Skip to content

Latest commit

 

History

History

log

training log description

Imagenet 1K classes

network gpus batch-size mxnet version description
resnet-18 2 x gtx 1080 256 73a0f6e learning rate decrease by 0.1 at [30, 60, 90] epoch, centos7 + openblas + cuda8.0 + cudnn v5.1
resnet-18-512batch 2 x k80 512 73a0f6e learning rate decrease by 0.1 at [60, 75, 90] epoch, centos7 + openblas + cuda7.5+cudnn v5/v5.1
resnet-34 2(3) x k80 512 73a0f6e learning rate decrease by 0.1 at [60, 75, 90] epoch, centos7 + openblas + cuda7.5+cudnn v5/v5.1, after epoch 95, using 3xk80 for trianing
resnet-50 3 x k80 256 73a0f6e learning rate decrease by 0.1 at [30, 60, 90] epoch, centos7 + openblas + cuda7.5+cudnn v5/v5.1
resnet-101 8 x m40 480 73a0f6e learning rate decrease by 0.1 at [60, 90, 120], and also disabled the recommend aug since epoch 120
resnet-152 8 x m40 256 73a0f6e learning rate decrease by 0.1 at [30, 60, 90] epoch, but disable the recommend aug since epoch 101
resnet-200 8 x m40 256 73a0f6e others are same as above, but disable the recommend aug since epoch 95, after 118 epoch using 3 x K80 and bn-mom=0.997, all training using --memonger

Cifar10

network gpus batch-size mxnet version description
resnet-164 2 gtx 1080 128 73a0f6e learning rate decrease by 0.1 at [220, 260, 280] epoch, cuda8.0 + cudnn v5.1