Matplotlib interfering with NumPy (on Windows)

Question

_{The issue described below has been replicated}

---

Say you have the following

import numpy as np
import matplotlib.pyplot as plt

x = np.random.randint(5, size=(100, 12), dtype=np.int64)
# [THERE IS ACTUALLY NO NEED TO SET THE DATA TYPE
# `x = np.random.rand(100, 12)` yields the same problem]

and you want to compute x's rank.

>>> np.linalg.matrix_rank(x)
12

Everything is fine. Let's restart a new session from scratch, whose underlying code this time is

import numpy as np
import matplotlib.pyplot as plt

np.random.seed(1010)  # <-----
x               = np.random.randint(5, size=(100, 12), dtype=np.int64)
x_vals = y_vals = np.arange(0, .5, .05)

plt.plot(x_vals, y_vals, linestyle='--')

print(np.linalg.matrix_rank(x))

This prints 0 (!!). And more surprisingly, the reason behind this is the value given to linestyle (!!). I mean, having linestyle='-' (solid) turns everything back to normal.

This is clearly an undesired behavior (I literally spent hours to locate precisely)... but still:

How ?

---

This occurs under Windows 10 with Python3.7.3

numpy==1.19.2      # since 1.19.0 actually
matplotlib==3.3.3  # between 3.1.3 and 3.3.3 for what I can tell

No problem under Linux (with the same environment)

---

Other precisions

1. This does not occur for all x's shape. Hard to tell exactly. However, this is not random either, and is monotonously linked to x's (vertical and/or horizontal) shape.
2. The issue is transposition-invariant.
3. x is exactly the same when compared to itself before plt.plot is called (compared using joblib.hash)

---

---

This question is more about leaving a trace than getting an answer. This is so weird that I had to write it somewhere. I've changed my linestyle...

The title of the question is sufficiently unequivocal to drive people with the same problem as mine here.

---

Another screen capture:

The GIF's code follows.

import numpy as np
import matplotlib.pyplot as plt
import joblib as jl

linestyles = [
    'solid', '-',
    'dotted',  # '.', => ValueError: '.' 
    'dashed', '--',
    'dashdot', '-.',
    ':', '', ' '    
]

for ls in linestyles:
    print(26*'*', f"linestyle='{ls}'")
    
    np.random.seed(1010)
    x  = np.random.rand(9, 5)
    h0 = jl.hash(x)
    
    x_vals = y_vals = np.arange(0, .5, .05)
    plt.plot(x_vals, y_vals, linestyle=ls)
    # plt.show()
    
    h1 = jl.hash(x)
    mr = np.linalg.matrix_rank(x)
    print(
        '\t', mr, (not mr)*'<---------------[!!!]'
    )
    print('\t', 'Has not changed:', h0 == h1)

Answer 1

This is a bug in numpy and has been corrected as of version 1.19.3:

There is a known problem with Windows 10 version=2004 and OpenBLAS svd that we are trying to debug. If you are running that Windows version you should use a NumPy version that links to the MKL library, earlier Windows versions are fine.

https://github.com/numpy/numpy/releases/tag/v1.19.2

NumPy 1.19.3 is a small maintenace release with two major improvements:

• Python 3.9 binary wheels on all supported platforms.
• OpenBLAS fixes for Windows 10 version 2004 fmod bug.

https://github.com/numpy/numpy/releases/tag/v1.19.3