I think this a bug. caption.py 140

[tigerbrother222]

incomplete_inds = [ind for ind, next_word in enumerate(next_word_inds) if next_word != word_map['']] incomplete_inds always is [0,1,2,3,4] .
and then complete_inds = list(set(range(len(next_word_inds))) - set(incomplete_inds)) complete_inds is empty so complete_seqs is null.

my Model out is error

[ArborealSky]

I have the same bug in caption.py . Did you fix it ?

[ArborealSky]

i fix it , you should consider the situation that your model is too weak to generate the <end> symbol for ending your sentences.

[ArborealSky]

please add code

if len(complete_seqs_scores) == 0:
        return seqs[0], 0

before

i = complete_seqs_scores.index(max(complete_seqs_scores))

in fuction caption_image_beam_search

[gsoykan]

why don't you use it like the following ? Is there any specific reason why you are returning 0 instead of matching attention map?

if len(complete_seqs_scores) == 0:
    return seqs[0], seqs_alpha[0]

[ArborealSky]

why don't you use it like the following ? Is there any specific reason why you are returning 0 instead of matching attention map?

if len(complete_seqs_scores) == 0:
    return seqs[0], seqs_alpha[0]

You are right. Actually, the code in my project doesn't need to care about seqs_alpha. So I just used '0' to replace its return. Your code are more complete and better.