Beam Search

Beam search implementation with examples optimizing BLEU score of COCO image captions.

Install

Clone this repo into a directory also containing the MS COCO Caption Evaluation repo.

Examples

See the example and results by running jupyter notebook coco-experiment.ipynb

Takeaways:

• Beam search initialized with the start-of-sentence token achieves a perfect BLEU score. (this also holds with 1 beam, i.e. greedy search)
• Considering words in the reference captions as actions instead of all words in the dictionary improves search speed by 1000 fold and produces identical results.