Zero-Shot Logit Adjustment

This project contains the pytorch implemention for Zero-Shot Logit Adjustment.

Dependencies

Python 3.7
Pytorch = 1.0.1
Prerequisites
Dataset: please download the dataset, and change --dataroot to your local path.
Semantic: The semantics for AWA2, SUN, and APY are available in the dataset. please download the 1024-D CUB semantic and save it to the data path.

Train and Test

Please run the following commands to test on different datasets:

python ./WGAN+ZLAP.py --dataset AWA2 --attSize 85 --nz 85 --syn_num 10 --ratio 1000
python ./WGAN+ZLAP.py --dataset SUN --attSize 102 --nz 102 --syn_num 10 --ratio 60
python ./WGAN+ZLAP.py --dataset APY --attSize 64 --nz 64 --syn_num 10 --ratio 300
python ./WGAN+ZLAP.py --dataset CUB --attSize 1024 --nz 1024 --syn_num 10 --ratio 30 --class_embedding sent

The meaning of these args is

--dataset: datasets, e.g: SUN.
--attSize: size of semantic descriptors.
--nz: size of the Gaussian noise.
--syn_num: synthetic number for each unseen class.
--reatio: hyperparameter to control the seen-unseen prior (see Sec. 4.4 of the paper)

Results

We test our method in WGAN and CE-GZSL, and here are the results.

Method	AWA2			CUB			SUN			APY
Method	A^S	A^U	A^H	A^S	A^U	A^H	A^S	A^U	A^H	A^S	A^U	A^H
f-CLSWGAN	57.7	71.0	63.7	59.4	63.3	61.3	46.2	35.2	40.0	32.5	57.2	41.5
ZLAPWGAN	65.4	82.2	72.8	73.0	64.8	68.7	50.1	38.0	43.2	40.2	53.8	46.0
CE-GZSL	65.3	75.0	69.9	66.9	65.9	66.4	52.4	34.3	41.5	28.3	65.8	39.6
ZLAP+CE-GZSL	64.8	80.9	72.0	71.2	66.2	68.6	50.9	35.7	42.0	38.3	60.9	47.0

Citation

If you recognize our work, please cite:

@inproceedings{ijcai2022-114,
    title     = {Zero-Shot Logit Adjustment},
    author    = {Chen, Dubing and Shen, Yuming and Zhang, Haofeng and Torr, Philip H.S.},
    booktitle = {Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, {IJCAI-22}},
    publisher = {International Joint Conferences on Artificial Intelligence Organization},
    editor    = {Lud De Raedt},
    pages     = {813--819},
    year      = {2022},
    month     = {7},
    note      = {Main Track}
    doi       = {10.24963/ijcai.2022/114},
    url       = {https://doi.org/10.24963/ijcai.2022/114},
    }

Acknowledgment

We acknowledge the prior works f-CLSWGAN and CE-GZSL for their contributions to our work.

cdb342 / IJCAI-2022-ZLA

readme