How to avoid copying data to input_blobs()?

[Hongxin Liu]

Hi everybody, 

    There is a thing that really makes me confused in classification. 

    Usually, we wrap input layer and copy data to net->input_blobs()[0]->mutable_cpu_data(), then use net->Forward() to pass it and get the output. 

    Is this necessary? I think copying data takes so much time especially in application like object detection(there are hundreds of thousands sub-windows to be evaluated). 

    Is there any method to avoid copying data?

Hongxin