pooled_rep

[Ali TT]

Hi, 

sorry to trouble you again, I read the code, but I can't what's the meaning of this step

pooled_rep = replace_masked_values(things_to_pool,answer_mask[...,None], -1e7).max(2)[0]#  things_to_pool :[96,4,29,1536]     answer_mask [96, 4, 29]

I found the shape of pooled_rep is different, things_to_pool[batch_size, answe_num, len, rnn_size*3], pooled_rep[batch_size, answe_num, batch_size, answe_num], could you please help to explain if it's possible, thanks a lot!!

best regards!

[Rowan Zellers]

Hi there,

See the "Reasoning" paragraph in the VCR paper. We max-pool the three bidirectional lstms over time, so the end result should be [batch_size, 4, rnn_size*3].

thanks,

Rowan