Zero-shot Learning PyTorch Implementation

Zero-shot classification, segmentation implementation in PyTorch

Supporting (re-)implementation

[1] SPNet

Zero-shot Learning?

There is a field called 'few shot learning', in which only few number of training samples are given.

Zero-shot learning is a recognition field, which predicts unseen classes without a positive sample.

Word-Embedding as External Knowledge

In order to recognize an unseen class, we need any external knowledge about the unseen class. Recently many works (e.g., [1]) leverages word-embedding as external knowledge. In this repository, we use (Wikipedia or Common Crawl) pretrained Fasttext [2].

Zero-Shot Classification

All classification models supported by Torchvision is available (e.g., ResNet).

Zero-Shot Segmentation

Deep base ResNet + DeeplabV2 + SPNet is supported.

References

[1]: http://openaccess.thecvf.com/content_CVPR_2019/papers/Xian_Semantic_Projection_Network_for_Zero-_and_Few-Label_Semantic_Segmentation_CVPR_2019_paper.pdf

[2]: https://fasttext.cc/docs/en/english-vectors.html