Pillow 8.3 and NumPy

[doublex]

Throws exception with Pillow 8.3: TypeError: __array__() takes 1 positional argument but 2 were given

with PIL.Image.open(filepath) as img:
    numpy.array( img, dtype=numpy.float32 )

[radarhere]

What version of Python are you using? What version of numpy?

[vgrabovets]

Same problem here.
Python 3.7.4
numpy 1.21.0

workaround:

with PIL.Image.open(filepath) as img:
    numpy.array(img).astype(np.float32)

[t-vi]

Works with just about anything. The problem is that pillow newly implemented __array__, but only without arguments. Numpy requires __array__ to take an optional dtype argument. This is used e.g. by numpy.array when the user passes a dtype argument. This should be with just about any numpy/python version.

[doublex]

fails with python 3.8 + 3.9, numpy 1.20 + 1.21

[t-vi]

I'll send a patch.

[carmocca]

We are seeing errors in 3.{6,7,8,9}

[ericpre]

Same issue in https://github.com/hyperspy/hyperspy/runs/2961516889 for all python version >=3.6 using the imageio library:

                # Apply palette
>               frame_paletted = np.array(im, np.uint8)
E               TypeError: __array__() takes 1 positional argument but 2 were given

/opt/hostedtoolcache/Python/3.9.5/x64/lib/python3.9/site-packages/imageio/plugins/pillow.py:745: TypeError

The python 3.6 build is with numpy 1.17.1

[cgebbe]

same here: windows, python=3.6.8, numpy=1.19.5, Pillow=8.3.0

[seberg]

Just to note, there is a small chance that some details changed in NumPy 1.19.x, but I doubt it matters. NumPy itself actually does not care whether you do return the correct dtype here, it is more of a hint at the moment. But, it is probably good habit to just return the requested dtype anyway.

[IceTDrinker]

Any chance of a hotfix release ?

Edit : Every torchvision user using the ToTensor transform (i.e. all the vision people) will encounter the issue (and their build systems)

[janosh]

Definitely needs a hot fix. This is all over the place.

[hugovk]

There'll be an 8.3.1 soon, keep an eye on #5578. Thank you for your patience!

[radarhere]

Pillow 8.3.1 has now been released with a fix for this.