skimage.transform.rescale fails for scale < 0.2

[pyscorcher]

Description

The function skimage.transform.rescale throws an error for for scale < 0.2, which is quite unexpected if you're just trying to use the rescale function. I get following error messages for these small scale factors:

SOME_PATH\skimage\transform\_warps.py:115: RuntimeWarning: divide by zero encountered in true_divide
  factors = (np.asarray(input_shape, dtype=float) /
Traceback (most recent call last):
  File "PATH_TO\my_script.py", line 5, in <module>
    img = transform.rescale(img, scale=0.1)  # error
  File "SOME_PATH\skimage\transform\_warps.py", line 290, in rescale
    return resize(image, output_shape, order=order, mode=mode, cval=cval,
  File "SOME_PATH\skimage\transform\_warps.py", line 147, in resize
    image = ndi.gaussian_filter(image, anti_aliasing_sigma,
  File "SOME_PATH\scipy\ndimage\filters.py", line 341, in gaussian_filter
    gaussian_filter1d(input, sigma, axis, order, output,
  File "SOME_PATH\scipy\ndimage\filters.py", line 257, in gaussian_filter1d
    lw = int(truncate * sd + 0.5)
OverflowError: cannot convert float infinity to integer
Process finished with exit code 1

Way to reproduce

from skimage import transform
import numpy as np
img = np.zeros((30, 40, 3))
img = transform.rescale(img, scale=0.2)  # ok
img = transform.rescale(img, scale=0.1)  # error

Version information

# Paste the output of the following python commands
from __future__ import print_function
import sys; print(sys.version)
import platform; print(platform.platform())
import skimage; print("scikit-image version: {}".format(skimage.__version__))
import numpy; print("numpy version: {}".format(numpy.__version__))

3.7.3 (default, Mar 27 2019, 17:13:21) [MSC v.1915 64 bit (AMD64)]
Windows-10-10.0.19041-SP0
scikit-image version: 0.14.2
numpy version: 1.16.2

[rfezzani]

Thank you @pyscorcher for the report, I can reproduce the bug.

In the example that you provided, the input image shape is (30, 40, 3) and when you apply a scale factor of 0.1, the output image shape is obtained as round(0.1 * [30, 40, 3]) = round([3, 4, 0.3]), which is [3, 4, 0]. This last 0 is the root of the problem.

I assume that this behavior is expected, but the raised error is extremely confusing 😕.

[hmaarrfk]

Thanks for getting to the bottom of that @rfezzani

@pyscorcher to help you with your work, there is a multichannel parameter that also exists in version 0.14.2
https://scikit-image.org/docs/0.14.x/api/skimage.transform.html#rescale

I think that should help you. Let us know!

[pyscorcher]

@rfezzani Thanks for explaining and for the PR!
@hmaarrfk That is very good to know, thank you very much!

[ntjess]

Is this issue fixed in 0.18.1?

I face something similar when (x,y) resized dimensions are 0:

from skimage import transform as trans
import numpy as np

x = np.zeros((1,1), bool)
y = trans.resize(x, (0,0))

This raises the unhelpful error message numpy.linalg.LinAlgError: SVD did not converge