search

KunpengLi1994 / VSRN

PyTorch code for ICCV'19 paper "Visual Semantic Reasoning for Image-Text Matching"
288 stars 47 forks source link

encoder_hidden #3

Closed kywen1119 closed 4 years ago

kywen1119 commented 4 years ago

I notice that the you refer to the dimension of "encoder_hidden" in your code DecoderRNN.py as (num_layers num_directions, batch_size, dim_hidden). However, if it's extracted by GRU, the dimension of hidden_state should be (num_layers num_directions, seq_len, dim_hidden). This confuses me when trying to run your codes.

KunpengLi1994 commented 4 years ago

It is obtained from this line of the Encoder. If you check the Pytorch GRU code, "hidden" actually represents the hidden state for the last time step not for all time steps.