jingli9111 / goru-tensorflow Goto Github PK

Hey I read your paper and was super interested in trying out the GORU in the charRNN task. However, when I use your code I get this error:

ValueError: Dimension size must be evenly divisible by 126 but is 6350 for 'RNN/rnn_decoder/rnn_decoder/multi_rnn_cell/cell_0/goru_cell/while/cond/cond/Reshape_2' (op: 'Reshape') with input shapes: [50,127], [3] and with input tensors computed as partial shapes: input[1] = [?,63,2].

I'm not really sure where this is coming from. Because it is trying to divide 6350 by 126, but the input shape is [50, 127] and 6350 is divisible by 127.

Thank you for your research in this area!

1. I use GORU instead of GRU in my 3-layers denoiser. But in comparison with GRU, GORU works very slow! Is it any way to speed up?

2. I can't outperform GRU in my tests with GORU. GORU is much worse, it stopped at 1.0 loss (I use Itakura-Saito distance), but GRU stopped at 0.5. Maybe I miss something? I just replaced GORU.GRUcell to GORU.GORUCell in my code. I have 512-FFTs audio features as input to neural net.

GORU loss: