Will it support tensorflow backend?

[unacao]

Is there any timeline for supporting models trained with TensorFlow backend?

[moof2k]

I have no plans to work on this anymore (at the moment anyways).

There are two issues here, Theano and TensorFlow store matrices differently (row major vs column major). They also implement convolution slightly differently, so the implementation is not directly interchangeable.

There is hope however, see https://github.com/keras-team/keras/wiki/Converting-convolution-kernels-from-Theano-to-TensorFlow-and-vice-versa

[arquolo]

I've forked this to rewrite with supporting models trained with TensorFlow backend in latest Keras (2.1.5), see arquolo/kerasify. Now it's simplier for contributing, because there're classes in separate source files. Also it uses some features from C++17.

[patrikhuber]

@arquolo I was interested in looking at your fork, but the link is a 404 now. Did you decide to make the project private (or give it up)?

[arquolo]

@patrikhuber No, I just moved it to gitlab

[patrikhuber]

@arquolo Cool, thanks for the link!

[XiuSdk]

Tensor tmp(weights_.dims_[0], in->dims_[1] - st_nj - st_pj, in->dims_[2] - st_nk - st_pk); the value of 'in->dims_[2] - st_nk - st_pk ' is negative number when i input the tensor(i,j,k), how to solve it?