Inverse does not account for padding

[fversaci]

Inverse does not account for padding

[fversaci]

In the 2D-Haar case, padding is added only when the number of rows or columns is odd and to obtain the original vector you just need to discard the last row/column.

However this information is not present in the transformed tensor, e.g., a tensor of shape [2,3,4] might derive from [4,6,1] (no padding) or from [3,5,1] (padding both in x and y), and therefore you cannot automatically recover the original tensor (unless, of course, you add the original tensor shape as an input to the process).