Caffe: Why add an 'Permute' layer before flattening the output?

[yw155]

I have two questions:

1. Why need to convert N * C * H * W to N * H * W * C, but not flattening it directly?
2. The default input image size is 300*300. If I input an image with a different size, do I need to adjust the pad value and the stride value in the convolution and pooling layer respectively?

[weiliu89]

@yw155

1) It is to make it easier to combine predictions from multiple layers.
2) You don't have to.

[yw155]

Thank you very much

[yxchng]

@weiliu89 @yw155 how does permuting make it easier to combine predictions from multiple layers?