Though the script get the correct output, I'm quite confused with the
StorageOrder staff, say I had to converted the input with
xyz_to_zyx to get the correct output.
So, what's the diffetence between
StorageOrder.orderZYX, and how to feed the input and weights to the network correctly?
@oraoto The X and Y dimensions are the height and width and the Z dimension is the channels. YXZ has the channels first in rows, then in columns (i.e. BGR BGR BGR…). With ZYX, the tensor is split into planes. The first WxH channels come from the first channel, then the second channel, etc. So the image would be split into planes of blue, green and red. For the NCSDK, all layers are implemented as YXZ.