pylearn2 library is not sustained any more, can it be adapted into keras or tensorflow?

[lxmwust]

as the title, it is difficult to use pylearn2 nowadays, since many libraries it depends have been updated into new versions, and the existing pylearn2 is not compatible with them.

[johnarevalo]

I think the heaviest part to migrate is the preprocessing that uses Local and Global contrast norm implementations of pylearn2. FWIW, I've partially moved the code from python2 to python3 (still experimental), including the necessary updates in pylearn2 and cnn-bcdr to make things work. You can try it with:

pip3 install --user --ignore-installed git+https://github.com/theano/[email protected]
git clone git://github.com/johnarevalo/pylearn2.git
cd pylearn2
python3 setup.py develop
pip3 install pillow
pip3 install sklearn
pip3 install matplotlib
pip3 install h5py
pip3 install shapely

This works only for preprocessing and oversampling. The training won't work. However, you could generate the preprocessed dataset and then implement the CNN in other framework. Recall that the CNN is pretty simple: 2 conv layers + 1 fully connected layer + logistic regression layer.

[lxmwust]

Thanks for the quick response. Well, I have tried twice to make the code runnable with anaconda virtual environment. It still doesn't work due to some compatible problems (such as theano 1.0 has changed its support for gpu from 0.9, and sclearn version problem, the error reported from theano or pylearn is not understandable).