1""" 2Functions for performing contractions with array elements which are objects. 3""" 4 5import numpy as np 6import functools 7import operator 8 9 10def object_einsum(eq, *arrays): 11 """A ``einsum`` implementation for ``numpy`` arrays with object dtype. 12 The loop is performed in python, meaning the objects themselves need 13 only to implement ``__mul__`` and ``__add__`` for the contraction to be 14 computed. This may be useful when, for example, computing expressions of 15 tensors with symbolic elements, but note it will be very slow when compared 16 to ``numpy.einsum`` and numeric data types! 17 18 Parameters 19 ---------- 20 eq : str 21 The contraction string, should specify output. 22 arrays : sequence of arrays 23 These can be any indexable arrays as long as addition and 24 multiplication is defined on the elements. 25 26 Returns 27 ------- 28 out : numpy.ndarray 29 The output tensor, with ``dtype=object``. 30 """ 31 32 # when called by ``opt_einsum`` we will always be given a full eq 33 lhs, output = eq.split('->') 34 inputs = lhs.split(',') 35 36 sizes = {} 37 for term, array in zip(inputs, arrays): 38 for k, d in zip(term, array.shape): 39 sizes[k] = d 40 41 out_size = tuple(sizes[k] for k in output) 42 out = np.empty(out_size, dtype=object) 43 44 inner = tuple(k for k in sizes if k not in output) 45 inner_size = tuple(sizes[k] for k in inner) 46 47 for coo_o in np.ndindex(*out_size): 48 49 coord = dict(zip(output, coo_o)) 50 51 def gen_inner_sum(): 52 for coo_i in np.ndindex(*inner_size): 53 coord.update(dict(zip(inner, coo_i))) 54 locs = (tuple(coord[k] for k in term) for term in inputs) 55 elements = (array[loc] for array, loc in zip(arrays, locs)) 56 yield functools.reduce(operator.mul, elements) 57 58 out[coo_o] = functools.reduce(operator.add, gen_inner_sum()) 59 60 return out 61