Anybody port this to TensorFlow or Pytorch? about deep_complex_networks

[obilaniu]

@austinmw It is my intention to port the vision code to PyTorch and torchvision, as Theano is being sunset and Kerosene has too many dependencies. I appreciate PyTorch's data loaders far more than Kerosene/Fuel.

It definitely won't happen within a week, but within a month probably.

When that happens I will push the new code here.

[lintongtong123]

mark, waiting for TensorFlow-Keras or PyTorch port

[harritaylor]

Hi, has there been any updates on this? I'd be happy to contribute if help is needed.

[chris1992212]

Has there been any updates now? I also want to transform it to Tensorflow, but I'm not sure whether I've done it right.

[planewave]

Same here, waiting for TensorFlow-Keras or PyTorch port

[veritas9872]

@lintongtong123 @planewave @harritaylor @obilaniu There is a very (extremely) basic implementation in Pytorch here.
@chris1992212 Also, a slightly less basic implementation in Tensorflow was made here.
I recall that there was another implementation in Keras using Tensorflow as its backend but I could not find the repo.
However, I think that it would be better to start from scratch and just copy some code from those two repositories as they are so basic. Only 2D convolutions are implemented and the Pytorch implementation does not even have proper weight initialization.

[JiachuanDENG]

conv2d part is trivial, Batch normalization should be the most difficult part to implement

[JesperDramsch]

I have a keras with Tensorflow fork of this. But it's mostly untested. I'd be happy to have contributions to it. https://github.com/JesperDramsch/keras-complex

It implements Conv1D to Con3D, Transpose, BN, Dense layers, Spectral Pooling. But it's very much research code from a non-CS major. Happy about help to get the standard up to par.
It's also TF1.x not TF2, so there's that.