Running on CPUs about sparseconvnet-archived HOT 5 CLOSED

I know this does not answer your question directly, but FYI, I've been running SparseConvNet on Amazon g2 instances (which do have GPUs) with a single minor modification to the code (the detection of the GPU does not work). The performance is OK, even from EBS (i.e. networked) drives rather than from the dedicated 80GB SSD dedicated to the instance. So that's a possible cloud-based solution.

from sparseconvnet-archived.

Samuel: Hello. As Jean-Olivier say, you can run SparseConvNet on g2 instances.
I am planning on producing a CPU version. My main motivation for doing that would be to make it easier to distribute trained networks, rather than for training. However, in some cases CPUs might be fast enough for training too.
If you email me ([email protected]) a few details about what kind of data you would be looking at, I will see if I can help.
Regards
Ben

from sparseconvnet-archived.

Great, will be in touch.

And yeah @jiho, g2 is certainly an option I've considered, and there's nothing wrong with it, just that I'd like to be able to run this network on other cloud providers (not all of which support GPUs).

from sparseconvnet-archived.

I have hacked SparseConvNet to run on the CPU:
https://github.com/btgraham/SparseConvNet-CPU
Unsurprisingly, the GPU version is quite a lot faster!

from sparseconvnet-archived.

Wow, excellent, thanks!

from sparseconvnet-archived.