convolutional-pose-machines-release
convolutional-pose-machines-release copied to clipboard
shihenw
why no BatchNorm layer?
Hello Wei, Great work. But why didn't you use BatchNorm layer which is said very good for speed and performence?
Hope Reply.
Hi Shih-En, since we're on the topic of training speed, may I ask here how many GPUs did you use to train your models?
We've tried adding BN layers after each CONV layers, but didn't get better convergence speed and accuracy. But there's a big speedup with VGG-pretrained model vs. training from scratch.
Given the VGG-pretrained model, I was able to get the result within 3 days with 2 Titan X's (pascal).
Thank you for info. I tried also BN for another project, and it got the same observation as yours: no speedup, no amelioration.
2016-12-30 2:01 GMT+01:00 Shih-En Wei [email protected]:
We've tried adding BN layers after each CONV layers, but didn't get better convergence speed and accuracy. But there's a big speedup with VGG-pretrained model vs. training from scratch.
Given the VGG-pretrained model, I was able to get the result within 3 days with 2 Titan X's (pascal).
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/shihenw/convolutional-pose-machines-release/issues/38#issuecomment-269717493, or mute the thread https://github.com/notifications/unsubscribe-auth/ARFboLsP8IYD_KWXFmTwB5SsNUGe-fjtks5rNFfzgaJpZM4LVS3N .