Question about Gumbel trick

Hi! Thanks for your solid and insightful work.

I am checking the implementation of data selection but didn't find the place corresponding to this equation. The implementation didn't use the exp operation and normalize by expsum, which I think is necessary for the Gumbel trick. Could you explain a little bit? Thank you very much!

QuRating/data_tools/select_subset.py

Lines 238 to 241 in 828b96a

    
               metrics = metrics / args.temperature 
        
           if args.sample and args.temperature != 0.0: 
        
               metrics += np.random.gumbel(size=len(metrics)) # Use topk-gumbel trick

	metrics = metrics / args.temperature

	if args.sample and args.temperature != 0.0:
	metrics += np.random.gumbel(size=len(metrics)) # Use topk-gumbel trick