These days run Vgg and googlenet really fast be abused cry, Vgg ran 2 weeks to converge to error rate 40%, then change local tyrants K40, run some test results to everyone to see, the first part share performance report, program run in Nvidia K40, video memory 12G, Memory 64G server, training and test data set built in own datasets and imagenet datasets
Training configuration: batchsize=128
Caffe's own imagenet with CuDNN model faster than googlenet with CuDNN
VGG16 layer without CUDNN slower than Caffe own imagenet with CUDNN model
VGG19 layer without CUDNN slower than Caffe own imagenet with CUDNN model
First, CAFFE comes with the configuration, use Cudnn
Forward Speed: 220ms
Backward Speed: 360ms
Second, the CAFFE comes with the configuration, does not use the CUDNN
Forward Speed: 300ms
Backward Speed: 410ms
Third, googlenet, use CUDNN
Forward Speed: 614ms
Backward Speed: 1377ms
Four, googlenet, do not use CUDNN
Forward Speed: 1145ms
Backward Speed: 2009ms
Five, VGG16 layer, use CUDNN
Forward Speed: 3101ms
Backward Speed: 8002ms
Six, VGG19 layer, using CUDNN
Forward Speed: 3972ms
Backward Speed: 8540ms
Here is a question, according to the discussion that Vgg CUDNN will be faster than the use of CUDNN, this need to verify, follow-up can update to everyone, temporary computing resources are scarce, unable to experiment
Also want to say is, vgg convergence extremely slow, do not suggest casually use Vgg do engineering, run a parameter on can, hehe, local tyrants old beauty, anticipate Vgg convergence takes 1 months.
Analysis of Googlenet,vgg operation performance of eccv2014,imagenet competition in deep learning