buleli-ai / gmn-zsl

This is the official repository for the Gradient Matching Generative Networks for Zero-Shot Learning paper published at IEEE CVPR 2019.

Paper link: CVPR Open Access

Bibtex entry:

@InProceedings{Sariyildiz_2019_CVPR,
	author = {Bulent Sariyildiz, Mert and Gokberk Cinbis, Ramazan},
	title = {Gradient Matching Generative Networks for Zero-Shot Learning},
	booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
	month = {June},
	year = {2019}
} 

After the camera ready deadline, we run several more experiments for a follow-up work. We simply enlarged the hyper-parameter space, for instance, (i) we further tuned the beta parameters of the ADAM optimizer and some of the GMN parameters in GMN training (ii) replaced ReLUs with LeakyReLUs in the generator and the discriminator networks, (iii) decayed learning rates during classifier training stages. Naturally, with these changes we obtained slightly higher scores on the CUB, SUN and AWA datasets. Besides, we found the followings:

• attribute concatenation works better than modeling latent noise spaces on the SUN dataset.
• When tuned to maximize the harmonic mean score, a linear classifier performs slightly better than a bilinear multi-modal embedding classifier in the AWA and SUN datasets. Therefore, to be fairly comparable with the SOTA approaches on the SUN dataset, we give our best results on the SUN dataset which are obtained by the attribute concatenation method and a linear classifier as in Felix et al. ECCV-2018 or Xian et al. CVPR-2018. We will publish our latest results on a follow-up work, meanwhile, we suggest practitioners to take the results reported in this repository as a baseline when comparing GMN with their own approaches.

The scripts provided in this repository are designed to directly re-produce the scores that we give below.

Zero-Shot Learning Scores

	CUB	SUN	AWA
Updated	67.0	61.1	72.0
Paper	64.3	63.6	71.9

Generalized Zero-Shot Learning Scores

	CUB			SUN			AWA
	u	s	h	u	s	h	u	s	h
Updated	54.7	58.4	56.5	50.3	37.2	42.8	57.1	81.3	67.1
Paper	56.1	54.3	55.2	53.2	33.0	40.7	61.1	71.3	65.8