search

forresti / caffe

Caffe
http://caffe.berkeleyvision.org/
Other
5 stars 2 forks source link

more graceful degradation for large images #8

Open sergeyk opened 10 years ago

sergeyk commented 10 years ago

The current behavior is: "had to remove first 5 scales, because they didn't fit in Patchwork plane."

Is there a plan to improve this situation? Maybe process large scales in their own planes. Maybe this is already done... Or dynamically switch to CPU mode just for that image?

moskewcz commented 10 years ago

large scales will already go in their own plane. the dropped scales are too large to fit in a plane by themselves. the dynamic CPU/GPU choice idea is interesting, and a simple concept, but sound like it might be complex / hard to actually implement on top of the existing codebase. it might be worth investigating what would be required: i.e. can it be done as a workaround externally already? if that works, it would be a good indicator that the idea is sound and would illustrate what might be needed to make it work internally.

beyond that, for this issue, we have some ideas, but no firm plans/solutions. it's related to the issues of memory-blowup with smaller strides. and of course this same issue is worse if we want to use even-bigger-than-2 scales.

running on a GPU with 'lots' of memory allows you to use a larger plane size (i.e. 2000x2000), which is probably the current best workaround (i.e what forrest does in practice). using CPU mode is another; on smaller GPUs (with less memory), the perf. penalty might not be too horrible anyway.

longer term, speeding up CPU mode and/or reducing GPU memory usage (in particular: removing the ic2col buffers) are the two obvious 'big ticket' ideas for better solutions. but, as usual, it's hard to prioritize these things. as a side note, we would probably want to merge with the main master branch before doing some of these things (in case they are done or the code in caffe has changed).

On Fri, Feb 14, 2014 at 11:37 AM, Sergey Karayev notifications@github.comwrote:

The current behavior is: "had to remove first 5 scales, because they didn't fit in Patchwork plane."

Is there a plan to improve this situation? Maybe process large scales in their own planes. Maybe this is already done... Or dynamically switch to CPU mode just for that image?

Reply to this email directly or view it on GitHubhttps://github.com/forresti/caffe/issues/8 .

sergeyk commented 10 years ago

Speaking of merging with master, here's one PR that's about to get merged into BVLC/master that you guys might like: https://github.com/BVLC/caffe/pull/99

On Fri, Feb 14, 2014 at 11:54 AM, moskewcz notifications@github.com wrote:

large scales will already go in their own plane. the dropped scales are too large to fit in a plane by themselves. the dynamic CPU/GPU choice idea is interesting, and a simple concept, but sound like it might be complex / hard to actually implement on top of the existing codebase. it might be worth investigating what would be required: i.e. can it be done as a workaround externally already? if that works, it would be a good indicator that the idea is sound and would illustrate what might be needed to make it work internally.

beyond that, for this issue, we have some ideas, but no firm plans/solutions. it's related to the issues of memory-blowup with smaller strides. and of course this same issue is worse if we want to use even-bigger-than-2 scales.

running on a GPU with 'lots' of memory allows you to use a larger plane size (i.e. 2000x2000), which is probably the current best workaround (i.e what forrest does in practice). using CPU mode is another; on smaller GPUs (with less memory), the perf. penalty might not be too horrible anyway.

longer term, speeding up CPU mode and/or reducing GPU memory usage (in particular: removing the ic2col buffers) are the two obvious 'big ticket' ideas for better solutions. but, as usual, it's hard to prioritize these things. as a side note, we would probably want to merge with the main master branch before doing some of these things (in case they are done or the code in caffe has changed).

On Fri, Feb 14, 2014 at 11:37 AM, Sergey Karayev notifications@github.comwrote:

The current behavior is: "had to remove first 5 scales, because they didn't fit in Patchwork plane."

Is there a plan to improve this situation? Maybe process large scales in their own planes. Maybe this is already done... Or dynamically switch to CPU mode just for that image?

Reply to this email directly or view it on GitHub< https://github.com/forresti/caffe/issues/8> .

— Reply to this email directly or view it on GitHubhttps://github.com/forresti/caffe/issues/8#issuecomment-35118548 .