generate_from_frequencies may use negative font size in some cases

[BLKSerene]

Description

I use generate_from_frequencies to generate word clouds with non-frequency data (test statistics, bayes factors, etc.) which could be negative, and in most cases it works. But in some cases, it raises an error.

It seems that the problem is that generate_from_frequencies would try to use a negative font size in some cases.

Steps/Code to Reproduce

Example:

import wordcloud
word_cloud = wordcloud.WordCloud()

word_cloud.generate_from_frequencies({'a': 3, 'b': -5, 'c': -1}) # OK!
word_cloud.generate_from_frequencies({'a': 3, 'b': -5}) # Error!

Expected Results

No error.

Actual Results

Traceback (most recent call last):
File "<pyshell#25>", line 1, in
word_cloud.generate_from_frequencies({'a': 3, 'b': -5})
File "D:\Python\lib\site-packages\wordcloud\wordcloud.py", line 453, in generate_from_frequencies
self.generate_from_frequencies(dict(frequencies[:2]),
File "D:\Python\lib\site-packages\wordcloud\wordcloud.py", line 509, in generate_from_frequencies
# find possible places using integral image:
File "D:\Python\lib\site-packages\PIL\ImageDraw.py", line 607, in textsize
return font.getsize(
File "D:\Python\lib\site-packages\PIL\ImageFont.py", line 859, in getsize
w, h = self.font.getsize(text)
File "D:\Python\lib\site-packages\PIL\ImageFont.py", line 483, in getsize
size, offset = self.font.getsize(text, "L", direction, features, language)
OSError: invalid argument

Versions

Windows-10-10.0.19044-SP0
Python 3.8.10 (tags/v3.8.10:3d8993a, May 3 2021, 11:48:03) [MSC v.1928 64 bit (AMD64)]
NumPy 1.22.3
matplotlib 3.5.1
wordcoud 1.8.2.2

[amueller]

Indeed, not planned, but I'm happy to keep it open in case someone wants to plan it - it's unlikely I'll look at it, but I'd be happy to accept a fix.

[BLKSerene]

@amueller I would be happy to open a PR for this. But should I modify the calculation logic of font size or just modify users' raw data? An easy fix would be adding the absolute value of the smallest negative values in data (plus a tiny value 1e-15 to avoid too many zeros) to all data points in case of negative numbers, e.g.
{'a': -2, 'b': -2, 'c': -1, 'd': 5}
would be changed to
{'a': 1e-15, 'b': 1e-15, 'c': 1 + 1e-15, 'd': 7 + 1e-15}
then everything works just as normal frequency data would do.

Another issue is that frequency data should generally contain only non-negative integers, so should I just patch the original generate_from_frequencies function to allow negative data as well or add another new function such as generate_from_data with the patch?