On Sat, 2015-08-08 at 10:57 +0100, Ghislain Vaillant wrote:
>
> Another naming convention we use for these elf executables is to use
> lib<name>-bin.
[...]
Thank you for advice, and I prefer the name caffe-{cpu,cuda},
because user will be able to find the caffe package with
$ apt list caffe* <- won't work for libcaffe-cpu-bin
$ apt search caffe <- both name ok
> Another quick question for something I might have overlooked:
>
> Is the cuda stuff really a different backend, i.e. it provides the exact
> same functionalities as the CPU library [1] and can exist on its own?
>
> or
>
> does it enhance the CPU library with GPU-accelerated speedups for a
> restricted subset of operations [2]?
>
> In case of [1] use libcaffe-cpu / libcaffe-cuda
> In case of [2] use libcaffe / libcaffe-cuda
The caffe-cpu has the same function with caffe-cuda, functionally
they are basically the same.
The issue is, training large convolutional neural networks [1]
on large dataset [2] with CPU is *nightmare*[3], while
the CUDA's parallel computing can significantly boost up training
speed when the image datasets are large.
So, caffe-cpu is suitable for computers that (ARCH != i386 | amd64),
and is suitable for small experiments that need not too much computation.
Meanwhile, caffe-cuda makes sense when user need to deal with a big
dataset/network. Moreover, users who really need performance will
find the way to compile caffe by themselves, hence I plan to provid
the custom target in d/rules (inspired by OpenBLAS's d/rules).
> FYI, someone might have already mentioned this, but if your package
> build-depends on any non-free stuff (like the cuda stack), your source
> package will not qualify for inclusion in main but in contrib instead.
Yes, I found way to pass the nvcc issue, as the last mail I sent said.
Binary packages in current control file were already marked as contrib.
Without caffe-cuda, caffe-cpu can enter the main section. However
it's more worthy that I provide caffe-cuda for researchers having
trouble compiling caffe by themselves.
My packaging work helps myself too :-)
[1] Yes, this is what caffe users do.
[2] i.e. ImageNet (Computer vision and pattern recognition)
[3] It takes centuries to see the result. :-)
Attachment:
signature.asc
Description: This is a digitally signed message part