about real dataset (labeled or unlabeled)

[lx7555]

Thanks for your work, I noticed that you validated the importance of using unlabeled real datasets. So how do I control in code to implement data states ((labeled or unlabeled)) for training.

Thank you!

[WangWenhao0716]

They should be controlled by train.py:

train_loader = get_train_loader(args, dataset_real, args.height, args.width,
                                            args.batch_size, args.workers, args.num_instances, iters,
                                            trainset=pseudo_labeled_dataset_relabel + dataset_syne.train)

pseudo_labeled_dataset_relabel denotes using the unlabeled dataset, and dataset_syne.train denotes using the labels.

[lx7555]

Thank you. As your suggestion,

train_loader = get_train_loader(args, dataset_real, args.height, args.width, args.batch_size, args.workers, args.num_instances, iters, trainset=pseudo_labeled_dataset_relabel)

mean real data using the unlabeled dataset (pseudo_labeled)

while train_loader = get_train_loader(args, dataset_real, args.height, args.width, args.batch_size, args.workers, args.num_instances, iters, trainset=pseudo_labeled_dataset_relabel + dataset_syne.train) mean real data using the labeled dataset (real_labeled)

[WangWenhao0716]

No.

train_loader = get_train_loader(args, dataset_real, args.height, args.width,
args.batch_size, args.workers, args.num_instances, iters,
trainset=pseudo_labeled_dataset_relabel)

means real data using the unlabeled dataset (pseudo_labeled)

train_loader = get_train_loader(args, dataset_real, args.height, args.width,
args.batch_size, args.workers, args.num_instances, iters,
trainset=pseudo_labeled_dataset_relabel + dataset_syne.train)

means real data using the unlabeled dataset (pseudo_labeled) while syne data using the labeled dataset.

[lx7555]

Thank you! So as you mentioned, how to use "adding labeled real-world data to synthetic data" in your paper? I mean how to change the label state for real-world data? Best wishes

[lx7555]

Maybe? train_loader = get_train_loader(args, dataset_real, args.height, args.width, args.batch_size, args.workers, args.num_instances, iters, trainset=dataset_real.train + dataset_syne.train

means real data using the labeled dataset (real-labeled) while syne data using the labeled dataset.

[WangWenhao0716]

Yes.

[lx7555]

Yes.

Thank you