These are chat archives for beniz/deepdetect

6th
Aug 2016
Emmanuel Benazera
@beniz
Aug 06 2016 07:31
@sixsamuraisoldier you need to apply the diff between our custom Caffe and the official one
if you need functionalities it is better to describe them directly
if you have your own custom Caffe, it is up to you I guess
Tapabrata Ghosh
@sixsamuraisoldier
Aug 06 2016 20:04
@beniz Thanks for the response. I'm looking for LSTM layers and if possible, an OpenCL backend.
Emmanuel Benazera
@beniz
Aug 06 2016 20:05
The lstm layer has been merged into Caffe
Tapabrata Ghosh
@sixsamuraisoldier
Aug 06 2016 20:05
@beniz I see, has it also been merged into your custom Caffe?
Emmanuel Benazera
@beniz
Aug 06 2016 20:06
yes, support from DD is pending, see #140
as for opencl it d be a good thing to have. There's a Caffe official branch for it.
It has not been merged on our side as it'd be better to have some company to step in to support it: maintaining both cuda and opencl together requires some man hours.
Tapabrata Ghosh
@sixsamuraisoldier
Aug 06 2016 20:08
Yeah, that's true.
If I was to take the OpenCL branch and apply the four changes listed on the README for your custom version could it work with DD?
Emmanuel Benazera
@beniz
Aug 06 2016 20:11
if you merge the opencl branch into the custom one , it could... for sure you'll have to put some code in, if only for fixes.
what are you looking for in fact?
Tapabrata Ghosh
@sixsamuraisoldier
Aug 06 2016 20:15
@beniz I just prefer not to be locked into CUDA and the
nvidia ecosystem
AMD cards are also a decent bit cheaper which is nice also.
Plus, two big things coming in that make me want to keep the possibilities open for AMD card usage.
1) Native FP16 support on Polaris consumer cards. Although Pascal supports native FP16, it's crippled on all consumer cards, including the Titan X. Although there's no compute acceleration, it effectively doubles memory size and bandwidth.
2) Vega cards coming in Q42016/Q2017 are rumored to support FP16 as well and there's a decent chance they might support 2:1 compute acceleration. More importantly however, it is confirmed that they will be using HBM2, which gives a very large bandwidth boost (up to 1TB/s) which would allow us to get work done quite a bit faster.
Emmanuel Benazera
@beniz
Aug 06 2016 20:37
OK, I knew for 1.
I'm all for AMD, the more options the better, the more open the better! On our side I am waiting for a commercial incentive, otherwise it is a lot of work, maintaining and writing the missing kernels.
I have plans to support FP16 via Caffe, but have just started studying the matter. We have a few additions coming first and that are unrelated.
Emmanuel Benazera
@beniz
Aug 06 2016 20:44
What are the patches of interest to you in our version of Caffe ? I feel like you don't need DD, only Caffe, no ?
Tapabrata Ghosh
@sixsamuraisoldier
Aug 06 2016 21:49
@beniz Actually, DD is probably the biggest thing I need. The ability to set up a server and then work from anywhere I have internet access with my laptop is very big and one of the reasons I started using DD. Also, I'm thinking about creating a web demo with DD that anyone can use, DD is the only open source deep learning REST API I could find.
Emmanuel Benazera
@beniz
Aug 06 2016 22:35
There are others bit they're a bit more narrow and/or proprietary
As for opencl support you can open an issue. I doubt there's enough interest just now, but at least it gives the opportunity to gather steam