TOHAN: A One-step Approach towards Few-shot Hypothesis Adaptation

Hi, this is the core code of our rencent work "TOHAN: A One-step Approach towards Few-shot Hypothesis Adaptation" (NeurIPS 2021 Spotlight, https://arxiv.org/abs/2106.06326). This work is done by

Haoang Chi (NUDT), haoangchi618@gmail.com
Dr. Feng Liu (UTS), feng.liu@uts.edu.au
Dr. Wenjing Yang (NUDT), wenjing.yang@nudt.edu.cn
Dr. Long Lan (NUDT), long.lan@nudt.edu.cn
Dr. Tongliang Liu (USYD), tongliang.liu@sydney.edu.au
Dr. Bo Han (HKBU), bhanml@comp.hkbu.edu.hk
Prof. William Cheung (HKBU), william@comp.hkbu.edu.hk
Prof. James Kwok (HKUST), jamesk@cse.ust.hk

Software version

Torch version is 1.7.1. Python version is 3.7.6. CUDA version is 11.0.

These python files, of cause, require some basic scientific computing python packages, e.g., numpy. I recommend users to install python via Anaconda (python 3.7.6), which can be downloaded from https://www.anaconda.com/distribution/#download-section . If you have installed Anaconda, then you do not need to worry about these basic packages.

After you install anaconda and pytorch (gpu), you can run codes successfully.

TOHAN

Please feel free to test the TOHAN method by running main.py.

Specifically, please run

CUDA_VISIBLE_DEVICES=0 python main.py

in your terminal (using the first GPU device).

Our pre-trained models can be downloaded in the following link.

https://drive.google.com/drive/folders/1IIRJSDvJ9WYVGbZRNOpkU0STdVfmcjfb?usp=sharing

Two bugs

We just found two bugs in the source codes regrettably.

We did not use encoder.eval() and classifier.eval() to completely freeze the source model in the generation process.
The implementation of the augmented L1 distance degraded into the naive L1 distance.

We have repaired the two bugs and the modified results are coming soon. Sorry for the trouble caused to you.

Original reproducible results

For digital tasks,

Task	Method	1	2	3	4	5	6	7
M2S	S+FADA	25.6	27.7	27.8	28.2	28.4	29.0	29.6
	T+FADA	25.3	26.3	28.9	29.1	29.2	31.9	32.4
	TOHAN	26.7	28.6	29.5	29.6	30.5	32.1	33.2
S2M	S+FADA	74.4	83.1	83.3	85.9	86.0	87.6	89.1
	T+FADA	74.2	81.6	83.4	82.0	86.2	87.2	88.2
	TOHAN	76.0	83.3	84.2	86.5	87.1	88.0	89.7
M2U	S+FADA	83.7	86.0	86.1	86.5	86.8	87.0	87.2
	T+FADA	84.2	84.2	85.2	85.2	86.0	86.8	87.2
	TOHAN	87.7	88.3	88.5	89.3	89.4	90.0	90.4
U2M	S+FADA	83.2	84.0	85.0	85.6	85.7	86.2	87.2
	T+FADA	82.9	83.9	84.7	85.4	85.6	86.3	86.6
	TOHAN	84.0	85.2	85.6	86.5	87.3	88.2	89.2
S2U	S+FADA	72.2	73.6	74.7	76.2	77.2	77.8	79.7
	T+FADA	71.7	74.3	74.5	75.9	77.7	76.8	79.7
	TOHAN	75.8	76.8	79.4	80.2	80.5	81.1	82.6
U2S	S+FADA	28.1	28.7	29.0	30.1	30.3	30.7	30.9
	T+FADA	27.5	27.9	28.4	29.4	29.5	30.2	30.4
	TOHAN	29.9	30.5	31.4	32.8	33.1	34.0	35.1

For objective tasks,		S+FADA	T+FADA	TOHAN
CF2SL	72.1	71.3	72.8
SL2CF	56.9	55.8	56.6

Modified results

For digital tasks,

Task	Method	1	2	3	4	5	6	7
M2S	S+FADA	27.2	29.0	28.5	28.7	29.4	30.4	32.1
	T+FADA	27.1	28.7	27.7	26.6	28.4	28.7	29.8
	TOHAN	28.0	29.7	28.6	28.8	30.2	31.8	33.4
S2M	S+FADA	81.0	84.1	84.6	85.1	87.8	88.1	90.3
	T+FADA	80.1	80.5	84.8	80.9	86.7	87.1	90.0
	TOHAN	80.6	83.4	85.8	88.9	88.2	90.6	91.1
M2U	S+FADA	81.8	84.9	86.5	89.9	90.8	91.9	92.1
	T+FADA	77.3	83.2	87.6	89.9	90.6	93.2	92.7
	TOHAN	83.8	86.8	88.6	89.8	91.0	93.2	93.0
U2M	S+FADA	83.2	83.5	82.0	83.3	85.1	85.3	85.5
	T+FADA	83.5	84.2	83.5	84.2	85.3	85.1	84.4
	TOHAN	83.9	84.1	84.2	83.9	86.0	86.0	86.5
S2U	S+FADA	83.9	85.8	88.8	91.4	91.0	91.3	92.3
	T+FADA	84.1	87.5	88.8	91.1	91.4	91.9	91.5
	TOHAN	84.4	88.8	89.2	92.1	91.6	92.2	92.8
U2S	S+FADA	26.5	27.3	27.2	29.5	30.2	30.7	31.1
	T+FADA	27.1	27.0	29.2	30.0	31.6	33.9	35.9
	TOHAN	28.3	28.9	32.2	30.8	32.0	34.8	37.4

For objective tasks,		S+FADA	T+FADA	TOHAN
CF2SL	72.2	71.0	72.5
SL2CF	57.1	55.3	57.0

After fixing the bugs, most of the results improved, and we will update a new arXiv version soon.

Citation

If you are using this code for your own researching, please consider citing

@inproceedings{chi2021tohan,
 author = {Chi, Haoang and Liu, Feng and Yang, Wenjing and Lan, Long and Liu, Tongliang and Han, Bo and Cheung, William and Kwok, James},
 booktitle = {Advances in Neural Information Processing Systems},
 title = {TOHAN: A One-step Approach towards Few-shot Hypothesis Adaptation},
 year = {2021}
}

Acknowledgment

This work was partially supported by the National Natural Science Foundation of China (No. 91948303-1, No. 61803375, No. 12002380, No. 62106278, No. 62101575, No. 61906210) and the National Grand R&D Plan (Grant No. 2020AAA0103501). FL would also like to thank Dr. Yanbin Liu for productive discussions.

Haoang97 / TOHAN

readme