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
if args.load_tar:
train_dir = os.path.join(args.data_path, 'train.tar')
train_transform = build_transform(True, args)
dataset_train = DatasetTar(train_dir, transform=train_transform)
args.nb_classes = 1000
val_transform = build_transform(False, args)
eval_dir = os.path.join(args.data_path, 'val.tar')
dataset_val = DatasetTar(eval_dir, transform=val_transform)
else:
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=args.pretrained,
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(args):
if args.apex:
if sys.version_info < (3, 0):
raise RuntimeError(
"Apex currently only supports Python 3. Aborting.")
if amp is None:
raise RuntimeError(
"Failed to import apex. Please install apex from https://www.github.com/nvidia/apex "
"to enable mixed-precision training.")
if args.output_dir:
utils.mkdir(args.output_dir)
utils.init_distributed_mode(args)
print(args)
device = torch.device(args.device)
torch.backends.cudnn.benchmark = True
# Data loading code
print("Loading data")
traindir = os.path.join(args.data_path, 'train')
valdir = os.path.join(args.data_path, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
print("Loading training data")
st = time.time()
cache_path = _get_cache_path(traindir)
if args.cache_dataset and os.path.exists(cache_path):
# Attention, as the transforms are also cached!
print("Loading dataset_train from {}".format(cache_path))
dataset, _ = torch.load(cache_path)
else:
dataset = torchvision.datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
if args.cache_dataset:
print("Saving dataset_train to {}".format(cache_path))
utils.mkdir(os.path.dirname(cache_path))
utils.save_on_master((dataset, traindir), cache_path)
print("Took", time.time() - st)
print("Loading validation data")
cache_path = _get_cache_path(valdir)
if args.cache_dataset and os.path.exists(cache_path):
# Attention, as the transforms are also cached!
print("Loading dataset_test from {}".format(cache_path))
dataset_test, _ = torch.load(cache_path)
else:
dataset_test = torchvision.datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
]))
if args.cache_dataset:
print("Saving dataset_test to {}".format(cache_path))
utils.mkdir(os.path.dirname(cache_path))
utils.save_on_master((dataset_test, valdir), cache_path)
print("Creating data loaders")
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
dataset)
test_sampler = torch.utils.data.distributed.DistributedSampler(
dataset_test)
else:
train_sampler = torch.utils.data.RandomSampler(dataset)
test_sampler = torch.utils.data.SequentialSampler(dataset_test)
data_loader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
sampler=train_sampler,
num_workers=args.workers,
pin_memory=True)
data_loader_test = torch.utils.data.DataLoader(dataset_test,
batch_size=args.batch_size,
sampler=test_sampler,
num_workers=args.workers,
pin_memory=True)
print("Creating model")
model = torchvision.models.__dict__[args.model](pretrained=args.pretrained)
model.to(device)
if args.distributed and args.sync_bn:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=args.lr_step_size,
gamma=args.lr_gamma)
if args.apex:
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.apex_opt_level)
if args.resume:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.test_only:
evaluate(model, criterion, data_loader_test, device=device)
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
train_one_epoch(model, criterion, optimizer, data_loader, device,
epoch, args.print_freq, args.apex)
lr_scheduler.step()
evaluate(model, criterion, data_loader_test, device=device)
if args.output_dir:
checkpoint = {
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
'args': args
}
utils.save_on_master(
checkpoint,
os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
utils.save_on_master(
checkpoint, os.path.join(args.output_dir, 'checkpoint.pth'))
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):
if args.apex:
if sys.version_info < (3, 0):
raise RuntimeError(
"Apex currently only supports Python 3. Aborting.")
if amp is None:
raise RuntimeError(
"Failed to import apex. Please install apex from https://www.github.com/nvidia/apex "
"to enable mixed-precision training.")
if args.output_dir:
utils.mkdir(args.output_dir)
utils.init_distributed_mode(args)
print(args)
print("torch version: ", torch.__version__)
print("torchvision version: ", torchvision.__version__)
device = torch.device(args.device)
torch.backends.cudnn.benchmark = True
# Data loading code
print("Loading data")
traindir = os.path.join(args.data_path, 'train_avi-480p')
valdir = os.path.join(args.data_path, 'val_avi-480p')
normalize = T.Normalize(mean=[0.43216, 0.394666, 0.37645],
std=[0.22803, 0.22145, 0.216989])
print("Loading training data")
st = time.time()
cache_path = _get_cache_path(traindir)
transform_train = torchvision.transforms.Compose([
T.ToFloatTensorInZeroOne(),
T.Resize((128, 171)),
T.RandomHorizontalFlip(), normalize,
T.RandomCrop((112, 112))
])
if args.cache_dataset and os.path.exists(cache_path):
print("Loading dataset_train from {}".format(cache_path))
dataset, _ = torch.load(cache_path)
dataset.transform = transform_train
else:
if args.distributed:
print("It is recommended to pre-compute the dataset cache "
"on a single-gpu first, as it will be faster")
dataset = torchvision.datasets.Kinetics400(
traindir,
frames_per_clip=args.clip_len,
step_between_clips=1,
transform=transform_train)
if args.cache_dataset:
print("Saving dataset_train to {}".format(cache_path))
utils.mkdir(os.path.dirname(cache_path))
utils.save_on_master((dataset, traindir), cache_path)
dataset.video_clips.compute_clips(args.clip_len, 1, frame_rate=15)
print("Took", time.time() - st)
print("Loading validation data")
cache_path = _get_cache_path(valdir)
transform_test = torchvision.transforms.Compose([
T.ToFloatTensorInZeroOne(),
T.Resize((128, 171)), normalize,
T.CenterCrop((112, 112))
])
if args.cache_dataset and os.path.exists(cache_path):
print("Loading dataset_test from {}".format(cache_path))
dataset_test, _ = torch.load(cache_path)
dataset_test.transform = transform_test
else:
if args.distributed:
print("It is recommended to pre-compute the dataset cache "
"on a single-gpu first, as it will be faster")
dataset_test = torchvision.datasets.Kinetics400(
valdir,
frames_per_clip=args.clip_len,
step_between_clips=1,
transform=transform_test)
if args.cache_dataset:
print("Saving dataset_test to {}".format(cache_path))
utils.mkdir(os.path.dirname(cache_path))
utils.save_on_master((dataset_test, valdir), cache_path)
dataset_test.video_clips.compute_clips(args.clip_len, 1, frame_rate=15)
print("Creating data loaders")
train_sampler = torchvision.datasets.video_utils.RandomClipSampler(
dataset.video_clips, args.clips_per_video)
test_sampler = UniformClipSampler(dataset_test.video_clips,
args.clips_per_video)
if args.distributed:
train_sampler = DistributedSampler(train_sampler)
test_sampler = DistributedSampler(test_sampler)
data_loader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
sampler=train_sampler,
num_workers=args.workers,
pin_memory=True,
collate_fn=collate_fn)
data_loader_test = torch.utils.data.DataLoader(dataset_test,
batch_size=args.batch_size,
sampler=test_sampler,
num_workers=args.workers,
pin_memory=True,
collate_fn=collate_fn)
print("Creating model")
# model = torchvision.models.video.__dict__[args.model](pretrained=args.pretrained)
model = torchvision.models.video.__dict__[args.model]()
model.to(device)
if args.distributed and args.sync_bn:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
criterion = nn.CrossEntropyLoss()
lr = args.lr * args.world_size
optimizer = torch.optim.SGD(model.parameters(),
lr=lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.apex:
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.apex_opt_level)
# convert scheduler to be per iteration, not per epoch, for warmup that lasts
# between different epochs
warmup_iters = args.lr_warmup_epochs * len(data_loader)
lr_milestones = [len(data_loader) * m for m in args.lr_milestones]
lr_scheduler = WarmupMultiStepLR(optimizer,
milestones=lr_milestones,
gamma=args.lr_gamma,
warmup_iters=warmup_iters,
warmup_factor=1e-5)
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
if args.resume:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.test_only:
evaluate(model, criterion, data_loader_test, device=device)
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
train_one_epoch(model, criterion, optimizer, lr_scheduler, data_loader,
device, epoch, args.print_freq, args.apex)
evaluate(model, criterion, data_loader_test, device=device)
if args.output_dir:
checkpoint = {
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
'args': args
}
utils.save_on_master(
checkpoint,
os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
utils.save_on_master(
checkpoint, os.path.join(args.output_dir, 'checkpoint.pth'))
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)
print(args)
device = torch.device(args.device)
# Data loading code
print("Loading data")
dataset, num_classes = get_dataset(args.dataset, "train", get_transform(train=True), args.data_path)
dataset_test, _ = get_dataset(args.dataset, "valid", get_transform(train=False), args.data_path)
print("Creating data loaders")
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(dataset)
test_sampler = torch.utils.data.distributed.DistributedSampler(dataset_test)
else:
train_sampler = torch.utils.data.RandomSampler(dataset)
test_sampler = torch.utils.data.SequentialSampler(dataset_test)
if args.aspect_ratio_group_factor >= 0:
group_ids = create_aspect_ratio_groups(dataset, k=args.aspect_ratio_group_factor)
train_batch_sampler = GroupedBatchSampler(train_sampler, group_ids, args.batch_size)
else:
train_batch_sampler = torch.utils.data.BatchSampler(
train_sampler, args.batch_size, drop_last=True)
data_loader = torch.utils.data.DataLoader(
dataset, batch_sampler=train_batch_sampler, num_workers=args.workers,
collate_fn=utils.collate_fn)
data_loader_test = torch.utils.data.DataLoader(
dataset_test, batch_size=1,
sampler=test_sampler, num_workers=args.workers,
collate_fn=utils.collate_fn)
print("Creating model")
kwargs = {
"trainable_backbone_layers": args.trainable_backbone_layers
}
if "rcnn" in args.model:
if args.rpn_score_thresh is not None:
kwargs["rpn_score_thresh"] = args.rpn_score_thresh
model = torchvision.models.detection.__dict__[args.model](num_classes=num_classes, pretrained=args.pretrained,
**kwargs)
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
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(
params, lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
# lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.lr_step_size, gamma=args.lr_gamma)
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.lr_steps, gamma=args.lr_gamma)
if args.resume:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.test_only:
evaluate(model, data_loader_test, device=device)
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
train_one_epoch(model, optimizer, data_loader, device, epoch, args.print_freq)
lr_scheduler.step()
if args.output_dir:
utils.save_on_master({
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'args': args,
'epoch': epoch},
os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
# evaluate after every epoch
evaluate(model, data_loader_test, device=device)
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):
if args.output_dir:
utils.mkdir(args.output_dir)
utils.init_distributed_mode(args)
print(args)
device = torch.device(args.device)
torch.backends.cudnn.benchmark = True
# Data loading code
print("Loading data")
traindir = os.path.join(args.data_path, args.train_dir)
valdir = os.path.join(args.data_path, args.val_dir)
print("Loading training data")
st = time.time()
cache_path = _get_cache_path(traindir)
transform_train = presets.VideoClassificationPresetTrain(crop_size=(112, 112), resize_size=(128, 171))
if args.cache_dataset and os.path.exists(cache_path):
print(f"Loading dataset_train from {cache_path}")
dataset, _ = torch.load(cache_path)
dataset.transform = transform_train
else:
if args.distributed:
print("It is recommended to pre-compute the dataset cache on a single-gpu first, as it will be faster")
dataset = torchvision.datasets.Kinetics400(
traindir,
frames_per_clip=args.clip_len,
step_between_clips=1,
transform=transform_train,
frame_rate=15,
extensions=(
"avi",
"mp4",
),
)
if args.cache_dataset:
print(f"Saving dataset_train to {cache_path}")
utils.mkdir(os.path.dirname(cache_path))
utils.save_on_master((dataset, traindir), cache_path)
print("Took", time.time() - st)
print("Loading validation data")
cache_path = _get_cache_path(valdir)
if args.weights and args.test_only:
weights = torchvision.models.get_weight(args.weights)
transform_test = weights.transforms()
else:
transform_test = presets.VideoClassificationPresetEval(crop_size=(112, 112), resize_size=(128, 171))
if args.cache_dataset and os.path.exists(cache_path):
print(f"Loading dataset_test from {cache_path}")
dataset_test, _ = torch.load(cache_path)
dataset_test.transform = transform_test
else:
if args.distributed:
print("It is recommended to pre-compute the dataset cache on a single-gpu first, as it will be faster")
dataset_test = torchvision.datasets.Kinetics400(
valdir,
frames_per_clip=args.clip_len,
step_between_clips=1,
transform=transform_test,
frame_rate=15,
extensions=(
"avi",
"mp4",
),
)
if args.cache_dataset:
print(f"Saving dataset_test to {cache_path}")
utils.mkdir(os.path.dirname(cache_path))
utils.save_on_master((dataset_test, valdir), cache_path)
print("Creating data loaders")
train_sampler = RandomClipSampler(dataset.video_clips, args.clips_per_video)
test_sampler = UniformClipSampler(dataset_test.video_clips, args.clips_per_video)
if args.distributed:
train_sampler = DistributedSampler(train_sampler)
test_sampler = DistributedSampler(test_sampler)
data_loader = torch.utils.data.DataLoader(
dataset,
batch_size=args.batch_size,
sampler=train_sampler,
num_workers=args.workers,
pin_memory=True,
collate_fn=collate_fn,
)
data_loader_test = torch.utils.data.DataLoader(
dataset_test,
batch_size=args.batch_size,
sampler=test_sampler,
num_workers=args.workers,
pin_memory=True,
collate_fn=collate_fn,
)
print("Creating model")
model = torchvision.models.video.__dict__[args.model](weights=args.weights)
model.to(device)
if args.distributed and args.sync_bn:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
criterion = nn.CrossEntropyLoss()
lr = args.lr * args.world_size
optimizer = torch.optim.SGD(model.parameters(), lr=lr, momentum=args.momentum, weight_decay=args.weight_decay)
scaler = torch.cuda.amp.GradScaler() if args.amp else None
# convert scheduler to be per iteration, not per epoch, for warmup that lasts
# between different epochs
iters_per_epoch = len(data_loader)
lr_milestones = [iters_per_epoch * (m - args.lr_warmup_epochs) for m in args.lr_milestones]
main_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=lr_milestones, gamma=args.lr_gamma)
if args.lr_warmup_epochs > 0:
warmup_iters = iters_per_epoch * args.lr_warmup_epochs
args.lr_warmup_method = args.lr_warmup_method.lower()
if args.lr_warmup_method == "linear":
warmup_lr_scheduler = torch.optim.lr_scheduler.LinearLR(
optimizer, start_factor=args.lr_warmup_decay, total_iters=warmup_iters
)
elif args.lr_warmup_method == "constant":
warmup_lr_scheduler = torch.optim.lr_scheduler.ConstantLR(
optimizer, factor=args.lr_warmup_decay, total_iters=warmup_iters
)
else:
raise RuntimeError(
f"Invalid warmup lr method '{args.lr_warmup_method}'. Only linear and constant are supported."
)
lr_scheduler = torch.optim.lr_scheduler.SequentialLR(
optimizer, schedulers=[warmup_lr_scheduler, main_lr_scheduler], milestones=[warmup_iters]
)
else:
lr_scheduler = main_lr_scheduler
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
model_without_ddp = model.module
if args.resume:
checkpoint = torch.load(args.resume, map_location="cpu")
model_without_ddp.load_state_dict(checkpoint["model"])
optimizer.load_state_dict(checkpoint["optimizer"])
lr_scheduler.load_state_dict(checkpoint["lr_scheduler"])
args.start_epoch = checkpoint["epoch"] + 1
if args.amp:
scaler.load_state_dict(checkpoint["scaler"])
if args.test_only:
evaluate(model, criterion, data_loader_test, device=device)
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
train_one_epoch(model, criterion, optimizer, lr_scheduler, data_loader, device, epoch, args.print_freq, scaler)
evaluate(model, criterion, data_loader_test, device=device)
if args.output_dir:
checkpoint = {
"model": model_without_ddp.state_dict(),
"optimizer": optimizer.state_dict(),
"lr_scheduler": lr_scheduler.state_dict(),
"epoch": epoch,
"args": args,
}
if args.amp:
checkpoint["scaler"] = scaler.state_dict()
utils.save_on_master(checkpoint, os.path.join(args.output_dir, f"model_{epoch}.pth"))
utils.save_on_master(checkpoint, os.path.join(args.output_dir, "checkpoint.pth"))
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print(f"Training time {total_time_str}")
def main(args):
if args.apex and amp is None:
raise RuntimeError(
"Failed to import apex. Please install apex from https://www.github.com/nvidia/apex "
"to enable mixed-precision training.")
if args.output_dir:
utils.mkdir(args.output_dir)
utils.init_distributed_mode(args)
log_string(args)
device = torch.device(args.device)
torch.backends.cudnn.benchmark = True
train_dir = os.path.join(args.data_path, 'train')
val_dir = os.path.join(args.data_path, 'val')
dataset, dataset_test, train_sampler, test_sampler = load_data(
train_dir, val_dir, args.cache_dataset, args.distributed)
data_loader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
sampler=train_sampler,
num_workers=args.workers,
pin_memory=True)
data_loader_test = torch.utils.data.DataLoader(dataset_test,
batch_size=args.batch_size,
sampler=test_sampler,
num_workers=args.workers,
pin_memory=True)
log_string("Creating model")
# model = torchvision.models.__dict__[args.model](pretrained=args.pretrained)
if args.model == 'resnet50':
model = resnet50(pretrained=False)
elif args.model == 'resnet101':
model = resnet101(pretrained=False)
model.to(device)
if args.distributed and args.sync_bn:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.apex:
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.apex_opt_level)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=args.lr_step_size,
gamma=args.lr_gamma)
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
if args.resume and not args.test_only:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.test_only:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint)
evaluate(model, criterion, data_loader_test, device=device)
return
log_string("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
train_one_epoch(model, criterion, optimizer, data_loader, device,
epoch, args.print_freq, args.apex)
lr_scheduler.step()
evaluate(model, criterion, data_loader_test, device=device)
if args.output_dir:
checkpoint = {
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
'args': args
}
utils.save_on_master(
checkpoint,
os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
utils.save_on_master(
checkpoint, os.path.join(args.output_dir, 'checkpoint.pth'))
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
log_string('Training time {}'.format(total_time_str))
def main(args):
if args.output_dir:
utils.mkdir(args.output_dir)
utils.init_distributed_mode(args)
print(args)
device = torch.device(args.device)
if args.use_deterministic_algorithms:
torch.backends.cudnn.benchmark = False
torch.use_deterministic_algorithms(True)
else:
torch.backends.cudnn.benchmark = True
dataset, num_classes = get_dataset(args.data_path, args.dataset, "train",
get_transform(True, args))
dataset_test, _ = get_dataset(args.data_path, args.dataset, "val",
get_transform(False, args))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
dataset)
test_sampler = torch.utils.data.distributed.DistributedSampler(
dataset_test, shuffle=False)
else:
train_sampler = torch.utils.data.RandomSampler(dataset)
test_sampler = torch.utils.data.SequentialSampler(dataset_test)
data_loader = torch.utils.data.DataLoader(
dataset,
batch_size=args.batch_size,
sampler=train_sampler,
num_workers=args.workers,
collate_fn=utils.collate_fn,
drop_last=True,
)
data_loader_test = torch.utils.data.DataLoader(dataset_test,
batch_size=1,
sampler=test_sampler,
num_workers=args.workers,
collate_fn=utils.collate_fn)
model = torchvision.models.segmentation.__dict__[args.model](
weights=args.weights,
weights_backbone=args.weights_backbone,
num_classes=num_classes,
aux_loss=args.aux_loss)
model.to(device)
if args.distributed:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
params_to_optimize = [
{
"params": [
p for p in model_without_ddp.backbone.parameters()
if p.requires_grad
]
},
{
"params": [
p for p in model_without_ddp.classifier.parameters()
if p.requires_grad
]
},
]
if args.aux_loss:
params = [
p for p in model_without_ddp.aux_classifier.parameters()
if p.requires_grad
]
params_to_optimize.append({"params": params, "lr": args.lr * 10})
optimizer = torch.optim.SGD(params_to_optimize,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
scaler = torch.cuda.amp.GradScaler() if args.amp else None
iters_per_epoch = len(data_loader)
main_lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, lambda x: (1 - x /
(iters_per_epoch *
(args.epochs - args.lr_warmup_epochs)))**0.9)
if args.lr_warmup_epochs > 0:
warmup_iters = iters_per_epoch * args.lr_warmup_epochs
args.lr_warmup_method = args.lr_warmup_method.lower()
if args.lr_warmup_method == "linear":
warmup_lr_scheduler = torch.optim.lr_scheduler.LinearLR(
optimizer,
start_factor=args.lr_warmup_decay,
total_iters=warmup_iters)
elif args.lr_warmup_method == "constant":
warmup_lr_scheduler = torch.optim.lr_scheduler.ConstantLR(
optimizer,
factor=args.lr_warmup_decay,
total_iters=warmup_iters)
else:
raise RuntimeError(
f"Invalid warmup lr method '{args.lr_warmup_method}'. Only linear and constant are supported."
)
lr_scheduler = torch.optim.lr_scheduler.SequentialLR(
optimizer,
schedulers=[warmup_lr_scheduler, main_lr_scheduler],
milestones=[warmup_iters])
else:
lr_scheduler = main_lr_scheduler
if args.resume:
checkpoint = torch.load(args.resume, map_location="cpu")
model_without_ddp.load_state_dict(checkpoint["model"],
strict=not args.test_only)
if not args.test_only:
optimizer.load_state_dict(checkpoint["optimizer"])
lr_scheduler.load_state_dict(checkpoint["lr_scheduler"])
args.start_epoch = checkpoint["epoch"] + 1
if args.amp:
scaler.load_state_dict(checkpoint["scaler"])
if args.test_only:
# We disable the cudnn benchmarking because it can noticeably affect the accuracy
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
confmat = evaluate(model,
data_loader_test,
device=device,
num_classes=num_classes)
print(confmat)
return
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
train_one_epoch(model, criterion, optimizer, data_loader, lr_scheduler,
device, epoch, args.print_freq, scaler)
confmat = evaluate(model,
data_loader_test,
device=device,
num_classes=num_classes)
print(confmat)
checkpoint = {
"model": model_without_ddp.state_dict(),
"optimizer": optimizer.state_dict(),
"lr_scheduler": lr_scheduler.state_dict(),
"epoch": epoch,
"args": args,
}
if args.amp:
checkpoint["scaler"] = scaler.state_dict()
utils.save_on_master(
checkpoint, os.path.join(args.output_dir, f"model_{epoch}.pth"))
utils.save_on_master(checkpoint,
os.path.join(args.output_dir, "checkpoint.pth"))
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print(f"Training time {total_time_str}")
def main(args):
if args.model_dir:
utils.mkdir(args.model_dir)
utils.init_distributed_mode(args)
print(args)
device = torch.device(args.device)
# determine the desired resolution
resolution = (args.resolution, args.resolution)
if "width" in args and "height" in args:
resolution = (args.height, args.width)
# load the train and val datasets
dataset, num_classes = get_dataset(
args.dataset, args.data, "train",
get_transform(train=True, resolution=resolution))
dataset_test, _ = get_dataset(
args.dataset, args.data, "val",
get_transform(train=False, resolution=resolution))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
dataset)
test_sampler = torch.utils.data.distributed.DistributedSampler(
dataset_test)
else:
train_sampler = torch.utils.data.RandomSampler(dataset)
test_sampler = torch.utils.data.SequentialSampler(dataset_test)
data_loader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
sampler=train_sampler,
num_workers=args.workers,
collate_fn=utils.collate_fn,
drop_last=True)
data_loader_test = torch.utils.data.DataLoader(dataset_test,
batch_size=1,
sampler=test_sampler,
num_workers=args.workers,
collate_fn=utils.collate_fn)
print("=> training with dataset: '{:s}' (train={:d}, val={:d})".format(
args.dataset, len(dataset), len(dataset_test)))
print("=> training with resolution: {:d}x{:d}, {:d} classes".format(
resolution[1], resolution[0], num_classes))
print("=> training with model: {:s}".format(args.arch))
# create the segmentation model
model = torchvision.models.segmentation.__dict__[args.arch](
num_classes=num_classes,
aux_loss=args.aux_loss,
pretrained=args.pretrained)
model.to(device)
if args.distributed:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.resume:
checkpoint = torch.load(args.resume, map_location='cpu')
model.load_state_dict(checkpoint['model'])
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
# eval-only mode
if args.test_only:
confmat = evaluate(model,
data_loader_test,
device=device,
num_classes=num_classes)
print(confmat)
return
# create the optimizer
params_to_optimize = [
{
"params": [
p for p in model_without_ddp.backbone.parameters()
if p.requires_grad
]
},
{
"params": [
p for p in model_without_ddp.classifier.parameters()
if p.requires_grad
]
},
]
if args.aux_loss:
params = [
p for p in model_without_ddp.aux_classifier.parameters()
if p.requires_grad
]
params_to_optimize.append({"params": params, "lr": args.lr * 10})
optimizer = torch.optim.SGD(params_to_optimize,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, lambda x: (1 - x / (len(data_loader) * args.epochs))**0.9)
# training loop
start_time = time.time()
best_IoU = 0.0
for epoch in range(args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
# train the model over the next epoc
train_one_epoch(model, criterion, optimizer, data_loader, lr_scheduler,
device, epoch, args.print_freq)
# test the model on the val dataset
confmat = evaluate(model,
data_loader_test,
device=device,
num_classes=num_classes)
print(confmat)
# save model checkpoint
checkpoint_path = os.path.join(args.model_dir,
'model_{}.pth'.format(epoch))
utils.save_on_master(
{
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': epoch,
'args': args,
'arch': args.arch,
'dataset': args.dataset,
'num_classes': num_classes,
'resolution': resolution,
'accuracy': confmat.acc_global,
'mean_IoU': confmat.mean_IoU
}, checkpoint_path)
print(
'saved checkpoint to: {:s} ({:.3f}% mean IoU, {:.3f}% accuracy)'.
format(checkpoint_path, confmat.mean_IoU, confmat.acc_global))
if confmat.mean_IoU > best_IoU:
best_IoU = confmat.mean_IoU
best_path = os.path.join(args.model_dir, 'model_best.pth')
shutil.copyfile(checkpoint_path, best_path)
print(
'saved best model to: {:s} ({:.3f}% mean IoU, {:.3f}% accuracy)'
.format(best_path, best_IoU, confmat.acc_global))
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)
print(args)
device = torch.device(args.device)
# Data loading code
print("Loading data")
img_dim = parse_resize(args.resize)
print(img_dim)
dataset, num_classes = get_dataset(args.image_dir, args.train_json_path,
img_dim)
dataset_test, _ = get_dataset(args.image_dir, args.val_json_path, img_dim)
writer = SummaryWriter(logdir=args.output_dir)
best_map = np.NINF
print("Creating data loaders")
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
dataset)
test_sampler = torch.utils.data.distributed.DistributedSampler(
dataset_test)
else:
train_sampler = torch.utils.data.RandomSampler(dataset)
test_sampler = torch.utils.data.SequentialSampler(dataset_test)
train_batch_sampler = torch.utils.data.BatchSampler(train_sampler,
args.batch_size,
drop_last=True)
data_loader = torch.utils.data.DataLoader(
dataset,
batch_sampler=train_batch_sampler,
num_workers=args.workers,
collate_fn=utils.collate_fn,
)
data_loader_test = torch.utils.data.DataLoader(
dataset_test,
batch_size=1,
sampler=test_sampler,
num_workers=args.workers,
collate_fn=utils.collate_fn,
)
print("Creating model")
model = torchvision.models.detection.__dict__[args.model](
num_classes=num_classes + 1, pretrained=args.pretrained)
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
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.lr_step_size, gamma=args.lr_gamma)
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer, milestones=args.lr_steps, gamma=args.lr_gamma)
if args.resume:
checkpoint = torch.load(args.resume, map_location="cpu")
model_without_ddp.load_state_dict(checkpoint["model"])
optimizer.load_state_dict(checkpoint["optimizer"])
lr_scheduler.load_state_dict(checkpoint["lr_scheduler"])
args.start_epoch = checkpoint["epoch"] + 1
if args.test_only:
evaluate(model, data_loader_test, device=device)
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
train_one_epoch(model, optimizer, data_loader, device, epoch,
args.print_freq, writer)
lr_scheduler.step()
# evaluate after every epoch
stats = evaluate(model, data_loader_test, device=device)
map_avg, map_50, map_75, map_small, map_medium, map_large = stats[:6]
if utils.is_main_process():
writer.add_scalar("eval/[email protected]:0.95", map_avg, epoch)
writer.add_scalar("eval/[email protected]", map_50, epoch)
writer.add_scalar("eval/[email protected]", map_75, epoch)
writer.add_scalar("eval/map_small", map_small, epoch)
writer.add_scalar("eval/map_medium", map_medium, epoch)
writer.add_scalar("eval/map_large", map_large, epoch)
if args.output_dir:
if map_50 > best_map:
utils.save_on_master(
{
"model": model_without_ddp.state_dict(),
"optimizer": optimizer.state_dict(),
"lr_scheduler": lr_scheduler.state_dict(),
"args": args,
"epoch": epoch,
},
os.path.join(args.output_dir,
"model_{}.pth".format(epoch)),
)
best_map = map_50
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print("Training time {}".format(total_time_str))