agupta74

In matchutils.py, tf.reduce_max and tf.reduce_mean (e.g Lines 150 and 151 -- there are other places in matchutils.py using these functions) are computed without masking. So if you vary the zero...

With regard to the word/char embeddings for "zero" word/char tokens in the zero-padded query/passage, should these zero word embeddings not be a zero vector ? Any non-zero values for zero...

I agree that you mask out the values. However for the mean calculation the number of padded zeros will impact the mean value even though the sum does not get...

Thanks ! Also please look into the max calculation issue as mentioned in my comments above

@zhiguowang Any update on this ? I can submit a PR for the fix if that helps so that you can verify the fix and perform any other testing that...