TensorFlow 2.0 plans about tensorsem

[vankesteren]

Hi Nick,

thank you for the interest! Yes, I am currently looking at several features to move towards in the future. Tensorflow 2.0 is definitely high up on the to-do list, and more generally optimizing the tensorflow backend (e.g., looking at whether using GPU compute can speed up everything, seeing if any other optimizations can be made in this file). Unfortunately right now I'm a bit busy with other projects and I'm not so very up-to-date on the whole tf1 - tf2 upgrade. I'm also not opposed to the idea of implementing the entire package in a different framework than tensorflow (torch, mxnet, or any other computation graph + autodiff + adam optimizer software) because I am not too sure about the whole R -> python -> tensorflow tower of cards -- but this requires quite some research.

If you know how to translate the current state of this package into tensorflow 2.0 I'd be very happy to receive a pull request!

One thing on the feature roadmap is casewise / stochastic gradient descent optimization. There is a (kind of old) branch called batch_processing which implements this, but it's pretty slow and it uses some intransparent data wrangling for the tfdatasets package.
Another thing I want to think about is how to let the user add their own tensors (penalties, alternative losses) to the computation graph without pre-programming them first as in these lines. Some kind of function pass-through in the tf_sem() function would be nice.

Ideally, there would be a CRAN release at some point too.

[nredell]

Bad follow-up on my part. Thanks for updating to version 2.0+. Looking forward to giving this a spin.