Skip to content


Folders and files

Last commit message
Last commit date

Latest commit



32 Commits

Repository files navigation

Please be reffered to our new repository for codes of generalized zero-shot learning (GZSL).

Modification history

10/25/2016: update the Demo code to clarify the hyper-parameter range

8/19/2016: correction for

Zero-shot learning codes & data

You can now download the codes (

You can also download a baseline method ConSE (

We provide the GoogLeNet features for three datasets (AwA, CUB, and SUN) in

Note that on SUN, which has 717 classes in total, we split into ~646 seen classes and ~71 unseen classes, following Lampert et al., PAMI 2014. This is a much harder task than some other work that tests on only 10 unseen classes. The results of our method on such a simpler task are in the Suppl., where we have accuracy around 90%.

Feel free to contact me and let me know if you have any question about the codes. (Email:

Extra details

Here we provide extra details for our CVPR 2016 paper:

Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha, "Synthesized classifiers for zero-shot learning," CVPR, 2016.

You can find links to our paper and the supplementary material via

Class splitting and semantic information for ImageNet

We provide the details of classes (seen, 2-hop, 3-hop, All) for the ImageNet zero-shot learning task as well as the word vectors in

You can also download it from

Features for ImageNet

For extracting the GoogLeNet features, we follow:

The name of the network should be [CAFFE_ROOT] + 'models/bvlc_googlenet/bvlc_googlenet.caffemodel'.

Note that for the 1000 seen classes, we extract features from images in ILSVRC2012. For the 20842 unseen classes, we extract features from images in ImageNet2011release.


No description, website, or topics provided.






No releases published


No packages published