Pylint: False positive: "Does not support item assignment" error with np.zeros_like

[Max16hr]

Hi there!
Pylint gives a false-positive error while using numpy.zeros_like().
This numpy method will return a ndarray object. But while trying to address this object by indexing, pylint will raise an error although python works fine.

Code to reproduce:.

import numpy as np

a = np.array([[1, 2, 3], [4, 5, 6]])
print(a)
b = np.zeros_like(a)
print(b)

index = (0, 1)

a[index] = 9
print(a)

b[index] = 9
print(b)

Current behavior

Pylint gives

'b' does not support item assignment

in line 13. But Python runs the program without any problem.

Same issue with np.full_like(), np.empty_like(), np.ones_like().

Expected behavior

No pylint error.
Workaround: In line 5, change to b = np.array(np.zeros_like(a)).

pylint --version output

pylint 2.2.2
astroid 2.1.0
Python 3.7.1

[PCManticore]

Thanks for the report!

[hippo91]

@Max16hr the bug arises because of a lack of some dunder methods (such as __getitem__, __setitem__ ...) in the ndarray class in astroid numpy brain and because of a wrong inference of zeros_like, full_like etc methods.
I'have got a fix but it still needs more work to test correctly all the ndarray methods and all the functions that return a ndarray.
I'm working on it.

[VasLem]

This bug has not been fixed, I can reproduce the same behavior.

[hippo91]

@VasLem thanks for your remark.
Can you please send a precise example of your code please?

[olivbau]

Here is an example

arr1 = np.array([1,2,3])
arr2 = np.ones_like(arr1, dtype=np.uint8)
arr2[arr1 >= 2] = 255

print(arr2)
# [  1 255 255]

Line 3 : 'arr2' does not support item assignmentpylint(unsupported-assignment-operation)
Pylint : 2.3.1

[DungMinhDao]

There are still bugs with pylint. Here is my example:

import numpy as np 

def check_grad(fn, gr, X):
    X_flat    = X.reshape(-1)           # convert X to 1d array -> l for loop needed
    shape_X   = X.shape                 # original shape of X
    num_grad  = np.zeros_like(X)        # numerical grad, shape = shape of X
    grad_flat = np.zeros_like(X_flat)   # 1d version of grad
    eps       = 1e-6                    # a small number, 1e-10 --> 1e-6 is usually good
    numElems  = X_flat.shape[0]         # number of elements in X
    # calculate numerical gradient
    for i in range(numElems):           # iterate over all elements of X
        Xp_flat      = X_flat.copy() 
        Xn_flat      = X_flat.copy() 
        Xp_flat[i]  += eps
        Xn_flat[i]  -= eps
        Xp           = Xp_flat.reshape(shape_X)
        Xn           = Xn_flat.reshape(shape_X)
        grad_flat[i] = (fn(Xp) - fn(Xn))/(2*eps)
    num_grad = grad_flat.reshape(shape_X)
    diff = np.linalg.norm(num_grad - gr(X))
    print('Difference between 2 methods should be small: ', diff)

m, n = 10, 20
A = np.random.rand(m, n)
X = np.random.rand(n, m)

def fn1(X):
    return np.trace(A.dot(X))

def gr1(X):
    return A.T 

check_grad(fn1, gr1, X)

In line 18: 'grad_flat' does not support item assignment
In line 20: Instances of 'tuple' has no 'reshape' member
However, everything turns out to be just fine with Python interpreter
Pylint: 2.3.1

[hippo91]

@olivbau
I can confirm that using pylint (2.3.1) and astroid (2.2.5) the message does not support item assignment is emitted.
However when using the master branch of astroid it is no more emitted.
I assume that you won't be facing this message anymore as soon as a new release of astroid will be published.

[hippo91]

@DungMinhDao
It appears that with the astroid's master branch those messages are not yet emitted.
So i assume you won't be facing this problem anymore as soon as a new astroid version will be published.

[RichardPflaum]

Is there any workaround? I could get rid of similar errors by adding numpy.ndarray to the --ignored-classes. Pylint seems to think the return of zeros_like() is a tuple though..

[hippo91]

@RichardHuenermann you are right. With the current version of astroid the zeros_like function return is inferred as a tuple. It is no more the case with the master branch of astroid thanks to PyCQA/astroid#664. With the next version of astroid your problem should disappear.