diyang opened a new issue #10805: SKIP RNN is incorrect in LSTnet
URL: https://github.com/apache/incubator-mxnet/issues/10805
 
 
   Thanks for your example and source code, it helps me a lot to develop my own 
LSTnet.
   However, I have noticed that there are some flaws in the implementation of 
Skip RNN of yours.
   
   The following is your way of implementing Skip RNN 
   `####################
       # Skip-RNN Component
       ####################
       stacked_rnn_cells = mx.rnn.SequentialRNNCell()
       for i, recurrent_cell in enumerate(skiprcells):
           stacked_rnn_cells.add(recurrent_cell)
           stacked_rnn_cells.add(mx.rnn.DropoutCell(dropout))
       outputs, states = stacked_rnn_cells.unroll(length=q, 
inputs=cnn_reg_features, merge_outputs=False)
   
       # Take output from cells p steps apart
       p = int(seasonal_period / time_interval)
       output_indices = list(range(0, q, p))
       outputs.reverse()
       skip_outputs = [outputs[i] for i in output_indices]
       skip_rnn_features = mx.sym.concat(*skip_outputs, dim=1)
   `
   What I have noticed is that this way will not actually create a skip rnn, 
and what's this RNN doing is to select the hidden output of the first hour of 
every day in a week.
   
   According to the paper, what it should like is that input variants regarding 
every hour of any given day should pair with the hidden output regarding the 
same hour of the previous day.
   mx.rnn.SequentialRNNCell can not handle this type of recurrent, you might 
need to make your own native way to formuate this type of RNN. 

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
[email protected]


With regards,
Apache Git Services

Reply via email to