Deep RNN models with Theano.
Why 'DerpRNN'? Because 'DeepRNN' was already taken!
This is a smallish Python module for deep recurrent neural networks using Theano. Evolving through time and the layers is performed all in a nested theano.scan loop, so the implementation is in this way a bit different than in e.g. Keras. Here's a list of some of the features (of the
- Hidden recurrent layers can be SRU (standard recurrent layer) or GRU (all same type and shape!)
tanh"read in" layer before the recurrent layers
- Readout layer can be
- There's also a fully translation invariant version (
InvariantDeepRNN, but it's not quite done yet...)
Please see the Demo notebook for example usage!
Also discussed here.
Requirements are the usual scientific python ones, plus of course Theano. Also
python-midi is needed for processing the midi data. You may also need need cython.
There's a seup script, so you should be able to install the module and the dependencies by
pip install git+https://github.com/harpone/DerpRNN.
If that fails, you can try cloning the repo by
git clone https://github.com/harpone/DerpRNN and then
python setup.py build_ext --inplace to compile the cython modules (although you may not need to do that at all, depending on your machine).