hughperkins/char-lstm

char-rnn, tweaked to use Element-Research rnn modules

char-lstm's Issues

Bug in the code structure for Training

Hello,

I used the base code for testing my own code :
Training data : Batch size Tensor of size 50 x 3, where length (row) of sequence is 3 (float value). The LSTM has one input and one output. Example one row of the training batch : torch.Tensor({{0.1},{0.2}{0.3}}).

When passing as a single chunk batch to Sequencer Node (net) : updateOutput : input:size() is used as length of sequence which should be 3 but in my case return 50, which is wrong. Hence instead of passing a batch , single row of batch if passed works as per the logic and LSTM node is called 3 times (time step) instead of 50 times..,, This could be a problem in your code as well else Sequencer.lua logic need to handle batch 2D tensors.

Please suggest if I am wrong in understanding.....

Organization

Nice job on this. I am watching in hopes it will be easier to plug in bidirectional lstm's and some of the other options with ER. Why not move out the data loading functions to separate files so there is an option for char / word level batching without having the main file so complex?

char-lstm

This repo would get more hits if it were renamed char-lstm. IMHO