2
1

Does anyone know if there is an implementation of RNNLM Toolkit by Tomas Mikolov on Python with Theano? Thanks for any clues!

asked Aug 06 '13 at 07:31

Ira%20Korshunova's gravatar image

Ira Korshunova
1315712

2

I don't know about any implementations for that specific toolkit but the documentation of the 'scan' function in theano gives examples of recurrent neural networks. As far as I know, the original RNNs by Tomas Mikolov are standard RNNs with softmax outputs and truncated backprop through time training.

(Aug 06 '13 at 07:45) Philemon Brakel

One Answer:

Take a look at Taylor Graham's implementation:

https://github.com/gwtaylor/theano-rnn

It's quite general so you can create language model feeding it with words using one-hot encoding. Though for large dictionaries it would be expensive. If it's a problem then you could consider character-level n-grams but it would require to learn much longer dependencies.

As for me theano is quite complicated at the first glance. So be patient while going through the code :).

What task are you trying to solve with your LM? I'm trying to implement char-based RNN right now. First I tried to do it with Theano but then found myself implementing pure C cuda code. Send me a message if you have any questions (my email could be found at my github profile: https://github.com/lightcaster ).

answered Aug 08 '13 at 08:16

Konstantin's gravatar image

Konstantin
34181218

edited Aug 08 '13 at 13:03

Leon%20Palafox's gravatar image

Leon Palafox ♦
40857194128

I took the liberty to edit your link, since it has a bad formatting and is not directing to the right site.

Thanks a lot for the great answer

(Aug 08 '13 at 13:02) Leon Palafox ♦

Thanks Leon. Due to Russian's brand new censorship system some of cloudflare's ips was blacklisted. So I have some problems with access to metaoptimize.com/qa and its page formatting (I can see only a bare html). Hope they unblock it soon.

(Aug 09 '13 at 03:19) Konstantin

Thank you very much! It's exactly what I was looking for

(Aug 09 '13 at 05:27) Ira Korshunova
Your answer
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.