named tensor axes

import numpy as np

def prepare_order(current, target):
    """
    precalc the reordering of axes
    this need only be done once
    note; we need to allow for ... operator still.
    left of ... in target have plus or minus sign respectively
    """
    def calc(t):
        r = tuple(i for i,v in enumerate(current) if v==t) #all locactions of t in current
        if len(r)==0:
            return None
        if len(r)==1:       #most common case; unpack for readability and performance
            return r[0]
        return r
    return map(calc, target)

def reform(arr, current, target):
    """
    use as_strided to map current axes layout to target axes layout
    prepare an order object
    then apply it
    """
    order = prepare_order(current, target)

    def calc(o):
        if o is None: return (1,0)                                  #broadcasting axis
        if isinstance(o, int): return arr.shape[o],arr.strides[o]   #transposed axes
        return arr.shape[o[0]],sum(arr.strides[i] for i in o)       #diag axis
    layout = map(calc, order)

    shape, strides = zip(*layout)
    return np.lib.index_tricks.as_strided(arr, shape, strides)

Q = np.arange((3*3*3)).reshape(3,3,3)

#reorder axes
print reform(Q, ['a','b','c'], ['b','c','a'])
#view higher order diagonal
print reform(Q, ['a','a','a'], ['a'])
#broadcast left and right
Q = np.arange(3)
print reform(Q, ['a'], ['a','b'])
print reform(Q, ['a'], ['b','a'])