I believe it is very useful to support fill with minimum value to the grid_sample function. Is it possible to support this new “padding” modes in future pytorch releases?
Because normalizing the input tensor to zero mean 1 std is very popular. In this case, the background intensity value is <0, and all the three “padding” modes (zero, border, reflect) introduce an obvious “frame” to the transformed images/volumes even if they are noise-free. This have direct negative effects of, e.g., image registration applications. If we have native support to fill a “Minimum” value, this problem can be easily solve. Thank you very much!