def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
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)
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'])
# standard PyTorch mean-std input image normalization
transform = T.Compose([
T.Resize(800),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
pimg = Image.open(args.source).convert('RGB')
t0 = time.time()
scores, boxes = detect(pimg, model, transform, device)
print(f'inference use {time.time() - t0} seconds.')
plot_results(pimg, scores, boxes)
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
# dataset_train = build_dataset(image_set='train', args=args)
dataset_val = build_dataset(image_set='val', args=args)
# if args.distributed:
# sampler_train = DistributedSampler(dataset_train)
# sampler_val = DistributedSampler(dataset_val, shuffle=False)
# else:
# sampler_train = torch.utils.data.RandomSampler(dataset_train)
# sampler_val = torch.utils.data.SequentialSampler(dataset_val)
# batch_sampler_train = torch.utils.data.BatchSampler(
# sampler_train, args.batch_size, drop_last=True)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
# data_loader_train = DataLoader(dataset_train, batch_sampler=batch_sampler_train,
# collate_fn=utils.collate_fn, num_workers=args.num_workers)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn_leimao,
num_workers=args.num_workers)
for inputs, labels in data_loader_val:
print("---------------------")
print(inputs.shape)
print(labels)
def load_model(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
# IPython.embed()
# IPython.embed()
os.system("sudo chmod -R 777 /home/shuxuang/.cache/")
model, criterion, postprocessors = build_model(
args) # use the same model as detr paper on coco
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)
# optimizer = torch.optim.AdamW(param_dicts, lr=args.lr,
# weight_decay=args.weight_decay)
# lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
# dataset_train = build_dataset(image_set='train', args=args)
# dataset_val = build_dataset(image_set='val', args=args)
# modify the dataset from coco to nvdata
# home_dir = os.environ["HOME"]
# # on local
# dataset_train_ = build_nvdataset(dataset_root=[
# os.path.join(os.environ["HOME"],'datasets/annotation_sql_nvidia'),
# os.path.join(os.environ["HOME"], 'datasets/frames_nvidia')],
# mode='train')
dataset_val = build_nvdataset(dataset_root=[
os.path.join(os.environ["HOME"], 'datasets/test'),
os.path.join(os.environ["HOME"], 'datasets/frames_nvidia')
],
mode='test')
# indices_50k =np.load(os.path.join(os.environ["HOME"],'datasets/id_1_criterion_Max_SSD_num_labels_50000.npy'))
# # on maglev
# dataset_train_ = build_nvdataset(dataset_root=[args.dataset_root_sql, args.dataset_root_img],
# mode='train')
# dataset_val = build_nvdataset(dataset_root=[args.dataset_root_test, args.dataset_root_sql],
# mode='test')
# indices_50k =np.load(os.path.join(args.root_indices))
# dataset_train = Subset(dataset_train_, indices_50k)
print("Validation samples: %d" % (len(dataset_val)))
# IPython.embed()
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
output_dir = Path(args.output_dir)
# args.resume = os.path.join(os.environ["HOME"], 'datasets/exps_detr_base/checkpoint0299.pth')
args.resume = '/home/shuxuang/datasets/exps_detr_base/checkpoint0299.pth'
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:
print('Loading model: %s' % args.resume)
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
# if args.eval:
# evaluate(model, dataset_val, postprocessors, device)
return model, dataset_val, postprocessors, device
# if __name__ == '__main__':
# parser = argparse.ArgumentParser('DETR training and evaluation script', parents=[get_args_parser()])
# args = parser.parse_args()
# if args.output_dir:
# Path(args.output_dir).mkdir(parents=True, exist_ok=True)
# main(args)
# CUDA_VISIBLE_DEVICES=1 dazel run //sandbox/williamz/detr:train_with_eval -- --eval
# test:
# CUDA_VISIBLE_DEVICES=1 dazel run //sandbox/williamz/detr:eval -- --eval --resume /home/shuxuang/experiments/train_output/checkpoint.pth
# get info:
# maglev workflows get 0cf25940-3f00-5c2c-a8e8-1571e986513b
# maglev volumes mount -n train-outputs -v 4977fea-0998-4d5b-b557-ff17605f2098 -p /home/shuxuang/experiments/train_output
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
print(args)
# Save our Wandb metadata
if not args.no_wb:
wandb.init(entity='dl-project',
project='dl-final-project',
name=args.wb_name,
notes=args.wb_notes,
reinit=True)
wandb.config.epochs = args.epochs
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)
model, criterion, postprocessors = build_model(args)
model.to(device)
# visualize_video(model, postprocessors)
model_without_ddp = model
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of trainable params:', n_parameters)
wandb.config.n_parameters = n_parameters
wandb.config.n_trainable_parameters = n_parameters # better name
# Log total # of model parameters (including frozen) to W&B
n_total_parameters = sum(p.numel() for p in model.parameters())
print('total number of parameters:', n_total_parameters)
wandb.config.n_total_parameters = n_total_parameters
dataset_train = build_dataset(image_set='train', args=args)
dataset_val = build_dataset(image_set='val', args=args)
# For visualization we want the raw images without any normalization or random resizing
dataset_val_without_resize = CocoDetection(
"data/coco/val2017",
annFile="data/coco/annotations/instances_val2017.json",
transforms=T.Compose([T.ToTensor()]))
# Save metadata about training + val datasets and batch size
wandb.config.len_dataset_train = len(dataset_train)
wandb.config.len_dataset_val = len(dataset_val)
wandb.config.batch_size = args.batch_size
if args.distributed:
if args.cache_mode:
sampler_train = samplers.NodeDistributedSampler(dataset_train)
sampler_val = samplers.NodeDistributedSampler(dataset_val,
shuffle=False)
else:
sampler_train = samplers.DistributedSampler(dataset_train)
sampler_val = samplers.DistributedSampler(dataset_val,
shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
pin_memory=True)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
pin_memory=True)
# lr_backbone_names = ["backbone.0", "backbone.neck", "input_proj", "transformer.encoder"]
def match_name_keywords(n, name_keywords):
out = False
for b in name_keywords:
if b in n:
out = True
break
return out
for n, p in model_without_ddp.named_parameters():
print(n)
param_dicts = [{
"params": [
p for n, p in model_without_ddp.named_parameters()
if not match_name_keywords(n, args.lr_backbone_names)
and not match_name_keywords(n, args.lr_linear_proj_names)
and p.requires_grad
],
"lr":
args.lr,
}, {
"params": [
p for n, p in model_without_ddp.named_parameters() if
match_name_keywords(n, args.lr_backbone_names) and p.requires_grad
],
"lr":
args.lr_backbone,
}, {
"params": [
p for n, p in model_without_ddp.named_parameters()
if match_name_keywords(n, args.lr_linear_proj_names)
and p.requires_grad
],
"lr":
args.lr * args.lr_linear_proj_mult,
}]
# Not sure if we should save all hyperparameters in wandb.config?
# just start with a few important ones
wandb.config.lr = args.lr
wandb.config.lr_backbone = args.lr_backbone
if args.sgd:
optimizer = torch.optim.SGD(param_dicts,
lr=args.lr,
momentum=0.9,
weight_decay=args.weight_decay)
else:
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.coco.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
else:
base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
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')
missing_keys, unexpected_keys = model_without_ddp.load_state_dict(
checkpoint['model'], strict=False)
unexpected_keys = [
k for k in unexpected_keys
if not (k.endswith('total_params') or k.endswith('total_ops'))
]
if len(missing_keys) > 0:
print('Missing Keys: {}'.format(missing_keys))
if len(unexpected_keys) > 0:
print('Unexpected Keys: {}'.format(unexpected_keys))
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
import copy
p_groups = copy.deepcopy(optimizer.param_groups)
optimizer.load_state_dict(checkpoint['optimizer'])
for pg, pg_old in zip(optimizer.param_groups, p_groups):
pg['lr'] = pg_old['lr']
pg['initial_lr'] = pg_old['initial_lr']
# print(optimizer.param_groups)
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
# todo: this is a hack for doing experiment that resume from checkpoint and also modify lr scheduler (e.g., decrease lr in advance).
args.override_resumed_lr_drop = True
if args.override_resumed_lr_drop:
print(
'Warning: (hack) args.override_resumed_lr_drop is set to True, so args.lr_drop would override lr_drop in resumed lr_scheduler.'
)
lr_scheduler.step_size = args.lr_drop
lr_scheduler.base_lrs = list(
map(lambda group: group['initial_lr'],
optimizer.param_groups))
lr_scheduler.step(lr_scheduler.last_epoch)
args.start_epoch = checkpoint['epoch'] + 1
# check the resumed model
if not args.eval:
test_stats, coco_evaluator = evaluate(model, criterion,
postprocessors,
data_loader_val, base_ds,
device, args.output_dir)
if args.eval:
print("Generating visualizations...")
visualize_bbox(model, postprocessors, data_loader_val, device,
dataset_val_without_resize)
test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device,
args.output_dir)
if args.output_dir:
utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval.pth")
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_file_for_wb = str(
output_dir / f'{wandb.run.id}_checkpoint{epoch:04}.pth')
checkpoint_paths = [
output_dir / 'checkpoint.pth', checkpoint_file_for_wb
]
# extra checkpoint before LR drop and every 5 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 5 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
# Save model checkpoint to W&B
wandb.save(checkpoint_file_for_wb)
# Generate visualizations for fixed(?) set of images every epoch
print("Generating visualizations...")
visualize_bbox(model, postprocessors, data_loader_val, device,
dataset_val_without_resize)
test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device,
args.output_dir)
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
}
# Save the COCO metrics properly
metric_name = [
"AP", "AP50", "AP75", "APs", "APm", "APl", "AR@1", "AR@10",
"AR@100", "ARs", "ARm", "ARl"
]
for i, metric_val in enumerate(log_stats["test_coco_eval_bbox"]):
log_stats[metric_name[i]] = metric_val
if not args.no_wb:
wandb.log(log_stats)
print("train_loss: ", log_stats['train_loss'])
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")
wandb.save(str(output_dir / "log.txt"))
# for evaluation logs
if coco_evaluator is not None:
(output_dir / 'eval').mkdir(exist_ok=True)
if "bbox" in coco_evaluator.coco_eval:
filenames = ['latest.pth']
eval_filename_for_wb = f'{wandb.run.id}_eval_{epoch:04}.pth'
eval_path_for_wb = str(output_dir / "eval" /
eval_filename_for_wb)
filenames = ['latest.pth', eval_filename_for_wb]
if epoch % 50 == 0:
filenames.append(f'{epoch:03}.pth')
for name in filenames:
torch.save(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval" / name)
# TODO not sure if this file will end up being too big
# I think it's the COCO precision/recall metrics
# in some format...
# let's track it just in case to start!
wandb.save(eval_path_for_wb)
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)
if args.output_dir is None:
args.output_dir = os.path.expanduser(
'~/Data/AI2Thor_detection_features/')
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
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)
model, criterion, postprocessors = build_model(args)
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
dataset_all = build_dataset(image_set='all', args=args)
if args.distributed:
sampler_all = DistributedSampler(dataset_all, shuffle=False)
else:
sampler_all = torch.utils.data.SequentialSampler(dataset_all)
data_loader_all = DataLoader(dataset_all,
args.batch_size,
sampler=sampler_all,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
base_ds = get_coco_api_from_dataset(dataset_all)
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'], strict=False)
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
print("Start extracting features")
start_time = time.time()
extract_feature(model, criterion, postprocessors, data_loader_all, base_ds,
device, args.output_dir)
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Extracting features time {}'.format(total_time_str))
print('Start combining files')
start_time = time.time()
data_dir = os.path.expanduser('~/Data/AI2Thor_offline_data_2.0.2/')
combine_files(args, data_dir)
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Combining files time {}'.format(total_time_str))
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
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)
model, criterion = build_model(args)
model.to(device)
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)
param_dicts = [
{"params": [p for n, p in model_without_ddp.named_parameters() if "backbone" not in n and p.requires_grad]},
{
"params": [p for n, p in model_without_ddp.named_parameters() if "backbone" in n and p.requires_grad],
"lr": args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts, lr=args.lr, weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
dataset_train = build_dataset(image_set='train', args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train, args.batch_size, drop_last=True)
data_loader_train = DataLoader(dataset_train, batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn, num_workers=args.num_workers)
# Load from pretrained DETR model.
assert args.num_queries == 100, args.num_queries
assert args.enc_layers == 6 and args.dec_layers == 6
assert args.backbone in ['resnet50', 'resnet101', 'swin'], args.backbone
if args.backbone == 'resnet50':
pretrain_model = './data/detr_coco/detr-r50-e632da11.pth'
elif args.backbone == 'resnet101':
pretrain_model = './data/detr_coco/detr-r101-2c7b67e5.pth'
else:
pretrain_model = None
if pretrain_model is not None:
pretrain_dict = torch.load(pretrain_model, map_location='cpu')['model']
my_model_dict = model_without_ddp.state_dict()
pretrain_dict = {k: v for k, v in pretrain_dict.items() if k in my_model_dict}
my_model_dict.update(pretrain_dict)
model_without_ddp.load_state_dict(my_model_dict)
output_dir = Path(args.output_dir)
if args.resume:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
if '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
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(
model, criterion, data_loader_train, optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 10 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 100 == 0:
checkpoint_paths.append(output_dir / f'checkpoint{epoch:04}.pth')
if (epoch + 1) > args.lr_drop and (epoch + 1) % 10 == 0:
checkpoint_paths.append(output_dir / f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
'epoch': epoch,
'n_parameters': n_parameters}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
# IPython.embed()
# IPython.embed()
# os.system("sudo chmod -R 777 /home/shuxuang/.cache/")
model, criterion, postprocessors = build_model(
args) # use the same model as detr paper on coco
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
# dataset_train = build_dataset(image_set='train', args=args)
# dataset_val = build_dataset(image_set='val', args=args)
# modify the dataset from coco to nvdata
# home_dir = os.environ["HOME"]
# dataset_train_ = build_nvdataset(dataset_root=[
# os.path.join(os.environ["HOME"],'datasets/annotation_sql_nvidia'),
# os.path.join(os.environ["HOME"], 'datasets/frames_nvidia')],
# mode='train')
# dataset_val = build_nvdataset(dataset_root=[
# os.path.join(os.environ["HOME"],'datasets/test'),
# os.path.join(os.environ["HOME"], 'datasets/frames_nvidia')],
# mode='test')
# indices_50k =np.load(os.path.join(os.environ["HOME"],'datasets/id_1_criterion_Max_SSD_num_labels_50000.npy'))
dataset_train = build_nvdataset(
dataset_root=[args.dataset_root_sql, args.dataset_root_img],
mode='train',
camera=args.camera)
dataset_val = build_nvdataset(
dataset_root=[args.dataset_root_test, args.dataset_root_test],
mode='test',
camera=args.camera)
if args.root_indices is not None:
indices_50k = np.load(os.path.join(args.root_indices))
# indices_50k =np.load(os.path.join(os.environ["HOME"],'datasets/id_1_criterion_Max_SSD_num_labels_50000.npy'))
dataset_train = Subset(dataset_train, indices_50k)
# IPython.embed()
print("Train samples: %d" % (len(dataset_train)))
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
# if args.dataset_file == "coco_panoptic":
# # We also evaluate AP during panoptic training, on original coco DS
# coco_val = datasets.coco.build("val", args)
# base_ds = get_coco_api_from_dataset(coco_val)
# elif args.dataset_file == "nvdata":
# coco_val = datasets.coco.build("val", args)
# base_ds = get_coco_api_from_dataset(coco_val)
# else:
# base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
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.eval:
# test_stats, coco_evaluator = evaluate_nvdata(model, criterion, postprocessors,
# data_loader_val, base_ds, device, args.output_dir)
# if args.output_dir:
# utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval, output_dir / "eval.pth")
# return
# if args.eval:
# evaluate(model, dataset_val, postprocessors, device)
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 100 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 50 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
# test_stats, coco_evaluator = evaluate_nvdata(
# model, criterion, postprocessors, data_loader_val, base_ds, device, args.output_dir
# )
# 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}
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
# for evaluation logs
# if coco_evaluator is not None:
# (output_dir / 'eval').mkdir(exist_ok=True)
# if "bbox" in coco_evaluator.coco_eval:
# filenames = ['latest.pth']
# if epoch % 50 == 0:
# filenames.append(f'{epoch:03}.pth')
# for name in filenames:
# torch.save(coco_evaluator.coco_eval["bbox"].eval,
# output_dir / "eval" / name)
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("git:\n {}\n".format(utils.get_sha()))
wandb.init(project="qpic-project",
entity="sangbaeklee",
group="experiment_qpic")
wandb.config = {
"learning_rate": args.lr,
"epochs": args.epochs,
"batch_size": args.batch_size,
}
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
model.to(device)
wandb.watch(model)
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
dataset_train = build_dataset(image_set='train', args=args)
dataset_val = build_dataset(image_set='val', args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
if not args.hoi:
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.coco.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
else:
base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
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
elif args.pretrained:
checkpoint = torch.load(args.pretrained, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'], strict=False)
if args.eval:
if args.hoi:
test_stats = evaluate_hoi(args.dataset_file, model, postprocessors,
data_loader_val,
args.subject_category_id, device)
return
else:
test_stats, coco_evaluator = evaluate(model, criterion,
postprocessors,
data_loader_val, base_ds,
device, args.output_dir)
if args.output_dir:
utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval.pth")
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 100 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 100 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
if args.hoi:
test_stats = evaluate_hoi(args.dataset_file, model, postprocessors,
data_loader_val,
args.subject_category_id, device)
coco_evaluator = None
else:
test_stats, coco_evaluator = evaluate(model, criterion,
postprocessors,
data_loader_val, base_ds,
device, args.output_dir)
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
}
#import pdb; pdb.set_trace()
if args.dataset_file == 'hico':
wandb.log({
"loss": train_stats['loss'],
"mAP": test_stats['mAP'],
"mAP rare": test_stats['mAP rare'],
"mAP non-rare": test_stats['mAP non-rare'],
"mean max recall": test_stats['mean max recall']
})
elif args.dataset_file == 'vcoco':
wandb.log({
"mAP_all": test_stats['mAP_all'],
"mAP_thesis": test_stats['mAP_thesis'],
"AP_hold_obj": test_stats['AP_hold_obj'],
"AP_stand": test_stats['AP_stand'],
"AP_sit_instr": test_stats['AP_sit_instr'],
"AP_ride_instr": test_stats['AP_ride_instr'],
"AP_walk": test_stats['AP_walk'],
"AP_look_obj": test_stats['AP_look_obj'],
"AP_hit_instr": test_stats['AP_hit_instr'],
"AP_hit_obj": test_stats['AP_hit_obj'],
"AP_eat_obj": test_stats['AP_eat_obj'],
"AP_eat_instr": test_stats['AP_eat_instr'],
"AP_jump_instr": test_stats['AP_jump_instr'],
"AP_lay_instr": test_stats['AP_lay_instr'],
"AP_talk_on_phone_instr": test_stats['AP_talk_on_phone_instr'],
"AP_carry_obj": test_stats['AP_carry_obj'],
"AP_throw_obj": test_stats['AP_throw_obj'],
"AP_catch_obj": test_stats['AP_catch_obj'],
"AP_cut_instr": test_stats['AP_cut_instr'],
"AP_cut_obj": test_stats['AP_cut_obj'],
"AP_run": test_stats['AP_run'],
"AP_work_on_computer_instr": test_stats['AP_work_on_computer_instr'],
"AP_ski_instr": test_stats['AP_ski_instr'],
"AP_surf_instr": test_stats['AP_surf_instr'],
"AP_skateboard_instr": test_stats['AP_skateboard_instr'],
"AP_smile": test_stats['AP_smile'],
"AP_drink_instr": test_stats['AP_drink_instr'],
"AP_kick_obj": test_stats['AP_kick_obj'],
"AP_point_instr": test_stats['AP_point_instr'],
"AP_read_obj": test_stats['AP_read_obj'],
"AP_snowboard_instr": test_stats['AP_snowboard_instr'],\
"loss" : train_stats['loss']
})
else:
continue
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")
# for evaluation logs
if coco_evaluator is not None:
(output_dir / 'eval').mkdir(exist_ok=True)
if "bbox" in coco_evaluator.coco_eval:
filenames = ['latest.pth']
if epoch % 50 == 0:
filenames.append(f'{epoch:03}.pth')
for name in filenames:
torch.save(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval" / name)
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("git:\n {}\n".format(utils.get_sha()))
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)
model, criterion, postprocessors = build_model(args)
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
# no validation ground truth for ytvos dataset
dataset_train = build_dataset(image_set='train', args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
output_dir = Path(args.output_dir)
# load coco pretrained weight
checkpoint = torch.load(args.pretrained_weights,
map_location='cpu')['model']
del checkpoint["vistr.class_embed.weight"]
del checkpoint["vistr.class_embed.bias"]
del checkpoint["vistr.query_embed.weight"]
model.module.load_state_dict(checkpoint, strict=False)
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
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 1 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
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)
# fix the seed for reproducibility
seed = args.seed + utils.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
model, criterion, postprocessors = build_model(args)
model.to(device)
model_without_ddp = model
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
image_set = 'fewshot' if args.fewshot_finetune else 'train'
dataset_train = build_dataset(image_set=image_set, args=args)
dataset_val = build_dataset(image_set='val', args=args)
dataset_support = build_support_dataset(image_set=image_set, args=args)
if args.distributed:
if args.cache_mode:
sampler_train = samplers.NodeDistributedSampler(dataset_train)
sampler_val = samplers.NodeDistributedSampler(dataset_val,
shuffle=False)
sampler_support = samplers.NodeDistributedSampler(dataset_support)
else:
sampler_train = samplers.DistributedSampler(dataset_train)
sampler_val = samplers.DistributedSampler(dataset_val,
shuffle=False)
sampler_support = samplers.DistributedSampler(dataset_support)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
sampler_support = torch.utils.data.RandomSampler(dataset_support)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=False)
loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
pin_memory=True)
loader_val = DataLoader(dataset_val,
batch_size=args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
pin_memory=True)
loader_support = DataLoader(dataset_support,
batch_size=1,
sampler=sampler_support,
drop_last=False,
num_workers=args.num_workers,
pin_memory=False)
def match_name_keywords(n, name_keywords):
out = False
for b in name_keywords:
if b in n:
out = True
break
return out
for n, p in model_without_ddp.named_parameters():
print(n)
if not args.fewshot_finetune:
param_dicts = [{
"params": [
p for n, p in model_without_ddp.named_parameters()
if not match_name_keywords(n, args.lr_backbone_names)
and not match_name_keywords(n, args.lr_linear_proj_names)
and p.requires_grad
],
"lr":
args.lr,
"initial_lr":
args.lr,
}, {
"params": [
p for n, p in model_without_ddp.named_parameters()
if match_name_keywords(n, args.lr_backbone_names)
and p.requires_grad
],
"lr":
args.lr_backbone,
"initial_lr":
args.lr_backbone,
}, {
"params": [
p for n, p in model_without_ddp.named_parameters()
if match_name_keywords(n, args.lr_linear_proj_names)
and p.requires_grad
],
"lr":
args.lr * args.lr_linear_proj_mult,
"initial_lr":
args.lr * args.lr_linear_proj_mult,
}]
else:
# For few-shot finetune stage, do not train sampling offsets, reference points, and embedding related parameters
param_dicts = [
{
"params":
[p for n, p in model_without_ddp.named_parameters()
if not match_name_keywords(n, args.lr_backbone_names) and \
not match_name_keywords(n, args.lr_linear_proj_names) and \
not match_name_keywords(n, args.embedding_related_names) and p.requires_grad],
"lr": args.lr,
"initial_lr": args.lr,
},
{
"params": [p for n, p in model_without_ddp.named_parameters() if match_name_keywords(n, args.lr_backbone_names) and p.requires_grad],
"lr": args.lr_backbone,
"initial_lr": args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts, weight_decay=args.weight_decay)
lr_scheduler = WarmupMultiStepLR(optimizer,
args.lr_drop_milestones,
gamma=0.1,
warmup_epochs=args.warmup_epochs,
warmup_factor=args.warmup_factor,
warmup_method='linear',
last_epoch=args.start_epoch - 1)
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu], find_unused_parameters=True)
model_without_ddp = model.module
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.dataset.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
else:
base_ds = get_coco_api_from_dataset(dataset_val)
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')
missing_keys, unexpected_keys = model_without_ddp.load_state_dict(
checkpoint['model'], strict=False)
unexpected_keys = [
k for k in unexpected_keys
if not (k.endswith('total_params') or k.endswith('total_ops'))
]
if len(missing_keys) > 0:
print('Missing Keys: {}'.format(missing_keys))
if len(unexpected_keys) > 0:
print('Unexpected Keys: {}'.format(unexpected_keys))
if args.fewshot_finetune:
if args.category_codes_cls_loss:
# Re-init weights of novel categories for few-shot finetune
novel_class_ids = datasets.get_class_ids(args.dataset_file,
type='novel')
if args.num_feature_levels == 1:
for novel_class_id in novel_class_ids:
nn.init.normal_(model_without_ddp.category_codes_cls.L.
weight[novel_class_id])
elif args.num_feature_levels > 1:
for classifier in model_without_ddp.category_codes_cls:
for novel_class_id in novel_class_ids:
nn.init.normal_(
classifier.L.weight[novel_class_id])
else:
raise RuntimeError
if args.eval:
# Evaluate only base categories
test_stats, coco_evaluator = evaluate(args,
model,
criterion,
postprocessors,
loader_val,
loader_support,
base_ds,
device,
type='base')
if args.output_dir:
utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval_base.pth")
# Evaluate only novel categories
test_stats, coco_evaluator = evaluate(args,
model,
criterion,
postprocessors,
loader_val,
loader_support,
base_ds,
device,
type='novel')
if args.output_dir:
utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval_novel.pth")
return
print("Start training...")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(args, model, criterion, loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
# Saving Checkpoints after each epoch
if args.output_dir and (not args.fewshot_finetune):
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,
'args': args,
}, checkpoint_path)
# Saving Checkpoints every args.save_every_epoch epoch(s)
if args.output_dir:
checkpoint_paths = []
if (epoch + 1) % args.save_every_epoch == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
# Evaluation and Logging
if (epoch + 1) % args.eval_every_epoch == 0:
if 'base' in args.dataset_file:
evaltype = 'base'
else:
evaltype = 'all'
if args.fewshot_finetune:
evaltype = 'novel'
test_stats, coco_evaluator = evaluate(args,
model,
criterion,
postprocessors,
loader_val,
loader_support,
base_ds,
device,
type=evaltype)
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,
'evaltype': evaltype
}
if args.output_dir and utils.is_main_process():
with (output_dir / "results.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
# for evaluation logs
if coco_evaluator is not None:
(output_dir / 'eval').mkdir(exist_ok=True)
if "bbox" in coco_evaluator.coco_eval:
filenames = ['latest.pth']
filenames.append(f'{epoch:03}.pth')
for name in filenames:
torch.save(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval" / name)
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("git:\n {}\n".format(utils.get_sha()))
output_dir = Path(args.output_dir)
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)
model, criterion, postprocessors = build_model(args)
model.to(device)
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)
param_dicts = [
{"params": [p for n, p in model_without_ddp.named_parameters() if "backbone" not in n and p.requires_grad]},
{
"params": [p for n, p in model_without_ddp.named_parameters() if "backbone" in n and p.requires_grad],
"lr": args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts, lr=args.lr, weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
if args.eval:
dataset_val = build_dataset(image_set=args.dataset, args=args)
if args.distributed:
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
data_loader_val = DataLoader(dataset_val, args.batch_size, sampler=sampler_val, drop_last=False, collate_fn=utils.collate_fn, num_workers=args.num_workers)
else:
dataset_train = build_dataset(image_set='train', args=args)
dataset_val = build_dataset(image_set='val', args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.SequentialSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train, args.batch_size, drop_last=True)
data_loader_train = DataLoader(dataset_train, batch_sampler=batch_sampler_train, collate_fn=utils.collate_fn, num_workers=args.num_workers)
data_loader_val = DataLoader(dataset_val, args.batch_size, sampler=sampler_val, drop_last=False, collate_fn=utils.collate_fn, num_workers=args.num_workers)
base_ds = get_coco_api_from_dataset(dataset_val)
if args.resume and args.frozen_weights:
assert False
elif args.resume:
if args.resume.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.resume, map_location='cpu', check_hash=True)
new_state_dict = {}
for k in checkpoint['model']:
if ("class_embed" in k) or ("bbox_embed" in k) or ("query_embed" in k):
continue
if ("input_proj" in k) and args.layer1_num != 3:
continue
new_state_dict[k] = checkpoint['model'][k]
# Compare load model and current model
current_param = [n for n,p in model_without_ddp.named_parameters()]
current_buffer = [n for n,p in model_without_ddp.named_buffers()]
load_param = new_state_dict.keys()
for p in load_param:
if p not in current_param and p not in current_buffer:
print(p, 'NOT appear in current model. ')
for p in current_param:
if p not in load_param:
print(p, 'NEW parameter. ')
model_without_ddp.load_state_dict(new_state_dict, strict=False)
else:
checkpoint = torch.load(args.resume, map_location='cpu')
# this is to compromise old implementation
new_state_dict = {}
for k in checkpoint['model']:
if "bbox_embed" in k:
print("bbox_embed from OLD implementation has been replaced with lines_embed")
new_state_dict["lines_embed."+'.'.join(k.split('.')[1:])] = checkpoint['model'][k]
else:
new_state_dict[k] = checkpoint['model'][k]
# compare resume model and current model
current_param = [n for n,p in model_without_ddp.named_parameters()]
current_buffer = [n for n,p in model_without_ddp.named_buffers()]
load_param = new_state_dict.keys()
#for p in load_param:
#if p not in current_param and p not in current_buffer:
#print(p, 'not been loaded to current model. Strict == False?')
for p in current_param:
if p not in load_param:
print(p, 'is a new parameter. Not found from load dict.')
# load model
model_without_ddp.load_state_dict(new_state_dict)
# load optimizer
if not args.no_opt and not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer'])
checkpoint['lr_scheduler']['step_size'] = args.lr_drop # change the lr_drop epoch
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
elif args.frozen_weights:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
new_state_dict = {}
for k in checkpoint['model']:
if "bbox_embed" in k:
new_state_dict["lines_embed."+'.'.join(k.split('.')[1:])] = checkpoint['model'][k]
else:
new_state_dict[k] = checkpoint['model'][k]
model_without_ddp.letr.load_state_dict(new_state_dict)
# params
encoder = {k:v for k,v in new_state_dict.items() if "encoder" in k}
decoder = {k:v for k,v in new_state_dict.items() if "decoder" in k}
class_embed = {k:v for k,v in new_state_dict.items() if "class_embed" in k}
line_embed = {k:v for k,v in new_state_dict.items() if "lines_embed" in k}
model_without_ddp.load_state_dict(encoder, strict=False)
model_without_ddp.load_state_dict(decoder, strict=False)
model_without_ddp.load_state_dict(class_embed, strict=False)
model_without_ddp.load_state_dict(line_embed, strict=False)
print('Finish load frozen_weights')
else:
print("NO RESUME. TRAIN FROM SCRATCH")
if args.eval:
test_stats = evaluate(model, criterion, postprocessors, data_loader_val, base_ds, device, args.output_dir, args)
#print('checkpoint'+ str(checkpoint['epoch']))
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, postprocessors, data_loader_train, optimizer, device, epoch, args.clip_max_norm, args)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoints/checkpoint.pth']
# extra checkpoint before LR drop and every several epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % args.save_freq == 0:
checkpoint_paths.append(output_dir / f'checkpoints/checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
test_stats = evaluate(model, criterion, postprocessors, data_loader_val, base_ds, device, args.output_dir, args)
log_stats = {**{f'train_{k}': format(v, ".6f") for k, v in train_stats.items()},
**{f'test_{k}': format(v, ".6f") 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)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
model.to(device)
checkpoint = torch.load(args.resume, map_location='cpu')
model.load_state_dict(checkpoint['model'], strict=False)
if torch.cuda.is_available():
model.cuda()
model.eval()
for img_file in os.listdir(args.imgs_dir):
t0 = time.time()
img_path = os.path.join(args.imgs_dir, img_file)
out_imgName = './visualize/' + 'out_' + img_file[:-4] + '.png'
im = Image.open(img_path)
# mean-std normalize the input image (batch-size: 1)
img = transform(im).unsqueeze(0)
img = img.cuda()
# propagate through the model
outputs = model(img)
out_logits, out_bbox = outputs['pred_logits'], outputs['pred_boxes']
prob = out_logits.sigmoid()
topk_values, topk_indexes = torch.topk(prob.view(
out_logits.shape[0], -1),
100,
dim=1)
scores = topk_values
topk_boxes = topk_indexes // out_logits.shape[2]
labels = topk_indexes % out_logits.shape[2]
boxes = box_ops.box_cxcywh_to_xyxy(out_bbox)
boxes = torch.gather(boxes, 1,
topk_boxes.unsqueeze(-1).repeat(1, 1, 4))
keep = scores[0] > 0.2
boxes = boxes[0, keep]
labels = labels[0, keep]
# and from relative [0, 1] to absolute [0, height] coordinates
im_h, im_w = im.size
#print('im_h,im_w',im_h,im_w)
target_sizes = torch.tensor([[im_w, im_h]])
target_sizes = target_sizes.cuda()
img_h, img_w = target_sizes.unbind(1)
scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1)
boxes = boxes * scale_fct[:, None, :]
print(time.time() - t0)
#plot_results
source_img = Image.open(img_path).convert("RGBA")
draw = ImageDraw.Draw(source_img)
print("Boxes", boxes, boxes.tolist())
for xmin, ymin, xmax, ymax in boxes[0].tolist():
draw.rectangle(((xmin, ymin), (xmax, ymax)), outline="red")
source_img.save(out_imgName, "png")
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
model.to(device)
checkpoint = torch.load(args.resume, map_location='cpu')
model.load_state_dict(checkpoint['model'], strict=False)
if torch.cuda.is_available():
model.cuda()
model.eval()
t0 = time.time()
im = Image.open(args.img_path)
print(im)
# mean-std normalize the input image (batch-size: 1)
img = transform(im).unsqueeze(0)
#print('img shape = ', img[0].shape)
img=img.cuda()
#print(img[0].cpu().numpy())
import cv2
# cv2.imwrite('/content/content/content/Deformable-DETR/explained /sample before forward/tensor0.png', img[0][0].cpu().numpy(), )
# cv2.imwrite( '/content/content/content/Deformable-DETR/explained /sample before forward/tensor1.png', img[0][1].cpu().numpy(),)
# cv2.imwrite( '/content/content/content/Deformable-DETR/explained /sample before forward/tensor2.png', img[0][2].cpu().numpy(),)
# cv2.imwrite( '/content/content/content/Deformable-DETR/explained /sample before forward/tensor3.png', img[0].cpu().numpy(),)
# propagate through the model
outputs = model(img)
print ('-----------------------------------------------------------------------------')
# mean-std normalize the input image (batch-size: 1)
img = transform(im).unsqueeze(0)
img=img.cuda()
# propagate through the model
outputs = model(img)
# keep only predictions with 0.7+ confidence
probas = outputs['pred_logits'].softmax(-1)[0, :, :-1]
keep = probas.max(-1).values > 0.6
# convert boxes from [0; 1] to image scales
bboxes_scaled = V.rescale_bboxes(outputs['pred_boxes'][0, keep], im.size)
V.plot_results(im, probas[keep], bboxes_scaled, '/content/Explain-Deformable-DETR/explained /output/1.png')
print ('keep.nonzero()', keep)
# use lists to store the outputs via up-values
conv_features, enc_attn_weights, dec_attn_weights = [], [], []
hooks = [
model.backbone[-2].register_forward_hook(
lambda self, input, output: conv_features.append(output)
),
model.transformer.encoder.layers[-1].self_attn.register_forward_hook(
lambda self, input, output: enc_attn_weights.append(output[0])
),
model.transformer.decoder.layers[-1].self_attn.register_forward_hook(
lambda self, input, output: dec_attn_weights.append(output[0])
),
]
# propagate through the model
outputs = model(img)
for hook in hooks:
hook.remove()
# don't need the list anymore
conv_features = conv_features[0]
enc_attn_weights = enc_attn_weights[0]
dec_attn_weights = dec_attn_weights[0]
h, w = conv_features['0'].tensors.shape[-2:]
print ('(dec_attn_weights[0, idx].view(h, w))', (dec_attn_weights.transpose(0, 1).shape))
dec_attn_weights = dec_attn_weights.transpose(0, 1).cpu()
print ('bboxes_scaled', bboxes_scaled.shape)
fig, axs = plt.subplots(ncols=len(bboxes_scaled), nrows=2, figsize=(22, 7))
colors = V.COLORS * 100
for idx, ax_i, (xmin, ymin, xmax, ymax) in zip(keep.nonzero(), axs.T, bboxes_scaled):
ax = ax_i[0]
ax.imshow(dec_attn_weights[0, idx].view(16, 16))
ax.axis('off')
ax.set_title(f'query id: {idx.item()}')
ax = ax_i[1]
ax.imshow(im)
ax.add_patch(plt.Rectangle((xmin, ymin), xmax - xmin, ymax - ymin,
fill=False, color='blue', linewidth=3))
ax.axis('off')
ax.set_title(V.CLASSES[probas[idx].argmax()])
fig.tight_layout()
fig.savefig('/content/Explain-Deformable-DETR/explained /encoder feature map/1.png')
# print ('model.transformer.encoder.layers[-1].self_attn', model.transformer.encoder.layers[-1].self_attn)
# print ('enc_attn_weights.shape', enc_attn_weights.shape)
# model.transformer.encoder.layers[-1].self_attn.register_forward_hook(printnorm)
# outputs = model(img)
print ('-----------------------------------------------------------------------------')
out_logits, out_bbox = outputs['pred_logits'], outputs['pred_boxes']
prob = out_logits.sigmoid()
topk_values, topk_indexes = torch.topk(prob.view(out_logits.shape[0], -1), 100, dim=1)
scores = topk_values
topk_boxes = topk_indexes // out_logits.shape[2]
labels = topk_indexes % out_logits.shape[2]
boxes = box_ops.box_cxcywh_to_xyxy(out_bbox)
boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1,1,4))
keep = scores[0] > 0.35
boxes = boxes[0, keep]
labels = labels[0, keep]
# and from relative [0, 1] to absolute [0, height] coordinates
im_h,im_w = im.size
#print('im_h,im_w',im_h,im_w)
target_sizes =torch.tensor([[im_w,im_h]])
target_sizes =target_sizes.cuda()
img_h, img_w = target_sizes.unbind(1)
scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1)
boxes = boxes * scale_fct[:, None, :]
print(time.time()-t0)
#plot_results
source_img = Image.open(args.img_path).convert("RGBA")
draw = ImageDraw.Draw(source_img)
print("Boxes",boxes,boxes.tolist())
for xmin, ymin, xmax, ymax in boxes[0].tolist():
draw.rectangle(((xmin, ymin), (xmax, ymax)), outline ="red")
source_img.save('test.png', "png")
results = [{'scores': s, 'labels': l, 'boxes': b} for s, l, b in zip(scores, labels, boxes)]
output_dir = "outputs"
device = "cuda"
seed = 42
start_epoch = 0
eval = False
num_workers = 2
# distributed
distributed = False
world_size = 1
dist_url = "env://"
# %%
args = Args()
utils.init_distributed_mode(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)
# %% PREPARE DATA
INSTRE_PATH = "/datadrive/crr/datasets/instre"
TRAIN_PATH = INSTRE_PATH + "/INSTRE-S-TRAIN"
TEST_PATH = INSTRE_PATH + "/INSTRE-S-TEST"
class_dirs_train = glob.glob(os.path.join(TRAIN_PATH, "*/"))
def main(args):
utils.init_distributed_mode(args)
print("args: {}".format(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)
# special process to control whether freeze backbone
args.model.train_backbone = args.lr_backbone > 0
model, criterion, postprocessors = build_model(args.model)
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
args.lr_drop,
gamma=args.lr_gamma)
dataset_train = build_dataset(
image_set='train',
args=args.dataset,
model_stride=model_without_ddp.backbone.stride)
dataset_val = build_dataset(image_set='val',
args=args.dataset,
model_stride=model_without_ddp.backbone.stride)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
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 '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
benchmark_test_parser = benchmark_test.get_args_parser()
benchmark_test_args = benchmark_test_parser.get_defaults()
benchmark_test_args.tracker.model = args.model # overwrite the parameters about network model
benchmark_test_args.result_path = Path(
os.path.join(args.output_dir, 'benchmark'))
benchmark_test_args.dataset_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), 'benchmark')
benchmark_eval_parser = benchmark_eval.get_args_parser()
benchmark_eval_args = benchmark_eval_parser.get_defaults()
benchmark_eval_args.tracker_path = benchmark_test_args.result_path
best_eao = 0
best_ar = [0, 10] # accuracy & robustness
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
# training
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every args.model_save_step epochs
if (epoch + 1) % args.lr_drop == 0 or (
epoch + 1) % args.model_save_step == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
# hack: only inference model
utils.save_on_master({'model': model_without_ddp.state_dict()},
output_dir / 'checkpoint_only_inference.pth')
# evalute
val_stats = evaluate(model, criterion, postprocessors, data_loader_val,
device, args.output_dir)
log_stats = {
'epoch': epoch,
**{f'train_{k}': v
for k, v in train_stats.items()},
**{f'val_{k}': v
for k, v in val_stats.items()}, '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")
# evualute with benchmark
if utils.is_main_process():
if (
epoch + 1
) % args.benchmark_test_step == 0 and epoch > args.benchmark_start_epoch:
tracker = build_tracker(benchmark_test_args.tracker,
model=model_without_ddp,
postprocessors=postprocessors)
benchmark_test_args.model_name = "epoch" + str(epoch)
benchmark_start_time = time.time()
benchmark_test.main(benchmark_test_args, tracker)
benchmark_time = time.time() - benchmark_start_time
benchmark_eval_args.model_name = "epoch" + str(epoch)
benchmark_eval_args.tracker_prefix = "epoch" + str(epoch)
eval_results = benchmark_eval.main(benchmark_eval_args)
eval_result = list(eval_results.values())[0]
if benchmark_test_args.dataset in ['VOT2018', 'VOT2019']:
if args.output_dir:
with (output_dir /
str("benchmark_" + benchmark_test_args.dataset +
".txt")).open("a") as f:
f.write("epoch: " + str(epoch) + ", best EAO: " +
str(best_eao) + ", " +
json.dumps(eval_result) + "\n")
if best_eao < eval_result['EAO']:
best_eao = eval_result['EAO']
if args.output_dir:
best_eao_int = int(best_eao * 1000)
# record: only inference model
utils.save_on_master(
{'model': model_without_ddp.state_dict()},
output_dir /
f'checkpoint{epoch:04}_best_eao_{best_eao_int:03}_only_inference.pth'
)
if best_ar[0] < eval_result['accuracy'] and best_ar[
1] > eval_result['robustness']:
best_ar[0] = eval_result['accuracy']
best_ar[1] = eval_result['robustness']
if args.output_dir:
best_accuracy_int = int(best_ar[0] * 1000)
best_robustness_int = int(best_ar[1] * 1000)
# record: only inference model
utils.save_on_master(
{'model': model_without_ddp.state_dict()},
output_dir /
f'checkpoint{epoch:04}_best_ar_{best_accuracy_int:03}_{best_robustness_int:03}_only_inference.pth'
)
print("benchmark time: {}".format(benchmark_time))
if args.distributed:
torch.distributed.barrier()
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("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
model.to(device)
model_without_ddp = model
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
dataset_val = build_dataset(image_set='val', args=args)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
base_ds = get_coco_api_from_dataset(dataset_val)
checkpoint = torch.load(args.resume)
model_without_ddp.load_state_dict(checkpoint['model'])
model_without_ddp.eval()
model_without_ddp.to(device)
header = 'Test:'
for samples, targets in data_loader_val:
samples = samples.to(device)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
outputs = model_without_ddp(samples)
probas = outputs['pred_logits'].softmax(-1)[0, :, :-1]
keep = probas.max(-1).values > 0.7
bboxes_scaled = rescale_bboxes(outputs['pred_boxes'][0, keep])
print(probas[keep], bboxes_scaled)
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
dataset_train = build_dataset(image_set="train", args=args)
dataset_val = build_dataset(image_set="val", args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(
dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
)
data_loader_val = DataLoader(
dataset_val,
args.batch_size if args.batch_size < 4 else 4,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
)
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.coco.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
elif args.dataset_file in ["cmdd", "cmdc", "wider"]:
base_ds = None
elif args.dataset_file == "MOT17":
base_ds = dataset_val
else:
base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location="cpu")
model_without_ddp.detr.load_state_dict(checkpoint["model"])
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")
# NOTE: this is Bruno's hack to load stuff in
model_dict = model_without_ddp.state_dict()
pretrained_dict = checkpoint["model"]
# hack for adding query stuff
if ("query_embed.query_embed.weight" in model_dict.keys()
and "query_embed.weight" in pretrained_dict.keys()):
pretrained_dict[
"query_embed.query_embed.weight"] = pretrained_dict[
"query_embed.weight"]
# 1. filter out unnecessary keys
pretrained_dict = {
k: v
for k, v in pretrained_dict.items() if k in model_dict
}
# if finetuning skip the linear stuff
if args.finetune:
pretrained_dict = {
k: v
for k, v in pretrained_dict.items()
if k not in ["class_embed.weight", "class_embed.bias"]
}
# 2. overwrite entries in the existing state dict
model_dict.update(pretrained_dict)
# 3. load new state dict
model_without_ddp.load_state_dict(model_dict)
if (not args.eval and not args.load_model_only
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.eval:
if args.test and args.dataset_file == "wider":
if args.resume:
s = args.resume.split("/")[:-1]
output_dir = "/" + os.path.join(*s)
else:
output_dir = args.output_dir
print("SAVING TEST WIDER TO ", output_dir)
test_wider(
model,
criterion,
postprocessors,
dataset_val,
data_loader_val,
device,
output_dir,
)
return
test_stats, coco_evaluator = evaluate(
model,
criterion,
postprocessors,
data_loader_val,
base_ds,
device,
args.output_dir,
dset_file=args.dataset_file,
)
if args.output_dir and coco_evaluator is not None:
utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval.pth")
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(
model,
criterion,
data_loader_train,
optimizer,
device,
epoch,
args.clip_max_norm,
)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / "checkpoint.pth"]
# extra checkpoint before LR drop and every 100 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 100 == 0:
checkpoint_paths.append(output_dir /
f"checkpoint{epoch:04}.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,
"args": args,
},
checkpoint_path,
)
test_stats, coco_evaluator = evaluate(
model,
criterion,
postprocessors,
data_loader_val,
base_ds,
device,
args.output_dir,
dset_file=args.dataset_file,
)
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")
# for evaluation logs
if coco_evaluator is not None:
(output_dir / "eval").mkdir(exist_ok=True)
if "bbox" in coco_evaluator.coco_eval:
filenames = ["latest.pth"]
if epoch % 50 == 0:
filenames.append(f"{epoch:03}.pth")
for name in filenames:
torch.save(
coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval" / name,
)
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, exp_cfg):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
print(args)
# device = torch.device('cuda')
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)
# model, criterion, postprocessors = build_model(args)
model = SMPLXNet(exp_cfg)
model.to(device)
model_without_ddp = model
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
# for n, p in model_without_ddp.named_parameters():
# print(n)
# dataset_train = build_dataset(image_set='train', args=args)
# dataset_val = build_dataset(image_set='val', args=args)
print('start build dataset')
datasets = make_all_datasets(exp_cfg, split='train')
# dataset_train = ConcatDataset(datasets['body'])
dataset_train = ConcatDataset(datasets['body'] + datasets['hand'] +
datasets['head'])
print('finish build dataset')
sample_weight = [
child_dataset.sample_weight for child_dataset in dataset_train.datasets
]
sample_weight = np.concatenate(sample_weight, axis=0)
sampler_train = torch.utils.data.sampler.WeightedRandomSampler(
sample_weight, len(dataset_train))
# sampler_val = torch.utils.data.SequentialSampler(dataset_val)
if args.distributed:
sampler_train = samplers.DistributedSampler(sampler_train)
# sampler_val = samplers.DistributedSampler(sampler_val, shuffle=False)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
collate_fn = functools.partial(collate_batch,
use_shared_memory=args.num_workers > 0,
return_full_imgs=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=collate_fn,
num_workers=args.num_workers,
pin_memory=True)
# data_loader_val = DataLoader(dataset_val, args.batch_size, sampler=sampler_val,
# drop_last=False, collate_fn=utils.collate_fn, num_workers=args.num_workers,
# pin_memory=True)
optim_cfg = exp_cfg.get('optim', {})
optimizer = build_optimizer(model, optim_cfg)
lr_scheduler = build_scheduler(optimizer, optim_cfg['scheduler'])
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
if args.pretrain:
checkpoint = torch.load(args.pretrain, map_location='cpu')
missing_keys, unexpected_keys = model_without_ddp.load_state_dict(
checkpoint['model'], strict=False)
unexpected_keys = [
k for k in unexpected_keys
if not (k.endswith('total_params') or k.endswith('total_ops'))
]
if len(missing_keys) > 0:
print('Missing Keys: {}'.format(missing_keys))
if len(unexpected_keys) > 0:
print('Unexpected Keys: {}'.format(unexpected_keys))
if args.resume:
checkpoint = torch.load(args.resume, map_location='cpu')
missing_keys, unexpected_keys = model_without_ddp.load_state_dict(
checkpoint['model'], strict=False)
unexpected_keys = [
k for k in unexpected_keys
if not (k.endswith('total_params') or k.endswith('total_ops'))
]
if len(missing_keys) > 0:
print('Missing Keys: {}'.format(missing_keys))
if len(unexpected_keys) > 0:
print('Unexpected Keys: {}'.format(unexpected_keys))
if 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
import copy
p_groups = copy.deepcopy(optimizer.param_groups)
optimizer.load_state_dict(checkpoint['optimizer'])
for pg, pg_old in zip(optimizer.param_groups, p_groups):
pg['lr'] = pg_old['lr']
pg['initial_lr'] = pg_old['initial_lr']
print(optimizer.param_groups)
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
lr_scheduler.step(lr_scheduler.last_epoch)
args.start_epoch = checkpoint['epoch'] + 1
print("Start training")
output_dir = Path(args.output_dir)
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, data_loader_train, optimizer,
device, epoch)
# print('DEBUG!!!!!!!!!'); train_stats = {}
lr_scheduler.step()
if args.output_dir:
if not os.path.exists(args.output_dir) and utils.is_main_process():
os.makedirs(args.output_dir)
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 1 epochs
if (epoch + 1) % args.save_freq == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
scaler = torch.cuda.amp.GradScaler(enabled=args.fp16)
if args.det_val:
assert args.eval, 'only support eval mode of detector for track'
model, criterion, postprocessors = build_model(args)
elif args.eval:
model, criterion, postprocessors = build_tracktest_model(args)
else:
model, criterion, postprocessors = build_tracktrain_model(args)
model.to(device)
model_without_ddp = model
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
dataset_train = build_dataset(image_set=args.track_train_split, args=args)
dataset_val = build_dataset(image_set=args.track_eval_split, args=args)
if args.distributed:
if args.cache_mode:
sampler_train = samplers.NodeDistributedSampler(dataset_train)
sampler_val = samplers.NodeDistributedSampler(dataset_val,
shuffle=False)
else:
sampler_train = samplers.DistributedSampler(dataset_train)
sampler_val = samplers.DistributedSampler(dataset_val,
shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
pin_memory=True)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
pin_memory=True)
# lr_backbone_names = ["backbone.0", "backbone.neck", "input_proj", "transformer.encoder"]
def match_name_keywords(n, name_keywords):
out = False
for b in name_keywords:
if b in n:
out = True
break
return out
for n, p in model_without_ddp.named_parameters():
print(n)
param_dicts = [{
"params": [
p for n, p in model_without_ddp.named_parameters()
if not match_name_keywords(n, args.lr_backbone_names)
and not match_name_keywords(n, args.lr_linear_proj_names)
and p.requires_grad
],
"lr":
args.lr,
}, {
"params": [
p for n, p in model_without_ddp.named_parameters() if
match_name_keywords(n, args.lr_backbone_names) and p.requires_grad
],
"lr":
args.lr_backbone,
}, {
"params": [
p for n, p in model_without_ddp.named_parameters()
if match_name_keywords(n, args.lr_linear_proj_names)
and p.requires_grad
],
"lr":
args.lr * args.lr_linear_proj_mult,
}]
if args.sgd:
optimizer = torch.optim.SGD(param_dicts,
lr=args.lr,
momentum=0.9,
weight_decay=args.weight_decay)
else:
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu], find_unused_parameters=True)
model_without_ddp = model.module
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.coco.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
else:
base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
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')
missing_keys, unexpected_keys = model_without_ddp.load_state_dict(
checkpoint['model'], strict=False)
unexpected_keys = [
k for k in unexpected_keys
if not (k.endswith('total_params') or k.endswith('total_ops'))
]
if len(missing_keys) > 0:
print('Missing Keys: {}'.format(missing_keys))
if len(unexpected_keys) > 0:
print('Unexpected Keys: {}'.format(unexpected_keys))
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
import copy
p_groups = copy.deepcopy(optimizer.param_groups)
optimizer.load_state_dict(checkpoint['optimizer'])
for pg, pg_old in zip(optimizer.param_groups, p_groups):
pg['lr'] = pg_old['lr']
pg['initial_lr'] = pg_old['initial_lr']
print(optimizer.param_groups)
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
# todo: this is a hack for doing experiment that resume from checkpoint and also modify lr scheduler (e.g., decrease lr in advance).
args.override_resumed_lr_drop = True
if args.override_resumed_lr_drop:
print(
'Warning: (hack) args.override_resumed_lr_drop is set to True, so args.lr_drop would override lr_drop in resumed lr_scheduler.'
)
lr_scheduler.step_size = args.lr_drop
lr_scheduler.base_lrs = list(
map(lambda group: group['initial_lr'],
optimizer.param_groups))
lr_scheduler.step(lr_scheduler.last_epoch)
args.start_epoch = checkpoint['epoch'] + 1
# check the resumed model
# if not args.eval:
# test_stats, coco_evaluator, _ = evaluate(
# model, criterion, postprocessors, data_loader_val, base_ds, device, args.output_dir
# )
if args.eval:
assert args.batch_size == 1, print("Now only support 1.")
tracker = Tracker(score_thresh=args.track_thresh)
test_stats, coco_evaluator, res_tracks = evaluate(model,
criterion,
postprocessors,
data_loader_val,
base_ds,
device,
args.output_dir,
tracker=tracker,
phase='eval',
det_val=args.det_val,
fp16=args.fp16)
if args.output_dir:
# utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval, output_dir / "eval.pth")
if res_tracks is not None:
print("Creating video index for {}.".format(args.dataset_file))
video_to_images = defaultdict(list)
video_names = defaultdict()
for _, info in dataset_val.coco.imgs.items():
video_to_images[info["video_id"]].append({
"image_id":
info["id"],
"frame_id":
info["frame_id"]
})
video_name = info["file_name"].split("/")[0]
if video_name not in video_names:
video_names[info["video_id"]] = video_name
assert len(video_to_images) == len(video_names)
# save mot results.
save_track(res_tracks, args.output_dir, video_to_images,
video_names, args.track_eval_split)
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model,
criterion,
data_loader_train,
optimizer,
device,
scaler,
epoch,
args.clip_max_norm,
fp16=args.fp16)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 5 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 5 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
if epoch % 10 == 0 or epoch > args.epochs - 5:
test_stats, coco_evaluator, _ = evaluate(model,
criterion,
postprocessors,
data_loader_val,
base_ds,
device,
args.output_dir,
fp16=args.fp16)
log_test_stats = {
**{f'test_{k}': v
for k, v in test_stats.items()}
}
log_stats.update(log_test_stats)
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")
# for evaluation logs
# if coco_evaluator is not None:
# (output_dir / 'eval').mkdir(exist_ok=True)
# if "bbox" in coco_evaluator.coco_eval:
# filenames = ['latest.pth']
# if epoch % 50 == 0:
# filenames.append(f'{epoch:03}.pth')
# for name in filenames:
# torch.save(coco_evaluator.coco_eval["bbox"].eval,
# output_dir / "eval" / name)
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("git:\n {}\n".format(utils.get_sha()))
# align with DETR format
args.dataset_file = 'ImageNet'
args.masks = None
# freeze cnn weights
args.lr_backbone = 0 if args.fre_cnn else args.lr
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)
model, criterion, postprocessors = build_model(args)
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
dataset_train = build_dataset(image_set='train', args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.updetr_collate_fn,
num_workers=args.num_workers)
print(len(data_loader_train) * args.epochs)
output_dir = Path(args.output_dir)
if args.resume:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
if lr_scheduler.step_size != args.lr_drop:
lr_scheduler.step_size = args.lr_drop
args.start_epoch = checkpoint['epoch'] + 1
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 20 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 20 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
Dataset = get_dataset(args.dataset, args.task)
f = open(args.data_cfg)
data_config = json.load(f)
trainset_paths = data_config['train']
dataset_root = data_config['root']
f.close()
normalize = T.Compose([
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
scales = [480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800]
transforms = T.Compose([
T.RandomHorizontalFlip(),
T.RandomSelect(
T.RandomResize(scales, max_size=1333),
T.Compose([
T.RandomResize([400, 500, 600]),
T.RandomSizeCrop(384, 600),
# T.RandomSizeCrop_MOT(384, 600),
T.RandomResize(scales, max_size=1333),
])),
normalize,
])
dataset_train = Dataset(args,
dataset_root,
trainset_paths, (1088, 608),
augment=True,
transforms=transforms)
args.nID = dataset_train.nID
model, criterion, postprocessors = build_model(args)
model.to(device)
model_without_ddp = model
# dataset_train = build_dataset(image_set='train', args=args)
# dataset_val = build_dataset(image_set='val', args=args)
if args.distributed:
if args.cache_mode:
sampler_train = samplers.NodeDistributedSampler(dataset_train)
# sampler_val = samplers.NodeDistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = samplers.DistributedSampler(dataset_train)
# sampler_val = samplers.DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
# sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
pin_memory=True)
# data_loader_val = DataLoader(dataset_val, args.batch_size, sampler=sampler_val,
# drop_last=False, collate_fn=utils.collate_fn, num_workers=args.num_workers,
# pin_memory=True)
# data_loader_train = torch.utils.data.DataLoader(
# dataset_train,
# batch_size=args.batch_size,
# shuffle=True,
# num_workers=args.num_workers,
# pin_memory=True,
# drop_last=True
# )
# lr_backbone_names = ["backbone.0", "backbone.neck", "input_proj", "transformer.encoder"]
def match_name_keywords(n, name_keywords):
out = False
for b in name_keywords:
if b in n:
out = True
break
return out
for n, p in model_without_ddp.named_parameters():
print(n)
# 用于将classifer不更新参数
# for name,p in model_without_ddp.named_parameters():
# if name.startswith('classifier'):
# p.requires_grad = False
param_dicts = [{
"params": [
p for n, p in model_without_ddp.named_parameters()
if not match_name_keywords(n, args.lr_backbone_names)
and not match_name_keywords(n, args.lr_linear_proj_names)
and p.requires_grad
],
"lr":
args.lr,
}, {
"params": [
p for n, p in model_without_ddp.named_parameters() if
match_name_keywords(n, args.lr_backbone_names) and p.requires_grad
],
"lr":
args.lr_backbone,
}, {
"params": [
p for n, p in model_without_ddp.named_parameters()
if match_name_keywords(n, args.lr_linear_proj_names)
and p.requires_grad
],
"lr":
args.lr * args.lr_linear_proj_mult,
}]
if args.sgd:
optimizer = torch.optim.SGD(param_dicts,
lr=args.lr,
momentum=0.9,
weight_decay=args.weight_decay)
else:
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
# optimizer.add_param_group({'params': criterion.parameters()})
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
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_dict = model_without_ddp.state_dict() #当前模型参数
pretrained_dict = {
k: v
for k, v in checkpoint['model'].items() if k not in [
"class_embed.0.weight", "class_embed.0.bias",
"class_embed.1.weight", "class_embed.1.bias",
"class_embed.2.weight", "class_embed.2.bias",
"class_embed.3.weight", "class_embed.3.bias",
"class_embed.4.weight", "class_embed.4.bias",
"class_embed.5.weight", "class_embed.5.bias"
]
}
model_dict.update(pretrained_dict)
# missing_keys, unexpected_keys = model_without_ddp.load_state_dict(checkpoint['model'], strict=False)
missing_keys, unexpected_keys = model_without_ddp.load_state_dict(
model_dict, strict=False)
unexpected_keys = [
k for k in unexpected_keys
if not (k.endswith('total_params') or k.endswith('total_ops'))
]
if len(missing_keys) > 0:
print('Missing Keys: {}'.format(missing_keys))
if len(unexpected_keys) > 0:
print('Unexpected Keys: {}'.format(unexpected_keys))
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
args.start_epoch = checkpoint['epoch'] + 1
# optimizer.load_state_dict(checkpoint['optimizer'])
# if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
# import copy
# p_groups = copy.deepcopy(optimizer.param_groups)
# # optimizer.load_state_dict(checkpoint['optimizer'])
# for pg, pg_old in zip(optimizer.param_groups, p_groups):
# pg['lr'] = pg_old['lr']
# pg['initial_lr'] = pg_old['initial_lr']
# # print(optimizer.param_groups)
# lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
# # todo: this is a hack for doing experiment that resume from checkpoint and also modify lr scheduler (e.g., decrease lr in advance).
# args.override_resumed_lr_drop = True
# if args.override_resumed_lr_drop:
# print('Warning: (hack) args.override_resumed_lr_drop is set to True, so args.lr_drop would override lr_drop in resumed lr_scheduler.')
# lr_scheduler.step_size = args.lr_drop
# lr_scheduler.base_lrs = list(map(lambda group: group['initial_lr'], optimizer.param_groups))
# lr_scheduler.step(lr_scheduler.last_epoch)
# model.add_module('id')
# [p for p in model.named_parameters() if not p[1].requires_grad]
# 用于将classifer不更新参数
# optimizer = torch.optim.SGD(filter(lambda x: "classifier" not in x[0], model.parameters()), lr=args.lr,
# momentum=0.9, weight_decay=1e-4)
# model.classifier.training = False
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(args, model, criterion,
data_loader_train, optimizer, device,
epoch, args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 5 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 5 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
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("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
print(args)
# IPython.embed()
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)
# IPython.embed()
# IPython.embed()
os.system("sudo chmod -R 777 /home/shuxuang/.cache/")
model, criterion, postprocessors = build_model(
args) # use the same model as detr paper on coco
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)
dataset_val = build_nvdataset(
dataset_root=[
os.path.join(os.environ["HOME"], 'datasets/detection-f'), #test
os.path.join(os.environ["HOME"], 'datasets/frames_nvidia')
],
mode='test',
camera=args.camera)
# dataset_val = build_nvdataset(dataset_root=[args.dataset_root_test, args.dataset_root_sql],
# mode='test', camera=args.camera)
print("Validation samples: %d" % (len(dataset_val)))
IPython.embed()
# compute how many boxes in the test dataset for each image
# accumulate_bboxes_numbers(dataset_val)
# dataset_train_ = build_nvdataset(dataset_root=[
# os.path.join(os.environ["HOME"],'datasets/annotation_sql_nvidia'),
# os.path.join(os.environ["HOME"], 'datasets/frames_nvidia')],
# mode='train', camera=args.camera)
# # indices_50k =np.load(os.path.join(os.environ["HOME"],'datasets/id_1_criterion_Max_SSD_num_labels_50000.npy'))
# # dataset_train = Subset(dataset_train_, indices_50k)
# print("Train samples: %d"%(len(dataset_train_)))
# print(len(dataset_val))
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
output_dir = Path(args.output_dir)
# args.resume = os.path.join(os.environ["HOME"], 'datasets/exps_detr_base/checkpoint0299.pth')
# args.resume = '/home/shuxuang/datasets/exps_detr_base/checkpoint0299.pth'
log_path = args.resume
log = os.path.join(args.resume, 'log.txt')
# read_log(log)
# IPython.embed()
args.resume = os.path.join(args.resume, 'checkpoint.pth')
print(args.resume)
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:
print('Loading model: %s' % args.resume)
checkpoint = torch.load(args.resume, map_location='cpu')
print('Load model from %d epoch' % checkpoint['epoch'])
model_without_ddp.load_state_dict(checkpoint['model'])
if args.eval:
vis_bboxes(model, dataset_val, postprocessors, device)
# inference_time(model, dataset_val, postprocessors, device)
return model, dataset_val, postprocessors, device
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
model.to(device)
model_without_ddp = model
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
dataset_train = build_dataset(image_set='train', args=args)
dataset_val = build_dataset(image_set='val', args=args)
dataset_test = build_dataset(image_set='test', args=args)
if args.distributed:
if args.cache_mode:
sampler_train = samplers.NodeDistributedSampler(dataset_train)
sampler_val = samplers.NodeDistributedSampler(dataset_val,
shuffle=False)
else:
sampler_train = samplers.DistributedSampler(dataset_train)
sampler_val = samplers.DistributedSampler(dataset_val,
shuffle=False)
sampler_test = samplers.DistributedSampler(dataset_test,
shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
sampler_test = torch.utils.data.SequentialSampler(dataset_test)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
pin_memory=True)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
pin_memory=True)
data_loader_test = DataLoader(dataset_test,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers,
pin_memory=True)
# lr_backbone_names = ["backbone.0", "backbone.neck", "input_proj", "transformer.encoder"]
def match_name_keywords(n, name_keywords):
out = False
for b in name_keywords:
if b in n:
out = True
break
return out
for n, p in model_without_ddp.named_parameters():
print(n)
param_dicts = [{
"params": [
p for n, p in model_without_ddp.named_parameters()
if not match_name_keywords(n, args.lr_backbone_names)
and not match_name_keywords(n, args.lr_linear_proj_names)
and p.requires_grad
],
"lr":
args.lr,
}, {
"params": [
p for n, p in model_without_ddp.named_parameters() if
match_name_keywords(n, args.lr_backbone_names) and p.requires_grad
],
"lr":
args.lr_backbone,
}, {
"params": [
p for n, p in model_without_ddp.named_parameters()
if match_name_keywords(n, args.lr_linear_proj_names)
and p.requires_grad
],
"lr":
args.lr * args.lr_linear_proj_mult,
}]
if args.sgd:
optimizer = torch.optim.SGD(param_dicts,
lr=args.lr,
momentum=0.9,
weight_decay=args.weight_decay)
else:
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.coco.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
else:
base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
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')
del checkpoint["model"]["transformer.decoder.class_embed.0.weight"]
del checkpoint["model"]["transformer.decoder.class_embed.0.bias"]
del checkpoint["model"]["transformer.decoder.class_embed.1.weight"]
del checkpoint["model"]["transformer.decoder.class_embed.1.bias"]
del checkpoint["model"]["transformer.decoder.class_embed.2.weight"]
del checkpoint["model"]["transformer.decoder.class_embed.2.bias"]
del checkpoint["model"]["transformer.decoder.class_embed.3.weight"]
del checkpoint["model"]["transformer.decoder.class_embed.3.bias"]
del checkpoint["model"]["transformer.decoder.class_embed.4.weight"]
del checkpoint["model"]["transformer.decoder.class_embed.4.bias"]
del checkpoint["model"]["transformer.decoder.class_embed.5.weight"]
del checkpoint["model"]["transformer.decoder.class_embed.5.bias"]
del checkpoint["model"]["transformer.decoder.class_embed.6.weight"]
del checkpoint["model"]["transformer.decoder.class_embed.6.bias"]
del checkpoint["model"]["class_embed.0.weight"]
del checkpoint["model"]["class_embed.0.bias"]
del checkpoint["model"]["class_embed.1.weight"]
del checkpoint["model"]["class_embed.1.bias"]
del checkpoint["model"]["class_embed.2.weight"]
del checkpoint["model"]["class_embed.2.bias"]
del checkpoint["model"]["class_embed.3.weight"]
del checkpoint["model"]["class_embed.3.bias"]
del checkpoint["model"]["class_embed.4.weight"]
del checkpoint["model"]["class_embed.4.bias"]
del checkpoint["model"]["class_embed.5.weight"]
del checkpoint["model"]["class_embed.5.bias"]
del checkpoint["model"]["class_embed.6.weight"]
del checkpoint["model"]["class_embed.6.bias"]
missing_keys, unexpected_keys = model_without_ddp.load_state_dict(
checkpoint['model'], strict=False)
unexpected_keys = [
k for k in unexpected_keys
if not (k.endswith('total_params') or k.endswith('total_ops'))
]
# if len(missing_keys) > 0:
# print('Missing Keys: {}'.format(missing_keys))
# if len(unexpected_keys) > 0:
# print('Unexpected Keys: {}'.format(unexpected_keys))
# if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
# import copy
# p_groups = copy.deepcopy(optimizer.param_groups)
# optimizer.load_state_dict(checkpoint['optimizer'])
# for pg, pg_old in zip(optimizer.param_groups, p_groups):
# pg['lr'] = pg_old['lr']
# pg['initial_lr'] = pg_old['initial_lr']
# #print(optimizer.param_groups)
# lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
# # todo: this is a hack for doing experiment that resume from checkpoint and also modify lr scheduler (e.g., decrease lr in advance).
# args.override_resumed_lr_drop = True
# if args.override_resumed_lr_drop:
# print('Warning: (hack) args.override_resumed_lr_drop is set to True, so args.lr_drop would override lr_drop in resumed lr_scheduler.')
# lr_scheduler.step_size = args.lr_drop
# lr_scheduler.base_lrs = list(map(lambda group: group['initial_lr'], optimizer.param_groups))
# lr_scheduler.step(lr_scheduler.last_epoch)
# args.start_epoch = checkpoint['epoch'] + 1
# # check the resumed model
if not args.eval:
test_stats, coco_evaluator = evaluate(model, criterion,
postprocessors,
data_loader_val, base_ds,
device, args.output_dir)
if args.eval:
test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device,
args.output_dir)
if args.output_dir:
utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval.pth")
return
if args.test:
test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
data_loader_test, base_ds,
device, args.output_dir)
if args.output_dir:
utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval.pth")
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 5 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 5 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device,
args.output_dir)
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")
# for evaluation logs
if coco_evaluator is not None:
(output_dir / 'eval').mkdir(exist_ok=True)
if "bbox" in coco_evaluator.coco_eval:
filenames = ['latest.pth']
if epoch % 50 == 0:
filenames.append(f'{epoch:03}.pth')
for name in filenames:
torch.save(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval" / name)
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("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
dataset_train = build_dataset(split='train', args=args)
dataset_val = build_dataset(split='val', args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.coco.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
elif args.dataset_file == "coco":
base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
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.eval:
# if 'coco' in args.dataset_file:
# test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
# data_loader_val, base_ds, device, args.output_dir)
# if args.output_dir:
# utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval, output_dir / "eval.pth")
# elif 'anet' == args.dataset_file:
# evaluate3d(model, postprocessors, data_loader_val, device, epoch=0)
# return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 100 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 100 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
if epoch % args.eval_freq == 0:
if 'coco' in args.dataset_file:
test_stats, coco_evaluator = evaluate(model, criterion,
postprocessors,
data_loader_val, base_ds,
device, args.output_dir)
elif 'anet' == args.dataset_file:
evaluate3d(model, postprocessors, data_loader_val, device,
epoch)
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
# IPython.embed()
# IPython.embed()
os.system("sudo chmod -R 777 /home/shuxuang/.cache/")
model, criterion, postprocessors = build_model(
args) # use the same model as detr paper on coco
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)
# optimizer = torch.optim.AdamW(param_dicts, lr=args.lr,
# weight_decay=args.weight_decay)
# lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
# dataset_train = build_dataset(image_set='train', args=args)
# dataset_val = build_dataset(image_set='val', args=args)
# modify the dataset from coco to nvdata
# home_dir = os.environ["HOME"]
# # on local
# dataset_train_ = build_nvdataset(dataset_root=[
# os.path.join(os.environ["HOME"],'datasets/annotation_sql_nvidia'),
# os.path.join(os.environ["HOME"], 'datasets/frames_nvidia')],
# mode='train')
# dataset_val = build_nvdataset(dataset_root=[
# os.path.join(os.environ["HOME"],'datasets/test'),
# os.path.join(os.environ["HOME"], 'datasets/frames_nvidia')],
# mode='test', camera=args.camera)
dataset_val = build_nvdataset(dataset_root=[
os.path.join(os.environ["HOME"], 'datasets/test'),
os.path.join(os.environ["HOME"], 'datasets/test')
],
mode='test',
camera=args.camera)
# indices_50k =np.load(os.path.join(os.environ["HOME"],'datasets/id_1_criterion_Max_SSD_num_labels_50000.npy'))
# # on maglev
# dataset_train_ = build_nvdataset(dataset_root=[args.dataset_root_sql, args.dataset_root_img],
# mode='train')
# dataset_val = build_nvdataset(dataset_root=[args.dataset_root_test, args.dataset_root_sql],
# mode='test', camera=args.camera)
# indices_50k =np.load(os.path.join(args.root_indices))
# dataset_train = Subset(dataset_train_, indices_50k)
print("Validation samples: %d" % (len(dataset_val)))
# IPython.embed()
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
output_dir = Path(args.output_dir)
# args.resume = os.path.join(os.environ["HOME"], 'datasets/exps_detr_base/checkpoint0299.pth')
# args.resume = '/home/shuxuang/datasets/exps_detr_base/checkpoint0299.pth'
print(args.resume)
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:
print('Loading model: %s' % args.resume)
checkpoint = torch.load(args.resume, map_location='cpu')
print('Load model from %d epoch' % (checkpoint['epoch'] + 1))
model_without_ddp.load_state_dict(checkpoint['model'])
if args.eval:
if args.dataset_file == 'nvdata':
evaluate(model, dataset_val, postprocessors, device)
else:
evaluate_5classes(model, dataset_val, postprocessors, device)
return model, dataset_val, postprocessors, device
def main(args):
utils.init_distributed_mode(args)
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
dataset_train = build_dataset(image_set='train', args=args)
dataset_val = build_dataset(image_set='val', args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
data_loader_val = DataLoader(dataset_val,
batch_size=1,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.coco.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
else:
base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
io.load_frozen(args, model_without_ddp)
output_dir = Path(args.output_dir)
if args.resume:
io.resume(args, model_without_ddp, optimizer, lr_scheduler)
elif args.finetune:
io.finetune(args, model_without_ddp)
if args.eval:
if args.output_dir and utils.is_main_process():
io.init_wandb(args.dataset_file + "-detr-eval",
model,
args,
n_parameters=n_parameters)
test_stats, evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device,
args.output_dir)
if args.output_dir:
io.save_on_master(evaluator.coco_eval["bbox"].eval,
output_dir / "eval.pth")
return
print("Start training")
start_time = time.time()
if args.output_dir and utils.is_main_process():
io.init_wandb(args.dataset_file + "-detr",
model,
args,
n_parameters=n_parameters)
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
# extra checkpoint before LR drop and every 100 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 100 == 0:
io.save_checkpoint(args, model_without_ddp, optimizer,
lr_scheduler, epoch)
test_stats, evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device,
args.output_dir, epoch)
if utils.is_main_process() and args.output_dir:
io.log_wandb(train_stats, test_stats)
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
# save final model
if utils.is_main_process() and args.output_dir:
io.save_on_master(model_without_ddp, output_dir / "model_final.pth")
print('Training time {}'.format(total_time_str))
def main(args):
# args = parser.parse_args()
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
print(args)
if args.seed is not None:
# random.seed(args.seed)
# torch.manual_seed(args.seed)
cudnn.deterministic = True
warnings.warn('You have chosen to seed training. '
'This will turn on the CUDNN deterministic setting, '
'which can slow down your training considerably! '
'You may see unexpected behavior when restarting '
'from checkpoints.')
# fix the seed for reproducibility
seed = args.seed + utils.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
##################################
# Logging setting
##################################
if args.output_dir and utils.is_main_process():
logging.basicConfig(
filename=os.path.join(args.output_dir, args.log_name),
filemode='w',
format=
'%(asctime)s: %(levelname)s: [%(filename)s:%(lineno)d]: %(message)s',
level=logging.INFO)
warnings.filterwarnings("ignore")
##################################
# Save to logging
##################################
if utils.is_main_process():
logging.info(str(args))
##################################
# Initialize dataset
##################################
if not args.evaluate:
# build_vocab_flag=True, # Takes a long time to build a vocab
train_dataset = GQATorchDataset(split='train_unbiased',
build_vocab_flag=False,
load_vocab_flag=False)
if args.distributed:
sampler_train = torch.utils.data.DistributedSampler(train_dataset)
else:
sampler_train = torch.utils.data.RandomSampler(train_dataset)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_sampler=batch_sampler_train,
collate_fn=GQATorchDataset_collate_fn,
num_workers=args.workers)
# Old version
# train_loader = torch.utils.data.DataLoader(
# train_dataset, batch_size=args.batch_size, shuffle=True,
# collate_fn=GQATorchDataset_collate_fn,
# num_workers=args.workers, pin_memory=True)
val_dataset_list = []
for eval_split in args.evaluate_sets:
val_dataset_list.append(
GQATorchDataset(split=eval_split,
build_vocab_flag=False,
load_vocab_flag=args.evaluate))
val_dataset = torch.utils.data.ConcatDataset(val_dataset_list)
if args.distributed:
sampler_val = torch.utils.data.DistributedSampler(val_dataset,
shuffle=False)
else:
sampler_val = torch.utils.data.SequentialSampler(val_dataset)
val_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=GQATorchDataset_collate_fn,
num_workers=args.workers)
# Old version
# val_loader = torch.utils.data.DataLoader(
# val_dataset,
# batch_size=args.batch_size, shuffle=False,
# collate_fn=GQATorchDataset_collate_fn,
# num_workers=args.workers, pin_memory=True)
##################################
# Initialize model
# - note: must init dataset first. Since we will use the vocab from the dataset
##################################
model = PipelineModel()
##################################
# Deploy model on GPU
##################################
model = model.to(device=cuda)
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)
##################################
# define optimizer (and scheduler)
##################################
# optimizer = torch.optim.SGD(model.parameters(), args.lr,
# momentum=args.momentum,
# weight_decay=args.weight_decay)
optimizer = torch.optim.Adam(
params=model.parameters(),
lr=args.lr,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=0, # weight_decay=args.weight_decay
amsgrad=False,
)
# optimizer = torch.optim.AdamW(
# params=model.parameters(),
# lr=args.lr,
# weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
model_without_ddp.load_state_dict(checkpoint['model'])
if not args.evaluate:
if 'optimizer' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer'])
if 'lr_scheduler' in checkpoint:
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
if 'epoch' in checkpoint:
args.start_epoch = checkpoint['epoch'] + 1
# checkpoint = torch.load(args.resume)
# args.start_epoch = checkpoint['epoch']
# model.load_state_dict(checkpoint['state_dict'])
# optimizer.load_state_dict(checkpoint['optimizer'])
# print("=> loaded checkpoint '{}' (epoch {})"
# .format(args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
# cudnn.benchmark = True
##################################
# Define loss functions (criterion)
##################################
# criterion = torch.nn.CrossEntropyLoss().cuda()
text_pad_idx = GQATorchDataset.TEXT.vocab.stoi[
GQATorchDataset.TEXT.pad_token]
criterion = {
"program":
torch.nn.CrossEntropyLoss(ignore_index=text_pad_idx).to(device=cuda),
"full_answer":
torch.nn.CrossEntropyLoss(ignore_index=text_pad_idx).to(device=cuda),
"short_answer":
torch.nn.CrossEntropyLoss().to(device=cuda),
# "short_answer": torch.nn.BCEWithLogitsLoss().to(device=cuda), # sigmoid
"execution_bitmap":
torch.nn.BCELoss().to(device=cuda),
}
##################################
# If Evaluate Only
##################################
if args.evaluate:
validate(val_loader, model, criterion, args, DUMP_RESULT=True)
return
##################################
# Main Training Loop
##################################
# best_acc1 = 0
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
##################################
# In distributed mode, calling the :meth`set_epoch(epoch) <set_epoch>` method
# at the beginning of each epoch before creating the DataLoader iterator is necessary
# to make shuffling work properly across multiple epochs.
# Otherwise, the same ordering will be always used.
##################################
sampler_train.set_epoch(epoch)
lr_scheduler.step()
# adjust_learning_rate(optimizer, epoch, args)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args)
# evaluate on validation set
if (epoch + 1) % 5 == 0:
validate(val_loader,
model,
criterion,
args,
FAST_VALIDATE_FLAG=False)
# # remember best acc@1 and save checkpoint
# save_checkpoint({
# 'epoch': epoch + 1,
# # 'arch': args.arch,
# 'state_dict': model.state_dict(),
# # 'best_acc1': best_acc1,
# 'optimizer' : optimizer.state_dict(),
# }, is_best)
if args.output_dir:
output_dir = pathlib.Path(args.output_dir)
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 100 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 100 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.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,
'args': args,
}, checkpoint_path)
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
dataset_train = build_dataset(image_set='train', args=args)
dataset_val = build_dataset(image_set='val', args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.coco.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
else:
base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
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'], strict=False)
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.eval:
test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device,
args.output_dir)
if args.output_dir:
utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval.pth")
return
#cab
writer = SummaryWriter("runs/" + args.tb_name)
best_value = 0
print("Start training, best_value is " + str(best_value))
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device,
args.output_dir)
#cab
for k, v in train_stats.items():
if isinstance(v, float):
writer.add_scalar(f'train_{k}', v, epoch)
new_value = 0
for k, v in test_stats.items():
if (isinstance(v, float)):
writer.add_scalar(f'test_{k}', v, epoch)
if (k == "coco_eval_bbox"):
new_value = v[0]
writer.add_scalar(
'Bbox Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ]',
v[0], epoch)
writer.add_scalar(
'Bbox Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ]',
v[1], epoch)
writer.add_scalar(
'Bbox Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ]',
v[2], epoch)
writer.add_scalar(
'Bbox Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ]',
v[3], epoch)
writer.add_scalar(
'Bbox Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ]',
v[4], epoch)
writer.add_scalar(
'Bbox Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ]',
v[5], epoch)
writer.add_scalar(
'Bbox Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ]',
v[6], epoch)
writer.add_scalar(
'Bbox Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ]',
v[7], epoch)
writer.add_scalar(
'Bbox Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ]',
v[8], epoch)
writer.add_scalar(
'Bbox Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ]',
v[9], epoch)
writer.add_scalar(
'Bbox Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ]',
v[10], epoch)
writer.add_scalar(
'Bbox Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ]',
v[11], epoch)
if (k == "coco_eval_masks"):
new_value = v[0]
writer.add_scalar(
'Mask Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ]',
v[0], epoch)
writer.add_scalar(
'Mask Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ]',
v[1], epoch)
writer.add_scalar(
'Mask Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ]',
v[2], epoch)
writer.add_scalar(
'Mask Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ]',
v[3], epoch)
writer.add_scalar(
'Mask Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ]',
v[4], epoch)
writer.add_scalar(
'Mask Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ]',
v[5], epoch)
writer.add_scalar(
'Mask Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ]',
v[6], epoch)
writer.add_scalar(
'Mask Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ]',
v[7], epoch)
writer.add_scalar(
'Mask Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ]',
v[8], epoch)
writer.add_scalar(
'Mask Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ]',
v[9], epoch)
writer.add_scalar(
'Mask Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ]',
v[10], epoch)
writer.add_scalar(
'Mask Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ]',
v[11], epoch)
print("Epoch finished, best_value is " + str(best_value))
save_pth = False
if best_value < new_value:
best_value = new_value
save_pth = True
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 100 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 100 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch:04}.pth')
if save_pth:
checkpoint_paths.append(output_dir / f'best.pth')
bestLog = open(output_dir / 'best_log.txt', 'w+')
bestLog.write(f'Saved model at epoch {epoch:04}\n')
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,
'args': args,
}, checkpoint_path)
#/cab
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")
# for evaluation logs
if coco_evaluator is not None:
(output_dir / 'eval').mkdir(exist_ok=True)
if "bbox" in coco_evaluator.coco_eval:
filenames = ['latest.pth']
if epoch % 50 == 0:
filenames.append(f'{epoch:03}.pth')
for name in filenames:
torch.save(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval" / name)
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("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
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)
model, criterion, postprocessors = build_model(args)
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)
param_dicts = [
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model_without_ddp.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts,
lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
args.lr_drop,
gamma=0.9)
dataset_train = build_dataset(image_set='train', args=args)
dataset_val = build_dataset(image_set='val', args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(sampler_train,
args.batch_size,
drop_last=True)
data_loader_train = DataLoader(dataset_train,
batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
data_loader_val = DataLoader(dataset_val,
args.batch_size,
sampler=sampler_val,
drop_last=False,
collate_fn=utils.collate_fn,
num_workers=args.num_workers)
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.coco.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
else:
base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
output_dir = Path(args.output_dir)
output_dir = output_dir / f"{args.backbone}_{args.transformer_type}"
if args.output_dir:
output_dir.mkdir(parents=True, exist_ok=True)
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.eval:
test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device,
args.output_dir)
if args.output_dir:
utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval.pth")
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(model, criterion, data_loader_train,
optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / f'checkpoint_{epoch}.pth']
# extra checkpoint before LR drop and every 100 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 100 == 0:
checkpoint_paths.append(output_dir /
f'checkpoint{epoch}_extra.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,
'args': args,
}, checkpoint_path)
test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device,
args.output_dir)
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")
# for evaluation logs
if coco_evaluator is not None:
(output_dir / 'eval').mkdir(exist_ok=True)
if "bbox" in coco_evaluator.coco_eval:
filenames = ['latest.pth']
if epoch % 50 == 0:
filenames.append(f'{epoch:03}.pth')
for name in filenames:
torch.save(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval" / name)
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)
device = torch.device(args.device)
args.batch_size = 1
print(args)
if args.dataset_config is not None:
# https://stackoverflow.com/a/16878364
d = vars(args)
with open(args.dataset_config, "r") as f:
cfg = json.load(f)
d.update(cfg)
# fix the seed for reproducibility
seed = args.seed + dist.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
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_args = checkpoint["args"]
for a in vars(args):
if a not in vars(model_args):
vars(model_args)[a] = vars(args)[a]
model_args.device = args.device
model, _, _, _, _ = build_model(model_args)
model.to(device)
with open(Path(args.lvis_minival_path) / "lvis_v1_minival.json", "r") as f:
lvis_val = json.load(f)
id2cat = {c["id"]: c for c in lvis_val["categories"]}
all_cats = sorted(list(id2cat.keys()))
# 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)
dset = build_dataset("lvis", image_set="minival", args=args)
sampler = DistributedSampler(dset, shuffle=False) if args.distributed else torch.utils.data.SequentialSampler(dset)
dataloader = DataLoader(
dset,
args.batch_size,
sampler=sampler,
drop_last=False,
collate_fn=partial(utils.collate_fn, False),
num_workers=args.num_workers,
)
if args.test:
evaluator = LvisDumper(fixed_ap=True, out_path=os.path.join(args.output_dir, "lvis_eval"))
else:
evaluator = LvisEvaluatorFixedAP(dset.lvis, fixed_ap=True)
postprocessor = build_postprocessors(args, "lvis")
model.load_state_dict(checkpoint["model_ema"], strict=False)
model.eval()
label_set = torch.as_tensor(list(all_cats))
print("label_set", len(label_set))
text_memories = []
print("encoding text...")
splits = torch.split(label_set, 32)
for split in tqdm(splits):
captions = [f"{clean_name(id2cat[l]['name'])}" for l in split.tolist()]
tokenized = model.transformer.tokenizer.batch_encode_plus(captions, padding="longest", return_tensors="pt").to(
device
)
encoded_text = model.transformer.text_encoder(**tokenized)
# Transpose memory because pytorch's attention expects sequence first
text_memory = encoded_text.last_hidden_state.transpose(0, 1)
# Invert attention mask that we get from huggingface because its the opposite in pytorch transformer
text_attention_mask = tokenized.attention_mask.ne(1).bool()
# Resize the encoder hidden states to be of the same d_model as the decoder
text_memory_resized = model.transformer.resizer(text_memory)
text_memories.append((text_attention_mask, text_memory_resized, tokenized))
for batch_dict in tqdm(dataloader):
samples = batch_dict["samples"].to(device)
targets = batch_dict["targets"]
targets = [
{
k: v.to(device) if k not in ["tokens_positive", "tokens", "dataset_name"] else v
for k, v in t.items()
if k != "caption"
}
for t in targets
]
assert len(targets) == 1
t = targets[0]
with torch.no_grad():
features, orig_pos = model.backbone(samples)
orig_src, orig_mask = features[-1].decompose()
res = []
for i in range(len(text_memories)):
# captions = [f"{clean_name(id2cat[l]['name'])}" for l in split.tolist()]
bs = len(splits[i])
src = orig_src.repeat(bs, 1, 1, 1)
mask = orig_mask.repeat(bs, 1, 1, 1)
pos = deepcopy(orig_pos)
pos[0] = pos[0].repeat(bs, 1, 1, 1)
memory_cache = model.transformer(
model.input_proj(src),
mask,
model.query_embed.weight,
pos[-1],
text_memories[i], # captions,
encode_and_save=True,
text_memory=None,
img_memory=None,
text_attention_mask=None,
)
out = model(samples, captions, encode_and_save=False, memory_cache=memory_cache)
orig_target_sizes = torch.stack([t["orig_size"] for _ in range(bs)], dim=0)
results = postprocessor["bbox"](out, orig_target_sizes)
assert len(results) == len(splits[i])
for j in range(len(results)):
results[j]["labels"] *= splits[i][j].item()
for output in results:
res.append((t["image_id"].item(), output))
evaluator.update(res)
evaluator.synchronize_between_processes()
evaluator.summarize()
