Caffe2: A New, Open-Source Deep Learning Framework From Facebook

[phoronix]

Phoronix: Caffe2: A New, Open-Source Deep Learning Framework From Facebook

Facebook just announced Caffe2, a new deep learning framework developed in cooperation with NVIDIA and other vendors...

http://www.phoronix.com/scan.php?page=news_item&px=Caffe2-Announced

[dstaubsauger]

Aaaaand as soon as Facebook gets on board, it includes their patents agreement in the github repo, that you have to accept and which restricts the bsd license. same situation as with react.js and their other open source projects.

http://react-etc.net/entry/your-license-to-use-react-js-can-be-revoked-if-you-compete-with-facebook

https://www.quora.com/Are-people-really-worried-about-using-React-and-ReactNative-with-the-language-thats-in-Facebooks-PATENTS-file

[Serafean]

Aaaaand as soon as Facebook gets on board, it includes their patents agreement in the github repo, that you have to accept and which restricts the bsd license. same situation as with react.js and their other open source projects.

http://react-etc.net/entry/your-license-to-use-react-js-can-be-revoked-if-you-compete-with-facebook

https://www.quora.com/Are-people-rea...s-PATENTS-file

I'll just leave this here : https://code.facebook.com/pages/8509..._fb_noscript=1
But still food for thought. (Does License language trump FAQ explaining said License?)

[nslay]

Meanwhile, Microsoft's CNTK has been doing distributed multi-GPU since forever. It's also one of the easier, if not easiest, ones to run in an HPC environment (compared to mxnet, the only other option that I know of for distributed multi-GPU).

But nobody seems to be using CNTK. They all stick to Caffe and it's probably because all the interesting work and models were developed on Caffe. Caffe is also a nightmare because every single grad student and/or scientist that touches it, forks it off and makes heavy modifications effectively resulting in completely different version of Caffe not compatible with the original. So, for example, if you wanted to build upon Faster R-CNN, you best bet get the forked version of Caffe mentioned in the paper. Holistically Nested Networks? Also another forked Caffe incompatible with the original. It's an absolute mess that infects many aspects of computer vision research because all the high impact papers build on Caffe. But some of the models have been ported to CNTK, Torch, mxnet, TensorFlow, but nobody cares and these remain not as well used (in computer vision at least). So Caffe2? What does it offer that conquers these problems? It may be doomed to be obscure like CNTK, Torch, TensorFlow and mxnet. No matter how easy it is use or technologically superior it is.

[fuzz]

Safe to assume it can't use OpenCL? Dead on arrival if so.

[Cape]

Why the fuck do you need an AI for FaceBook mobile app?

[Guest]

Why the fuck do you need an AI for FaceBook mobile app?

So that it can learn more optimal times to tap your mic/camera for less battery drain.

[starshipeleven]

Why the fuck do you need an AI for FaceBook mobile app?

To learn user behaviour in some field or whatever. Consider that it is not running in the app, and that it's a specialist AI not a generalist. Specialist AIs are the programs that auto-tag images with the correct content for example. Or that pattern user behaviour and suggest the best full answer as auto-completion in chats (Google Hello is a good example).

[Serafean]

Why the fuck do you need an AI for FaceBook mobile app?

because http://www.truthhawk.com/is-facebook...-free-society/

[M@yeulC]

Safe to assume it can't use OpenCL? Dead on arrival if so.

Well, there is some CUDA code in there, so there's that.
Then, it looks like they advertise RaspberryPi support, but probably CPU only.

If you plan to use GPU instead of CPU only, then you should install NVIDIA CUDA and cuDNN, a GPU-accelerated library of primitives for deep neural networks. NVIDIA's detailed instructions or if you're feeling lucky try the quick install set of commands below.

Yes, it's a shame. Unfortunately, this seems to be the norm rather than not. I guess nVidia has enough money to spare to invest in their lock-in strategy.