How to recreate performance results

237 views
Skip to first unread message

robert.jo...@gmail.com

unread,
Nov 17, 2014, 9:38:11 AM11/17/14
to caffe...@googlegroups.com
I'm trying to recreate the K40 performance results reported in http://caffe.berkeleyvision.org/performance_hardware.html
Best settings with ECC off and maximum clock speed in standard Caffe:
Training is 26.5 secs / 20 iterations (5,120 images)
Testing is 100 secs / validation set (50,000 images)

What commands are used for the training phase, and the testing phase ?

I'm assuming the training phase was run with the script
    ./examples/cifar10/train_quick.sh

How was the testing phase run ?
Was this run with "caffe test", or "caffe time" ?

Bob Francis

Evan Shelhamer

unread,
Nov 17, 2014, 6:03:03 PM11/17/14
to robert.jo...@gmail.com, caffe...@googlegroups.com
The performance numbers are for CaffeNet, the ILSVRC classifier. The training and testing times are measured by inspecting the log of `caffe train` so that the training and test nets are both executed as they are in real world usage.

The `caffe time` result should in any case be quite close.

Evan Shelhamer

--
You received this message because you are subscribed to the Google Groups "Caffe Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users...@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/6d8bed41-3fc4-493c-8ee1-3ff728c0a208%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Andriy Lysak

unread,
Nov 17, 2014, 6:39:57 PM11/17/14
to caffe...@googlegroups.com
I actually ran my own data set on the ImageNet with an EVGA GTX680 (which is the same as GTX 770, they were re-branded ) and got the same performance as the results on the hardware page for the GTX 770.

So you might want to look at this:

 http://caffe.berkeleyvision.org/gathered/examples/imagenet.html
Reply all
Reply to author
Forward
0 new messages