank == 0:
_logger.info(
f'Model {safe_model_name(args.model)} created, param count:{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.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 == 'apex':
# 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)
optimizer = create_optimizer_v2(model, **optimizer_kwargs(cfg=args))
# 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.')
# 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)
# 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 DDP wrapper
model_ema = ModelEmaV2(
model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else None)
if args.resume:
load_checkpoint(model_ema.module, args.resume, use_ema=True)
# setup distributed training
if args.distributed:
if has_apex and use_amp == 'apex':
# 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],
broadcast_buffers=not args.no_ddp_bb)
# NOTE: EMA model does not need to be wrapped by DDP
# setup learning rate schedule and starting epoch
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))
# create the train and eval datasets
dataset_train = create_dataset(args.dataset,
root=args.data_dir,
split=args.train_split,
is_training=True,
class_map=args.class_map,
download=args.dataset_download,
batch_size=args.batch_size,
repeats=args.epoch_repeats)
dataset_eval = create_dataset(args.dataset,
root=args.data_dir,
split=args.val_split,
is_training=False,
class_map=args.class_map,
download=args.dataset_download,
batch_size=args.batch_size)
# 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)
# 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 pipeiine
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_repeats=args.aug_repeats,
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,
worker_seeding=args.worker_seeding,
)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size or 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 loss function
if args.jsd_loss:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits,
smoothing=args.smoothing)
elif mixup_active:
# smoothing is handled with mixup target transform which outputs sparse, soft targets
if args.bce_loss:
train_loss_fn = BinaryCrossEntropy(
target_threshold=args.bce_target_thresh)
else:
train_loss_fn = SoftTargetCrossEntropy()
elif args.smoothing:
if args.bce_loss:
train_loss_fn = BinaryCrossEntropy(
smoothing=args.smoothing,
target_threshold=args.bce_target_thresh)
else:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing)
else:
train_loss_fn = nn.CrossEntropyLoss()
train_loss_fn = train_loss_fn.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 = None
if args.rank == 0:
if args.experiment:
exp_name = args.experiment
else:
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
safe_model_name(args.model),
str(data_config['input_size'][-1])
])
output_dir = get_outdir(
args.output if args.output else './output/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,
max_history=args.checkpoint_hist)
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 and hasattr(loader_train.sampler, 'set_epoch'):
loader_train.sampler.set_epoch(epoch)
train_metrics = train_one_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.module,
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])
if output_dir is not None:
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None,
log_wandb=args.log_wandb and has_wandb)
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 _train_loop(self, train_loader, val_loader, time_limit=math.inf):
start_tic = time.time()
# setup loss function
if self.mixup_active:
# smoothing is handled with mixup target transform
train_loss_fn = SoftTargetCrossEntropy()
elif self._augmentation_cfg.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(smoothing=self._augmentation_cfg.smoothing)
else:
train_loss_fn = nn.CrossEntropyLoss()
validate_loss_fn = nn.CrossEntropyLoss()
train_loss_fn = train_loss_fn.to(self.ctx[0])
validate_loss_fn = validate_loss_fn.to(self.ctx[0])
eval_metric = self._misc_cfg.eval_metric
if self._problem_type == REGRESSION:
train_loss_fn = nn.MSELoss()
validate_loss_fn = nn.MSELoss()
eval_metric = 'rmse'
early_stopper = EarlyStopperOnPlateau(
patience=self._train_cfg.early_stop_patience,
min_delta=self._train_cfg.early_stop_min_delta,
baseline_value=self._train_cfg.early_stop_baseline,
max_value=self._train_cfg.early_stop_max_value)
self._logger.info('Start training from [Epoch %d]', max(self._train_cfg.start_epoch, self.epoch))
self._time_elapsed += time.time() - start_tic
for self.epoch in range(max(self.start_epoch, self.epoch), self.epochs):
epoch = self.epoch
if self._best_acc >= 1.0:
self._logger.info('[Epoch {}] Early stopping as acc is reaching 1.0'.format(epoch))
break
should_stop, stop_message = early_stopper.get_early_stop_advice()
if should_stop:
self._logger.info('[Epoch {}] '.format(epoch) + stop_message)
break
train_metrics = self.train_one_epoch(
epoch, self.net, train_loader, self._optimizer, train_loss_fn,
lr_scheduler=self._lr_scheduler, output_dir=self._logdir,
amp_autocast=self._amp_autocast, loss_scaler=self._loss_scaler, model_ema=self._model_ema, mixup_fn=self._mixup_fn, time_limit=time_limit)
# reaching time limit, exit early
if train_metrics['time_limit']:
self._logger.warning(f'`time_limit={time_limit}` reached, exit early...')
return {'train_acc': train_metrics['train_acc'], 'valid_acc': self._best_acc,
'time': self._time_elapsed, 'checkpoint': self._cp_name}
post_tic = time.time()
eval_metrics = self.validate(self.net, val_loader, validate_loss_fn, amp_autocast=self._amp_autocast)
if self._model_ema is not None and not self._model_ema_cfg.model_ema_force_cpu:
ema_eval_metrics = self.validate(
self._model_ema.module, val_loader, validate_loss_fn, amp_autocast=self._amp_autocast)
eval_metrics = ema_eval_metrics
if self._problem_type == REGRESSION:
val_acc = eval_metrics['rmse']
if self._reporter:
self._reporter(epoch=epoch, acc_reward=-val_acc)
early_stopper.update(-val_acc)
if -val_acc > self._best_acc:
self._cp_name = os.path.join(self._logdir, _BEST_CHECKPOINT_FILE)
self._logger.info('[Epoch %d] Current best rmse: %f vs previous %f, saved to %s',
self.epoch, val_acc, -self._best_acc, self._cp_name)
self.save(self._cp_name)
self._best_acc = -val_acc
else:
val_acc = eval_metrics['top1']
if self._reporter:
self._reporter(epoch=epoch, acc_reward=val_acc)
early_stopper.update(val_acc)
if val_acc > self._best_acc:
self._cp_name = os.path.join(self._logdir, _BEST_CHECKPOINT_FILE)
self._logger.info('[Epoch %d] Current best top-1: %f vs previous %f, saved to %s',
self.epoch, val_acc, self._best_acc, self._cp_name)
self.save(self._cp_name)
self._best_acc = val_acc
if self._lr_scheduler is not None:
# step LR for next epoch
self._lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
self._time_elapsed += time.time() - post_tic
if 'accuracy' in train_metrics:
return {'train_acc': train_metrics['accuracy'], 'valid_acc': self._best_acc,
'time': self._time_elapsed, 'checkpoint': self._cp_name}
# rmse
else:
if self._problem_type == REGRESSION:
return {'train_score': train_metrics['rmse'], 'valid_score': -self._best_acc,
'time': self._time_elapsed, 'checkpoint': self._cp_name}
# mixup
else:
return {'train_score': train_metrics['rmse'], 'valid_acc': self._best_acc,
'time': self._time_elapsed, 'checkpoint': self._cp_name}
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))
num_workers=args.workers,
crop_pct=crop_pct,
pin_memory=args.pin_mem,
tf_preprocessing=args.tf_preprocessing)
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)
validate_loss_fn = nn.CrossEntropyLoss()
# elif 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
optimizer = create_optimizer(args, model_raw)
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
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(config):
dataset_train, dataset_val, data_loader_train, data_loader_val, mixup_fn = build_loader(config)
logger.info(f"Creating model:{config.MODEL.TYPE}/{config.MODEL.NAME}")
model = build_model(config)
model.cuda()
# logger.info(str(model))
optimizer = build_optimizer(config, model)
if config.AMP_OPT_LEVEL != "O0":
model, optimizer = amp.initialize(model, optimizer, opt_level=config.AMP_OPT_LEVEL)
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[config.LOCAL_RANK], broadcast_buffers=False)
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)
logger.info(f"number of params: {n_parameters}")
if hasattr(model_without_ddp, 'flops'):
flops = model_without_ddp.flops()
logger.info(f"number of GFLOPs: {flops / 1e9}")
lr_scheduler = build_scheduler(config, optimizer, len(data_loader_train))
if config.AUG.MIXUP > 0.:
# smoothing is handled with mixup label transform
criterion = SoftTargetCrossEntropy()
elif config.DATA.DATASET == "plant":
criterion = BCELoss()
elif config.MODEL.LABEL_SMOOTHING > 0.:
criterion = LabelSmoothingCrossEntropy(smoothing=config.MODEL.LABEL_SMOOTHING)
else:
criterion = torch.nn.CrossEntropyLoss()
max_accuracy = 0.0
if config.TRAIN.AUTO_RESUME:
resume_file = auto_resume_helper(config.OUTPUT)
if resume_file:
if config.MODEL.RESUME:
logger.warning(f"auto-resume changing resume file from {config.MODEL.RESUME} to {resume_file}")
config.defrost()
config.MODEL.RESUME = resume_file
config.freeze()
logger.info(f'auto resuming from {resume_file}')
else:
logger.info(f'no checkpoint found in {config.OUTPUT}, ignoring auto resume')
class_type = "single_class" if config.DATA.DATASET == "imagenet" else "multi_class"
if config.MODEL.RESUME:
max_accuracy = load_checkpoint(config, model_without_ddp, optimizer, lr_scheduler, logger)
acc1, acc5, loss = validate(config, data_loader_val, model, type=class_type)
if class_type == "single_class":
logger.info(f"Accuracy of the network on the {len(dataset_val)} test images: {acc1:.1f}%")
else:
logger.info(f"f1_score of the network on the {len(dataset_val)} test images: {acc1:.1f}")
if config.EVAL_MODE:
return
if config.THROUGHPUT_MODE:
throughput(data_loader_val, model, logger)
return
logger.info("Start training")
start_time = time.time()
for epoch in range(config.TRAIN.START_EPOCH, config.TRAIN.EPOCHS):
data_loader_train.sampler.set_epoch(epoch)
train_one_epoch(config, model, criterion, data_loader_train, optimizer, epoch, mixup_fn, lr_scheduler)
if dist.get_rank() == 0 and (epoch % config.SAVE_FREQ == 0 or epoch == (config.TRAIN.EPOCHS - 1)):
save_checkpoint(config, epoch, model_without_ddp, max_accuracy, optimizer, lr_scheduler, logger)
acc1, acc5, loss = validate(config, data_loader_val, model, type=class_type)
if not (dist.get_rank() == 0 and (epoch % config.SAVE_FREQ == 0 or epoch == (config.TRAIN.EPOCHS - 1))):
if acc1 > max_accuracy:
save_checkpoint(config, epoch, model_without_ddp, max_accuracy, optimizer, lr_scheduler, logger)
max_accuracy = max(max_accuracy, acc1)
if class_type == "single_class":
logger.info(f"Accuracy of the network on the {len(dataset_val)} test images: {acc1:.3f}%")
logger.info(f'Max accuracy: {max_accuracy:.3f}%')
else:
logger.info(f"f1_score of the network on the {len(dataset_val)} test images: {acc1:.3f}")
logger.info(f'Max f1_score: {max_accuracy:.3f}')
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
logger.info('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:
_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)
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:
_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)
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)
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()
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():
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():
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:
print(
'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:
print(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
print('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)
print('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)
if args.num_gpu > 1:
if args.amp:
print(
'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:
if args.distributed and args.sync_bn and has_apex:
model = convert_syncbn_model(model)
model.cuda()
optimizer = create_optimizer(args, model)
if optimizer_state is not None:
optimizer.load_state_dict(optimizer_state)
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
print('AMP enabled')
else:
use_amp = False
print('AMP disabled')
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=args.resume)
if args.distributed:
model = DDP(model, delay_allreduce=True)
if model_ema is not None and not args.model_ema_force_cpu:
# must also distribute EMA model to allow validation
model_ema.ema = DDP(model_ema.ema, delay_allreduce=True)
model_ema.ema_has_module = True
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
if start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
print('Scheduled epochs: ', num_epochs)
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
print('Error: training folder does not exist at: %s' % 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,
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, 'validation')
if not os.path.isdir(eval_dir):
print('Error: validation folder does not exist at: %s' % 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)
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:
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:
print('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
def main(config):
dataset_train, dataset_val, data_loader_train, data_loader_val, mixup_fn = build_loader(
config)
if config.DISTILL.DO_DISTILL:
logger.info(
f"Loading teacher model:{config.MODEL.TYPE}/{config.DISTILL.TEACHER}"
)
model_checkpoint_name = os.path.basename(config.DISTILL.TEACHER)
if 'regnety_160' in model_checkpoint_name:
model_teacher = create_model(
'regnety_160',
pretrained=False,
num_classes=config.MODEL.NUM_CLASSES,
global_pool='avg',
)
if config.DISTILL.TEACHER.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(
config.DISTILL.TEACHER,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(config.DISTILL.TEACHER,
map_location='cpu')
model_teacher.load_state_dict(checkpoint['model'])
model_teacher.cuda()
model_teacher.eval()
del checkpoint
torch.cuda.empty_cache()
else:
if 'base' in model_checkpoint_name:
teacher_type = 'base'
elif 'large' in model_checkpoint_name:
teacher_type = 'large'
else:
teacher_type = None
model_teacher = load_teacher_model(type=teacher_type)
model_teacher.cuda()
model_teacher = torch.nn.parallel.DistributedDataParallel(
model_teacher,
device_ids=[config.LOCAL_RANK],
broadcast_buffers=False)
checkpoint = torch.load(config.DISTILL.TEACHER, map_location='cpu')
msg = model_teacher.module.load_state_dict(checkpoint['model'],
strict=False)
logger.info(msg)
del checkpoint
torch.cuda.empty_cache()
logger.info(f"Creating model:{config.MODEL.TYPE}/{config.MODEL.NAME}")
model = build_model(config)
model.cuda()
logger.info(str(model))
optimizer = build_optimizer(config, model)
if config.AMP_OPT_LEVEL != "O0":
model, optimizer = amp.initialize(model,
optimizer,
opt_level=config.AMP_OPT_LEVEL)
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[config.LOCAL_RANK],
broadcast_buffers=False,
find_unused_parameters=True)
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
logger.info(f"number of params: {n_parameters}")
if hasattr(model_without_ddp, 'flops'):
flops = model_without_ddp.flops()
logger.info(f"number of GFLOPs: {flops / 1e9}")
lr_scheduler = build_scheduler(config, optimizer, len(data_loader_train))
criterion_soft = soft_cross_entropy
criterion_attn = cal_relation_loss
criterion_hidden = cal_hidden_relation_loss if config.DISTILL.HIDDEN_RELATION else cal_hidden_loss
if config.AUG.MIXUP > 0.:
# smoothing is handled with mixup label transform
criterion_truth = SoftTargetCrossEntropy()
elif config.MODEL.LABEL_SMOOTHING > 0.:
criterion_truth = LabelSmoothingCrossEntropy(
smoothing=config.MODEL.LABEL_SMOOTHING)
else:
criterion_truth = torch.nn.CrossEntropyLoss()
max_accuracy = 0.0
if config.TRAIN.AUTO_RESUME:
resume_file = auto_resume_helper(config.OUTPUT)
if resume_file:
if config.MODEL.RESUME:
logger.warning(
f"auto-resume changing resume file from {config.MODEL.RESUME} to {resume_file}"
)
config.defrost()
config.MODEL.RESUME = resume_file
config.DISTILL.RESUME_WEIGHT_ONLY = False
config.freeze()
logger.info(f'auto resuming from {resume_file}')
else:
logger.info(
f'no checkpoint found in {config.OUTPUT}, ignoring auto resume'
)
if config.MODEL.RESUME:
max_accuracy = load_checkpoint(config, model_without_ddp, optimizer,
lr_scheduler, logger)
acc1, acc5, loss = validate(config, data_loader_val, model, logger)
logger.info(
f"Accuracy of the network on the {len(dataset_val)} test images: {acc1:.1f}%"
)
if config.EVAL_MODE:
return
if config.THROUGHPUT_MODE:
throughput(data_loader_val, model, logger)
return
logger.info("Start training")
start_time = time.time()
for epoch in range(config.TRAIN.START_EPOCH, config.TRAIN.EPOCHS):
data_loader_train.sampler.set_epoch(epoch)
if config.DISTILL.DO_DISTILL:
train_one_epoch_distill(config,
model,
model_teacher,
data_loader_train,
optimizer,
epoch,
mixup_fn,
lr_scheduler,
criterion_soft=criterion_soft,
criterion_truth=criterion_truth,
criterion_attn=criterion_attn,
criterion_hidden=criterion_hidden)
else:
train_one_epoch(config, model, criterion_truth, data_loader_train,
optimizer, epoch, mixup_fn, lr_scheduler)
if dist.get_rank() == 0 and (epoch % config.SAVE_FREQ == 0
or epoch == (config.TRAIN.EPOCHS - 1)):
save_checkpoint(config, epoch, model_without_ddp, max_accuracy,
optimizer, lr_scheduler, logger)
if epoch % config.EVAL_FREQ == 0 or epoch == config.TRAIN.EPOCHS - 1:
acc1, acc5, loss = validate(config, data_loader_val, model, logger)
logger.info(
f"Accuracy of the network on the {len(dataset_val)} test images: {acc1:.1f}%"
)
max_accuracy = max(max_accuracy, acc1)
logger.info(f'Max accuracy: {max_accuracy:.2f}%')
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
logger.info('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
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
# 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")
torch.manual_seed(args.seed + args.rank)
####################################################################################
# Start - SparseML optional load weights from SparseZoo
####################################################################################
if args.initial_checkpoint == "zoo":
# Load checkpoint from base weights associated with given SparseZoo recipe
if args.sparseml_recipe.startswith("zoo:"):
args.initial_checkpoint = Zoo.download_recipe_base_framework_files(
args.sparseml_recipe,
extensions=[".pth.tar", ".pth"]
)[0]
else:
raise ValueError(
"Attempting to load weights from SparseZoo recipe, but not given a "
"SparseZoo recipe stub. When initial-checkpoint is set to 'zoo'. "
"sparseml-recipe must start with 'zoo:' and be a SparseZoo model "
f"stub. sparseml-recipe was set to {args.sparseml_recipe}"
)
elif args.initial_checkpoint.startswith("zoo:"):
# Load weights from a SparseZoo model stub
zoo_model = Zoo.load_model_from_stub(args.initial_checkpoint)
args.initial_checkpoint = zoo_model.download_framework_files(extensions=[".pth"])
####################################################################################
# End - SparseML optional load weights from SparseZoo
####################################################################################
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,
scriptable=args.torchscript,
checkpoint_path=args.initial_checkpoint)
if args.num_classes is None:
assert hasattr(model, 'num_classes'), 'Model must have `num_classes` attr if not set on cmd line/config.'
args.num_classes = model.num_classes # FIXME handle model default vs config num_classes more elegantly
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.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)
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.')
# 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)
# 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:
load_checkpoint(model_ema.module, args.resume, use_ema=True)
# 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 learning rate schedule and starting epoch
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)
# create the train and eval datasets
dataset_train = create_dataset(
args.dataset, root=args.data_dir, split=args.train_split, is_training=True, batch_size=args.batch_size)
dataset_eval = create_dataset(
args.dataset, root=args.data_dir, split=args.val_split, is_training=False, batch_size=args.batch_size)
# 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)
# 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 pipeiine
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 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=output_dir, decreasing=decreasing, max_history=args.checkpoint_hist)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
####################################################################################
# Start SparseML Integration
####################################################################################
sparseml_loggers = (
[PythonLogger(), TensorBoardLogger(log_path=output_dir)]
if output_dir
else None
)
manager = ScheduledModifierManager.from_yaml(args.sparseml_recipe)
optimizer = ScheduledOptimizer(
optimizer,
model,
manager,
steps_per_epoch=len(loader_train),
loggers=sparseml_loggers
)
# override lr scheduler if recipe makes any LR updates
if any("LearningRate" in str(modifier) for modifier in manager.modifiers):
_logger.info("Disabling timm LR scheduler, managing LR using SparseML recipe")
lr_scheduler = None
if manager.max_epochs:
_logger.info(
f"Overriding max_epochs to {manager.max_epochs} from SparseML recipe"
)
num_epochs = manager.max_epochs or num_epochs
####################################################################################
# End SparseML Integration
####################################################################################
if args.local_rank == 0:
_logger.info('Scheduled epochs: {}'.format(num_epochs))
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed and hasattr(loader_train.sampler, 'set_epoch'):
loader_train.sampler.set_epoch(epoch)
train_metrics = train_one_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.module, 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)
#################################################################################
# Start SparseML ONNX Export
#################################################################################
if output_dir:
_logger.info(
f"training complete, exporting ONNX to {output_dir}/model.onnx"
)
exporter = ModuleExporter(model, output_dir)
exporter.export_onnx(torch.randn((1, *data_config["input_size"])))
#################################################################################
# End SparseML ONNX Export
#################################################################################
except KeyboardInterrupt:
pass
if best_metric is not None:
_logger.info('*** Best metric: {0} (epoch {1})'.format(best_metric, best_epoch))
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(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:
print(retrain_config)
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(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)
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=args.batch_size // 2,
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("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)
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(file=output_dir / "config.yaml", mode='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}
trainer = AFSupernetTrainer(
model, criterion, data_loader_train, data_loader_val,
optimizer, device, args.epochs, loss_scaler,
args.clip_grad, model_ema, mixup_fn,
args.amp, teacher_model, teacher_loss,choices, args.mode, retrain_config, 0., output_dir, lr_scheduler,
)
if args.eval:
trainer._validate_one_epoch(-1)
return
trainer.fit()
