The following variables seem to be similar but they aren't and I don't understand why:
import ujson import numpy as np arr = np.array([1, 2, 3, 4]) arr_1 = arr.tolist() arr_2 = list(arr) arr_1 == arr_2 # True ujson.dumps({'arr': arr_1}) # it works ujson.dumps({'arr': arr_2}) # it does not work (OverflowError: Maximum recursion level reached) I am using Python-3.5.6, ujson-1.35 and numpy-1.16.4.
Thank you a lot for your help!!
numpy has its own numeric data types for different levels of precision.
These are created in ways that allow easy comparison with regular Python integers.
np.int32(3) == 3 # True [np.int32(1), 4] == [1, np.int32(4)] # True (Lists are equal in Python if all elements at the same index are equal)
That is why your arr_1 == arr_2.
They cannot readily be serialized to json, but the tolist method converts them to regular Python numeric types which do allow serialization.
The key differences between tolist() and list() are
The tolist() method converts a multidimensional array into a nested list whereas list() converts it to a list of arrays. For example,
import numpy as np # create a 2-D array array1 = np.array([[1, 2], [3, 4]]) # convert a 2-D array to nested list list1 = array1.tolist() # convert a 2-D array to a list of arrays list2 = list(array1) print('Using array.tolist(): ', list1) print('Using list(array): ', list2) 
tolistconverts it all the way.list()just iterates on the first dimension. In this casearr_2is a list ofnp.intobjects, where asarr_1is a list of python integers.ujsonhas a recursion error. Perhaps it tries to represent theint32number as a dictionary-like object and fails. If you use the standardjsonlibrary you will see that the problem is that it does not know how to serialize numbers stored in thenumpyint32type. See this question