Run error about cmd

[wzell]

Hallo jindongwang,
nice to see Your interest in our project!

Unfortunately our code only supports the theano backend and not tensorflow. However, theano is only used in the file domain_regularizer.py. One simple fix would be to use the keras backend functionality, wherever theano is used (indicated by 'T.'), forcing it to use tensorflow. I.e. using

K.cast(i+2,'int32')

instead of

T.cast(i+2,'int32')

and

K.square(r - x2)

instead of

T.sqr(r - x2)

I have tested this solution with python 2.7 and keras 1.1.0. More details on the requirements can be found in requirements.txt. Downgrading keras works well if you are using pip with

pip install keras==1.1 --no-dependencies

and python 3.5 should work.

If You have futher problems I would be happy to help.

[jindongwang]

Thank you @wzell for your kind reply!
I tried your way and replaced all "T.x" with "K.x" under Keras 1.1, but it still did not work. Here's some of the error msg:

[wzell]

I think this has something to do with the change of the keras backend from theano to tensorflow, see e.g.
keras-team/keras#3945.
The change in the json file should help.

[jindongwang]

Thanks for that! In fact I'm very interested in the CMD metric and want to use it independently as MMD. Is this possible?

[wzell]

Yes, this works. See exp_office.py for an example.