Bidirectional lstm example #2093

[xlvector]

This is an just simple example for bidirectional lstm.

[piiswrong]

@pluskid
We need a LSTM symbol implemented in python under mx.sym.LSTM so that it will be easier to replace LSTM in all examples with cudnn v5 lstm when we support it

[pluskid]

@piiswrong yes agreed. The currently used LSTM in ptb and speech are a variant with some simplification. An implementation of a full version needs the broadcast multiplication feature. Need to check what variant of the LSTM cell is cuDNN v5 providing. What is our current status of supporting cuDNN v5? I am reluctant to upgrade having heard that they breaks a lot of APIs.... also on servers I do not have permission to upgrade cuda driver to v8.

[antinucleon]

@pluskid We now support everything but LSTM cell in CuDNN v5. CUDA 7.5 is fine for CuDNN v5.

[antinucleon]

This file should not be modified.

[xlvector]

thanks, I will modify it

[xlvector]

Glad to hear that mxnet will have a full support of LSTM.

Another think I want to ask is, is it easy to integrate baidu wrap-ctc without using torch module?

[tqchen]

this is ready to be merged after combine the commits. see instructions http://mxnet.readthedocs.io/en/latest/how_to/contribute.html#how-to-combine-multiple-commits-into-one

[xlvector]

sorry to response later. I combined the commits but find its already merged.

Thanks all.

[giorgioercixu]

@xlvector I am working on make this example python2/3 compatible. However, I find that the data folder and gen_data.py are missing in the current master branch. Could you take a look at this?

[xlvector]

Thanks. The gen_data.py is very simple, it just generate whitespace seprated numbers in every line.

something like

for i in range(1000):
print ' '.join([x for x in sorted([random.randint(0, 100) for k in range 10)]

You can add this script in your PR.

[giorgioercixu]

ok I will put this gen_data into the lstm_sort.py directly. Thanks.