Diverse Beam Search

Beam search, the standard work-horse for decoding outputs from neural sequence models like RNNs produces generic and uninteresting sequences. This is inadequate for AI tasks with inherent ambiguity — for example, there can be multiple correct ways of describing the contents of an image. To overcome this we propose a diversity-promoting replacement, Diverse Beam Search that produces sequences that are significantly different — with runtime and memory requirements comparable to beam search.

Arxiv Paper link: https://arxiv.org/abs/1610.02424

Diverse Beam Search Demo: http://dbs.cloudcv.org/captioning

Installing / Getting started

We use RabbitMQ to queue the submitted jobs. Also, we use Redis as backend for realtime communication using websockets.

All the instructions for setting diverse beam search from scratch can be found here

Interested in Contributing?

Cloud-CV always welcomes new contributors to learn the new cutting edge technologies. If you'd like to contribute, please fork the repository and use a feature branch. Pull requests are warmly welcome.

If you have more questions about the project, then you can talk to us on our Gitter Channel.

Acknowledgements

• Original Diverse Beam Search Code
• NeuralTalk2
• PyTorch