Tagged image for cuDNN

[ruffsl]

I know accessing releases of the cuDNN GPU Accelerated Deep Learning currently requires membership to the Accelerated Computing Developer Program, but would it be possible to have a tag of the cuda image with cuDNN included? Or would this be done under the DIGITS image you where talking of doing, @flx42 ?

[flx42]

Seriously, are you reading our minds? :)

I have good news: cuDNN v4 is going to be available soon, and it won't require any registration.

[ruffsl]

How are you anticipating to arrange the tag names, something like this?

Tag	From
cuda:7.5/latest	ubuntu:14.04
cuda:7.5-cudnn	cuda:7.5

Or maybe like:

Tag	From
cuda:7.5/latest	ubuntu:14.04
cudnn:4.0	cuda:7.5

i.e. would you bump cudnn up to its own repo name, or host it under a matching cuda or Digits tag? I'm thinking if a version of cudnn always ships with a given version of cuda, keeping it as a cuda tag might be more centralized.

[flx42]

We had this discussion 1 hour ago with @3XX0, are you spying on us? :)
Actually, cuDNN is independent from the toolkit, the reference to the toolkit in the cuDNN archive is the minimum toolkit version to use this version of cuDNN. cuDNN v2 requires 6.5+, v3 requires 7.0+ and v4 will also be 7.0+. So if we take your first example, we will need cuda:7.5-cuddn-v3, cuda:7.5-cudnn-v4.
With the second example, we will need cudnn:v3-7.5 and cudnn:v4-7.5

We can postpone this discussion to the moment when we start working on DockerHub integration. Until then, you can tag the image as you see fit.

[flx42]

cuDNN v2 is now available and I just submitted a Dockerfile based on CUDA 7.0.
Now you can simply use those two lines:

$ docker build -t cuda:7.0 github.com/NVIDIA/nvidia-docker#:ubuntu/cuda/7.0
$ docker build -t cudnn github.com/NVIDIA/nvidia-docker#:ubuntu/cuda/7.0/cudnn

And then you can simply do FROM cudnn

[ruffsl]

Awesome, thanks! I'll go test this on tensorflow.
Update: It works! I'll have to cleanup the mods to the tensorflow build though.
Haha, I may have an idea of what you guys are thinking because like you know, I've done this before. :D

[mtngld]

Thanks for adding cuDNN v2, do you plan including cuDNN v3 too or only the future v4? Caffe is not backward compatible with v2 which makes it impossible to use this docker for using Caffe.

[flx42]

I will ask around concerning v3. But don't worry, v4 will be released soon and it will work with Caffe.

[flx42]

I just added cuDNN v3 for CUDA 7.5!
You can now use the Makefile to build it:

make 7.5-cudnn3-devel 7.5-cudnn3-runtime