LLP-VAT

Pytorch implementation of LLP-VAT

• Kuen-Han Tsai and Hsuan-Tien Lin. Learning from label proportions with consistency regularization. In Proceedings of the Asian Conference on Machine Learning (ACML), November 2020 [ pdf ]

Environment

• Python version: 3.6.2
• GPU: GeForce GTX 1080
• Prerequisite:

  pip install -r requirements.txt
  

Usage

Make sure to generate the LLP data before running the experiment of LLP-VAT.

Preprocessing

python -m llp_vat.preprocessing --dataset_name cifar10 --alg uniform --bag_size 64

Required arguments:

Parameter	Description
--dataset_name	svhn, cifar10 or cifar100
--alg	the bag creation algorithm, uniform or kmeans

Optional arguments:

Parameter	Description
--obj_dir	path to the proccessed object directory
--dataset_dir	path to the raw data directory

Arugments for the bag creation algorithm:

Algorithm	Parameter	Description
uniform	-b, --bag_size	number of instances in each bag
uniform	--replacement	whether the sample is with replacement
kmeans	--k, --n_clusters	number of clusters to be used
kmeans	--reduction	number of dimensions to keep
uniform, kmeans	--seed	pass an int for reproducible results

Experiment

python -m llp_vat.main --dataset_name cifar10 --alg uniform -b 64

Required arguments:

Parameter	Description
--dataset_name	svhn, cifar10 or cifar100
--alg	the bag creation algorithm, uniform or kmeans

Optional arguments:

Parameter	Description	Default
--obj_dir	path to the proccessed object directory	./obj
--dataset_dir	path to the raw data directory	./obj/dataset
--result_dir	path to the result directory	./results
--num_epochs	number of training epochs	400
--lr	value of learning rate	0.0003
--optimizer	adam or sgd	adam
--valid	ratio of the validation set	0.1
--seed	pass an int for reproducible results	0
--consistency_type	vat, pi or none	vat
--consistency	consistecny loss weight	0.05

Citation

@InProceedings{pmlr-v129-tsai20a, 
    title = {Learning from Label Proportions with Consistency Regularization}, 
    author = {Tsai, Kuen-Han and Lin, Hsuan-Tien}, 
    booktitle = {Proceedings of The 12th Asian Conference on Machine Learning}, 
    year = {2020} 
}