Purpose of mask_ph about soft-attention-image-captioning HOT 3 CLOSED

each element in the mask variable is either 1 or 0 depending on the ground truth sentence length.
The function finalCaptions() in utils.py computes the sentence length, producing a mask of 1s equal to the length of the sentence and zeros until the mask sentence length.

from soft-attention-image-captioning.

hi, sorry for the super late reply

The reason to do masking is to exclude the trailing captions. If we didn't mask out the redundant captions, the loss would include these information which basically is the wrong loss calculation.

AFAIK This is a common approach to deal with the inability of variable length output with LSTM.

Thanks!

from soft-attention-image-captioning.

Hi Mark,
That makes perfect sense! Thanks. I'm going to check out training a model today
Best wishes!

from soft-attention-image-captioning.