Modify a Lua script to report peak memory usage for each CUDA GPU or each neural net layer

Supply me with a modification of the jcjohnson/neural-style code ([url removed, login to view]) found on Github at [url removed, login to view] so that it reports peak memory usage either for each GPU involved in a multi-gpu run, or for each layer of the neural net. Also, if possible, in cases where the program aborts because it exceeded the memory capacity of one of the GPUs, report how much memory was requested and from which GPU.

If you know of a way to obtain this information through some means other than a modification of the Lua script, that would be fine, too. Note, though, that methods such as repeatedly calling the nvidia-smi utility from the command line won't do, because they could miss narrow time slices of temporarily high memory usage by one or more of the GPUs.

If you plan to report memory usage per GPU but don't own one (much less a system with multiple GPUs) I'd be happy to test any code you write for you.

Thanks!

CUDA 脚本语言

项目ID： #14968456

关于项目