How to merge specific axes without ambuigity with numpy.ndarray

Question

Basically I want to reshape tensors represented by numpy.ndarray.

For example, I want to do something like this (latex notation)

A_{i,j,k,l,m,n,p} -> A_{i,jk,lm,np}

or

A_{i,j,k,l,m,n,p} -> A_{ij,k,l,m,np}

where A is an ndarray. i,j,k,... denotes the original axes.

so the new axis 2 becomes the "flattened" version of axis 2 and 3, etc. If I simply use numpy.reshape, I don't think it knows what axes I want to merge, so it seems ambiguous and error prone.

Is there any neat way of doing this rather than creating another ndarray manually?

Answer 1

Using reshape is never ambiguous. It doesn't change the memory-layout of the data.

Indexing is always done using the strides determined by the shape. The right-most axis has stride 1, while the axes to the left have strides given by the product of the sizes to their right.

That means for you: as long as you collect neighboring axes, it will do the "right" thing.