deepcl_predict

[Armin234]

Hi,
I'm new on deepcl and trained the mnist as described in the doku without error, but I don't understand how use the deepcl_predict

C:\Users\Armin\Desktop\deepcl-win64-v11.3.1\dist\bin>mnist-to-pipe.exe ..\data\mnist\train-images-idx3-ubyte 10 | deepcl_predict.exe
Something went wrong: imageSize doesnt match imageSizeCheck, image not square

[hughperkins]

images need to be square. It's a design decision.

[OliverUrbann]

Nevertheless, this tool is somewhat unclear. I got it running but don't understand the output:

ManifestLoaderv1 checking format for /home/civic/dev/DeepCL-bin/manifest
matched: 1
ManifestLoaderv1 checking format for /home/civic/dev/DeepCL-bin/manifest
matched: 1
read file it=0
N is: 4030
read file it=1
doing alloc N=4030
manifest /home/civic/dev/DeepCL-bin/manifest read. N=4030 planes=1 size=16 labels? 0
statefultimer v0.7
forward try kernel 0
... not plausibly optimal, skipping
forward try kernel 1
... seems valid
ForwardAuto: kernel 1 230ms
forward try kernel 0
... not plausibly optimal, skipping
forward try kernel 1
... seems valid
ForwardAuto: kernel 1 228ms
forward try kernel 0
... not plausibly optimal, skipping
forward try kernel 1
... seems valid
ForwardAuto: kernel 1 6ms
forward try kernel 0
... not plausibly optimal, skipping
forward try kernel 1
... seems valid
ForwardAuto: kernel 1 7ms

What I would expect is just a single time for everything or per image. But does this output mean?

[hughperkins]

First few images, it is deciding which kernel to use. It tries each one, and sees which runs fastest. It does this for each layer. I found there is no easy heuristic to decide which works best for which inage size on which hardware, so just try all of them :P

…

On Sun, Apr 1, 2018, 09:33 Oliver Urbann ***@***.***> wrote: Nevertheless, this tool is somewhat unclear. I got it running but don't understand the output: ManifestLoaderv1 checking format for /home/civic/dev/DeepCL-bin/manifest matched: 1 ManifestLoaderv1 checking format for /home/civic/dev/DeepCL-bin/manifest matched: 1 read file it=0 N is: 4030 read file it=1 doing alloc N=4030 manifest /home/civic/dev/DeepCL-bin/manifest read. N=4030 planes=1 size=16 labels? 0 statefultimer v0.7 forward try kernel 0 ... not plausibly optimal, skipping forward try kernel 1 ... seems valid ForwardAuto: kernel 1 230ms forward try kernel 0 ... not plausibly optimal, skipping forward try kernel 1 ... seems valid ForwardAuto: kernel 1 228ms forward try kernel 0 ... not plausibly optimal, skipping forward try kernel 1 ... seems valid ForwardAuto: kernel 1 6ms forward try kernel 0 ... not plausibly optimal, skipping forward try kernel 1 ... seems valid ForwardAuto: kernel 1 7ms What I would expect is just a single time for everything or per image. But does this output mean? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#145 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAHiqGwcT7PddPx5i20syei6Mzdxd_49ks5tkNcigaJpZM4S6UgY> .

[OliverUrbann]

Oh thx for that quick answer. So to see how long the prediction of all 4030 takes I have just to add 230+228+6+7, right?

[hughperkins]

Well, there is a ton of other overhead around the place. What i would do is run for maybe 1-3 seconds to warm up, then run for another 3-5 seconds, and see how many images go through in these last 3-5 seconds, and divide by however long you waited. But yeah, approximately, you are right, (though you might need to take into account the batch size)

…

On Sun, Apr 1, 2018, 09:40 Oliver Urbann ***@***.***> wrote: Oh thx for that quick answer. So to see how long the prediction of all 4030 takes I have just to add 230+228+6+7, right? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#145 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAHiqN6YJntibq-gFu2LMPIBOnzHKkElks5tkNjTgaJpZM4S6UgY> .