Bug summary

When creating a histogram out of float16 data, the bins are also calculated in float16. The lower precision can cause two errors:

1. Gaps between certain bins.
2. Two neighboring bins overlap each other (only visible when alpha < 1)

Code for reproduction

importnumpyasnpimportmatplotlib.pyplotaspltvalues=np.clip(np.random.normal(0.5,0.3,size=1000),0,1).astype(np.float16)plt.hist(values,bins=100,alpha=0.5)plt.show()

Actual outcome

Expected outcome

Created byplt.hist(values.astype(np.float32), bins=100, alpha=0.5) plt.show()

Additional information

Possible solution
Calculate the bins in float32:

• Determine minimal and maximal value in float16.
• Convert min and max to float32.
• Calculate the bin edges.

Theoretical possible, but unwanted solution
Convert data into float32 before calculating the histogram. This behavior does not make a lot of sense, as float16 is mostly used because of memory limitations (arrays with billions of values can easily take several gigabytes).

Operating system

Windows 10

Matplotlib Version

3.4.3

Matplotlib Backend

TkAgg

Python version

3.7.1

Jupyter version

No response

Installation

pip