原因：
original Caffe
It looks like that Batchnorm layer and adjacent Scale layer are
integrated in single Batchnorm layer in NVIDIA caffe. However original
BVLC/caffe doesn’t integrate them. That means any networks and
pretrained models using batchnorm and scale layers with original Caffe
cannot use in NVIDIA caffe and DIGITS.
即：
包含：Batchnorm层和adjacent Scale层分开的；
NVIDIA Caffe/DIGITS包含：Batchnorm层和adjacent Scale层合并在一个BN层中；
所以在预测过程中不能使用NVIDIA Caffe，也就是GPU。
解决方法：修改deploy.prototxt
左边为修改前，右边为修改后
