Where is nvcc inside nvidia/caffe?

[Mottotime]

I can not find nvcc inside the nvidia/caffe image.

Would you please add it into the image and provide some explanation of how nvidia-docker works? Because I find it little mysterious to understand how nvidia-docker talks to host GPU device.

Besides, I suggest installing caffe with source code. Because examples of caffe and cuda are valuable tools for beginners.

Thank you very much.

[moconnor725]

nvcc is the compiler. Once caffe is built you wouldn't need the compiler any more.

nvidia docker enables applications that can talk to the driver to run in a container by mapping the driver into the environment.

The nvidia/caffe container isn't really meant as a substitution for a shell environment. It's not a virtual machine. The nvidia/caffe image is available for users to build upon (like DIGITS) or for users who want to execute a caffe command in a container (like caffe train).

[flx42]

You can find more information on the wiki, especially this page:
https://github.com/NVIDIA/nvidia-docker/wiki/Why%20NVIDIA%20Docker#motivation

[flx42]

@moconnor725 is right. To elaborate a bit more: we cannot foresee all possible use cases so we decided to provide a limited number of images for Caffe and DIGITS. For CUDA we have many combinations (base OS + version + runtime/devel + cudnn version) but that's OK since that's the base image.
Caffe is useful as a base image for less people that CUDA, so we didn't want to do the same (e.g.: base OS + cudnn version + caffe version + source/no-source). So we decided to have a minimal Caffe image instead of integrating all the sources.
If your particular use case is different then you can either build on top of the existing Caffe image (and apt-get nvcc) or on top of the cudnn image and build Caffe from source. You can take a look at the Dockerfile for BVLC/caffe (not NVIDIA/caffe) for the latter:
https://github.com/BVLC/caffe/blob/master/docker/standalone/gpu/Dockerfile

[Mottotime]

Thank you very much, @moconnor725 @flx42 :)

As a user, I need a full development environment of CUDA and Caffe inside docker because I want to quickly build, test and deploy my model in a consistent environment. I think it would be convenient to begin with CUDA and Caffe examples to test the hardware/software environment.

I will take @flx42 's suggestions and build a container based on these Dockerfiles.

Again, thank you for your kind explanation and advices.