conv2d
Tensorflow
x, y: [batch, height, width, depth]
core: [height, width, depth_in, depth_out]
iOS BNNS
x, y: [batch, depth, height, width]
core: [depth_out, depth_in, height, width]
1st full-conn after conv2d/pooling
Tensorflow
weights: [[height, width, depth].squeeze, size_y]
iOS
weights: [size_y, [depth, height, width].squeeze]
