CNTK
CNTK copied to clipboard
Is CNTK still alive?
I'm evaluating CNTK for a commercial project and I'm a bit confused. Apparently, CNTK doesn't support current-gen NVidia GPUs (RTX 2070, RTX 2080) at all (Issue #3566) and there's no fix for this.
Is this project still actively maintained?
@nikie78 I’m also curious. In my head, I’m still thinking should I even care about following CNTK? Or should we all move to PyTorch and use ONNX?
I’ve the feeling inside Microsoft they care about ONNX and PyTorch. It’s a pity because I still think CNTK is cool 😎 It’s only that nobody is writing papers making use of CNTK.
I have also dedicated time in learning CNTK and I have used it in some of my works. I think that if CTNK is not going to be maintained any longer we should be notified so we can look for a replacement.
They don't comment here much but CNTK continues to be updated. @nikie78, regarding your issue, have you tried a build of the latest update to confirm this issue hasn't been addressed? Admittedly they could mention it, but I suspect they have been looking into it. There is an update to Cuda 10...
I don't have the hardware to test but for C# users pre-release versions of the Nuget are available to test.
I can confirm that master supports RTX-series for the C++ API. Still waiting for official release though.
You can checkout this comment by kedeng, he's one of the maintainers. The strategy has been shifted to focus more on onnx.
From what i observe on master, onnx is being actively integrated in to cntk, most recently the entire onnxruntime has just been added in. So in that sense, cntk will have leading integration with onnx, which is a good thing. Because any functions in the onnx standards will be incorporated in cntk too.
@delzac “ Because any functions in the onnx standards will be incorporated in cntk too.” Does it mean if some functions which are in PyTorch goes to ONNX, then CNTK will have to incorporate that? If so, it’s nice 👍
@arijit17 I think its safe to say that if its part of onnx standard, then yes, cntk will eventually have it.
You can take a look here to see how onnx is being converted to cntk. Most onnx op have a cntk equivalent, with some exception.
The maintainers are working with constraints. While the expected release date of ver 2.7 has slipped, it doesn't mean that it has been postponed indefinitely. Please be patient.
Quoting https://github.com/microsoft/CNTK#disclaimer
Today’s 2.7 release will be the last main release of CNTK. We may have some subsequent minor releases for bug fixes, but these will be evaluated on a case-by-case basis. There are no plans for new feature development post this release.