v0.12 TypeError

[jbogaardt]

Describe the bug
The following assertion worked until v0.12. Now it returns a TypeError.

To Reproduce

import sparse
import numpy as np


a = np.random.rand(1000,1000)
a[a<.9] = np.nan
a = sparse.COO(a, fill_value=np.nan)

a.fill_value = 0.0
a = sparse.COO(a)

assert a == a

[hameerabbasi]

a.fill_value = 0.0

This line right here is the issue. This attribute should be read-only and private. However, if you're sure this is what you want, you can potentially use a.fill_value = np.zeros((), dtype=a.dtype)[()] instead, the difference being it gives you a NumPy scalar instead of a Python one.

[hameerabbasi]

Usual caveats about it being unsupported apply.

[jbogaardt]

Thank you @hameerabbasi . I definitely don't want to adopt unsupported conventions. So the question becomes - Is there any "supported" method for altering (even if copying the array) the fill_value of a sparse array?

[hameerabbasi]

The best way would be to construct a new COO with the same coords and data, but a new fill-value. AFAIK this won't incur copies.

[jbogaardt]

Awesome, that works!

import sparse
import numpy as np

a = np.random.rand(1000,1000)
a[a<.9] = np.nan
a = sparse.COO(a, fill_value=np.nan)
a = sparse.COO(coords=a.coords, data=a.data, fill_value = 0.0)

assert a == a

[hameerabbasi]

Another bit of unsupported behaviour there is a == a. You should really write assert (a == a).all().

[hameerabbasi]

Is it okay to close this, @jbogaardt?

[jbogaardt]

Yes, thank you.