Pylearn2用于时间序列或序列预测的示例

Question

Can Pylearn2 be used for time series or sequence prediction of continuous numerical data? Can an LSTM recurrent neural network in Pylearn2 be used for this? If so, can someone post an example code in Pylearn2/Theano/Python?

Answer 1

My understanding is that PyLearn2 is still not great for any kind of recurrent network, though I believe they are intending to improve support for these kinds of models.

Having said that, there is experimental support, including an LSTM implementation.

Take a look in the PyLearn2 source code in the directory pylearn2/sandbox/rnn , and in particular at the contents of pylearn2/sandbox/rnn/models/rnn.py where you'll find an LSTM implementation.

Because of its experimental nature, this code may not work properly, may not be supported fully, and the documentation may be incomplete or inaccurate.

If you're willing to forego the intended ease of use benefits of PyLearn2 and work at a more detailled level then recurrent neural newtworks can be implemented just fine in Theano. There are many tutorials for this, including:

• LSTM Networks for Sentiment Analysis
• Modeling and generating sequences of polyphonic music with the RNN-RBM
• Recurrent Neural Networks with Word Embeddings