TypeError: Cannot cast ufunc add output from dtype('float64') to dtype('uint8') with casting rule 'same_kind'

[dessskris]

I'm using Gizeh library and I get the above error upon installing the latest version of Numpy. There was no error with Numpy version 1.08.

File "animation/target_animation.py", line 161, in draw
    fill = gizeh.ImagePattern(self.bg.data, self.bg.pos, filter='best')
  File "build/bdist.linux-x86_64/egg/gizeh/gizeh.py", line 295, in __init__
  File "build/bdist.linux-x86_64/egg/gizeh/gizeh.py", line 50, in from_image
TypeError: Cannot cast ufunc add output from dtype('float64') to dtype('uint8') with casting rule 'same_kind'

Is there a workaround or would you fix this issue please?

[charris]

You need to make the cast explicit using the casting argument in the np.add ufunc. The behavior Gizeh depends on has been deprecated since 1.7.

[charris]

That is casting='unsafe'.

[vishalgolcha]

where should i change this @charris ?

[seberg]

Typically this is code such as a += b and you need to make it np.add(a, b, out=a, casting="unsafe") (if you are sure you want the unsafe cast behaviour. Or, unless b is huge and it is safe, maybe just cast b first.

[vishalgolcha]

I get that I was talking specifically for the changes to be made to gizeh/moviepy .

[mantielero]

In my case, I had a similar issue. It was solved as suggested by seberg. I replaced line 40 in gizeh.py:

arr += image.flatten()

with line:

arr = np.add(arr, image.flatten(), out=arr, casting="unsafe")

I hope this helps somebody else.

[FilipSavic95]

I had this issue when subtracting like this:
yn -= self.y_mean

My error went away when I did the subtraction via numpy.subtract, like so:
yn = np.subtract(yn, self.y_mean, out=yn, casting="unsafe")

Thanks a lot @seberg 👍

[eric-wieser]

@FilipSavic95: Could also do yn -= self.y_mean.astype(yn.dtype)

[FilipSavic95]

@eric-wieser Yeah, I've just tried your solution and it works. Thanks for posting it. 👍

[seberg]

The question is which precision you want to use for the operation itself. the unsafe casting will do the operation in the larger (rhs) precision (or the combined safe dtype) the other option will do the cast and thus the operation in the lower precision.

[WarrenWeckesser]

This is a downstream issue (see Zulko/gizeh#25), so closing.

[ocaleanam]

Hello, looking for this specific error with search engines usually leads to the solutions above, either on this issue here or on StackExchange.

However I've encountered it differently using np.insert

I had a massive array containing huge integers, as well as an array of indices and another of values to insert in the first.
That's why all three arrays had dtype=np.uint (uint64)

If you go by :

a = np.arange(100)
b = np.array([12, 25, 3, 0, 46, 87])
c = np.array([7, 99, 97, 85, 23, 45])

np.insert(a, b, c)

...there's no problem, values of c are inserted in a at positions given by b.

However, if the indices array has dtype=np.uint :
np.insert(a, b.astype(np.uint), c)

---------------------------------------------------------------------------
UFuncTypeError                            Traceback (most recent call last)
<ipython-input-4-bc3dd4e8f3bc> in <module>
----> 1 np.insert(a, b.astype(np.uint64), c)

<__array_function__ internals> in insert(*args, **kwargs)

~/.local/lib/python3.8/site-packages/numpy/lib/function_base.py in insert(arr, obj, values, axis)
   4738     numnew = len(indices)
   4739     order = indices.argsort(kind='mergesort')   # stable sort
-> 4740     indices[order] += np.arange(numnew)
   4741 
   4742     newshape[axis] += numnew

UFuncTypeError: Cannot cast ufunc 'add' output from dtype('float64') to dtype('uint64') with casting rule 'same_kind'

Cast indices array to int64 to solve :
np.insert(a, b.astype(np.int64), c)

I didn't understand why the error was mentioning float64 whereas everything was uint64. I was trying to look for uint64 overflow or some other weird implicit conversion before I realized indices must be signed int64.

Hope it'll help if someone else gets here for the same reason.

[seberg]

What you are seeing is that uint64 + int64 returns float64 in NumPy, which is somewhat weird.

[ocaleanam]

Indeed, I was so pissed when I realized I'd lost so much time for this