farmingyard
About speed
Hi, thanks for your work. when I test and train the net, it seems more slowly than alexNet or caffeNet under the same configuration, is this normal? What was your result? thanks!
@jianan86 I haven't trained alexnet or caffenet, so I can't compare their speed with shufflenet, but i can give you the shufflenet's training speed on my card: GTX1080, shufflenet_1x_g3_deploy.prototxt, batchsize=64, 40 iters, cost time is about 16 seconds.
@farmingyard Here is my test for shufflenet with caffe_time (TITAN xp) Average Forward pass: 9.85734 ms. Average Backward pass: 9.51034 ms. And the test for mobilenet you recommended (from https://github.com/farmingyard/caffe-mobilenet ) Average Forward pass: 5.08076 ms. Average Backward pass: 7.28298 ms. But as the paper says that Shufflenet performs better. Did you test it before? or any suggestion. Thank you a lot.
@farmingyard @MrAprils I met the same problem as MrAprils Here is my test for shufflenet with caffe_time(GTX1080) Average Forward pass: 8.30279 ms. Average Backward pass: 9.11459 ms. Average Forward-Backward: 17.7004 ms. mobilenet: Average Forward pass: 6.63855 ms. Average Backward pass: 15.8861 ms. Average Forward-Backward: 22.6294 ms. mobilenet seems to perform better,i don't know why? thank you!
@farmingyard @jianan86 hello ,have you figure out why shuffle net test slower than mobile_net_1by2 and alexnet ? I use Tesla M40, and batch size 1, the average forward time is shuffle net(13ms), mobile_net_1by2 (7ms), alexnet(3.7ms), any suggestion to improve the implementation speed ?
@jianan86 @farmingyard @MrAprils @lvboodvl @ @ @xiaomr Did you log the time of the following two model. One is the time of original framework,like VGG(caffe time and caffe train/test time).The another one is the the shuffleNet-VGG's framework,I mean that you only change the convolutions operation's format from the original VGG to the ShuffleNet-convolutions operation's format. Any information is welcome
