search

malmaud / TensorFlow.jl

A Julia wrapper for TensorFlow
Other
886 stars 110 forks source link

Support 2.1.0 and/or 1.15.2 #513

Open PallHaraldsson opened 5 years ago

PallHaraldsson commented 5 years ago

https://github.com/tensorflow/tensorflow/releases/

1.13.2 has some security issues fixed. [EDIT: 1.15 is also out.]

diversable commented 4 years ago

Is there any current work being done to support TF 2.0? Or have the maintainers moved on to other things...? I understand everyone is busy - I’m just wondering. And if you need help, well, I’m not used to working with C API’s, but I’ll do what I can - if others are interested in keeping this library up-to-date. Lemme know!

PallHaraldsson commented 4 years ago

Possibly helpful: https://discourse.julialang.org/t/state-of-deep-learning-in-julia/28049

Hi,

@n0c1cept0r, @chengchingwen while I opened this, and I've only tried TensorFlow once and never from Julia I don't have any immediate needs for this wrapper, and likely will not look more into this, at least not soon.

Maybe it's not either or, I just don't know, but (Julia-native) Flux.jl and Knet.jl seem to be the future or current use for deep learning in Julia. See also XLA.jl for TPUs. It seems all of these are alternatives to mainstream (Ptyhon) deep learning libraries. Ið'm not sure, maybe some combination of libraries is possible, including TensorFlow.jl?

Most recently I found (days ago, while in development since 2017 if I recall), not sure if good alternative/competitor: https://github.com/hshindo/Merlin.jl

Let me know if there is still use for Tensorflow[.jl] (for new projects). It seemed excellent, and better with this wrapper. For sure for those who need, 2.0 seems the future, I guess more work to support here, while 1.14 I guess simple to resolve.

Today I also found new (PyTourch related) https://github.com/dfdx/Lilith.jl (and seemingly outdated/very old Strada.jl).

chengchingwen commented 4 years ago

@PallHaraldsson Hi, I don't really use TensorFlow.jl, but I need to read a bunch of Tensorflow checkpoint files in Julia and turn them in to BSONs (for my package Transformers.jl). Since people told me that reading directly from the checkpoint file will kill myself, I turn into using the TensorFlow.jl to read those file. The new TF2 also change a few api, so the best way for me is to update the TensorFlow.jl with TF2 support.

PallHaraldsson commented 4 years ago

I'm not sure I can help, I guess updated TensorFlow.jl would, not up to speed on these checkpoint files or if other software can/does support those. Note new ML libraries. I've found 3 in recent days (one of not new; see updated comment above, I had link to wrong library).

PallHaraldsson commented 4 years ago

Is there any current work being done to support TF 2.0?

@n0c1cept0r, I was trying and almost got TF 2.1 to work (i.e. for TF 2.1 in Python, code here needs further changes for the download link seemingly); and got 1.15.0, not 1.15.2, to work locally, seemingly. See my PR.

PallHaraldsson commented 4 years ago

I found "Note: There is no libtensorflow support for TensorFlow 2 yet. It is expected in a future release."

It seems it provides the C API, TensorFlow.jl relies on, so supporting TF2 is more work than I like to take on now. I'm not sure if there's an easy workaround.

Sinansi commented 4 years ago

Any updates on this issue? Are you still working on Tensorflow 2 support?

PallHaraldsson commented 4 years ago

I'm not working on it. It seems I was almost there and simple to finish from what I documented, but I didn't really have any need for TF myself, so I was just looking a bit into this.