BUG: Rounding on float16 type not working

[Genarito]

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• (optional) I have confirmed this bug exists on the master branch of pandas.

---

Problem description

I'm trying to round a column to 4 decimal places to store them in the DB. I must use np.float16 type to work with as I need the smallest possible precision, but it's not working (see output section), if I don't cast the values and leave them as np.float64 type It'll work correctly.
I think it's a bug of Pandas as if I run:

np.array([0.37324, 0.12321, 0.23]).astype(np.float16).round(4)

It works as expected, so It doesn't seem to be a problem for Numpy.

Code Sample

import pandas as pd
import numpy as np
df = pd.DataFrame(np.array([0.37324, 0.12321, 0.23]), columns=['A'])
df.astype(np.float16).round(4)

Running:

pd.DataFrame(np.array([0.37324, 0.12321, 0.23]).astype(np.float16), columns=['A']).round({'A': 4})

Doesn't work either as it has the same output.

Current Output

          A
0  0.373291
1  0.123230
2  0.229980

Expected Output

          A
0  0.3732
1  0.1232
2  0.23

Output of pd.show_versions()

INSTALLED VERSIONS

commit : None
python : 3.7.5.final.0
python-bits : 64
OS : Linux
OS-release : 5.3.0-62-generic
machine : x86_64
processor : x86_64
byteorder : little
LC_ALL : None
LANG : es_AR.UTF-8
LOCALE : es_AR.UTF-8
pandas : 1.0.5
numpy : 1.18.1
pytz : 2019.3
dateutil : 2.8.1
pip : 18.1
setuptools : 40.8.0
Cython : None
pytest : 5.3.5
hypothesis : None
sphinx : None
blosc : None
feather : None
xlsxwriter : None
lxml.etree : None
html5lib : None
pymysql : None
psycopg2 : None
jinja2 : 2.11.1
IPython : None
pandas_datareader: None
bs4 : None
bottleneck : None
fastparquet : None
gcsfs : None
lxml.etree : None
matplotlib : None
numexpr : None
odfpy : None
openpyxl : None
pandas_gbq : None
pyarrow : None
pytables : None
pytest : 5.3.5
pyxlsb : None
s3fs : None
scipy : 1.4.1
sqlalchemy : None
tables : None
tabulate : None
xarray : None
xlrd : None
xlwt : None
xlsxwriter : None
numba : None

[jreback]

we have very little support for float16

furthermore i suspect this is really just a display issue; convert the series to numpy and see

[Genarito]

Hello!
First of all, thanks for responding!
Secondly, when it converts to a Numpy array, the rounding works, I confirmed it above in the topic description. However, it would be great to have a built-in solution to handle this kind of rounding at float16.
I'm inserting the DataFrame in the DB with PyMongo, I need all the DF structure, but every time I put the rounded data in it the values return to their original representation and are inserted with more than 4 decimals. So I'm not sure if it's just a display problem.

[erfannariman]

furthermore i suspect this is really just a display issue; convert the series to numpy and see

This is actually the case. Converting it back to numpy gives us the correct result, so the values underneath are actually rounded correct:

df = pd.DataFrame(np.array([0.37324, 0.12321, 0.23]), columns=['A'])
df.astype(np.float16).round(4).to_numpy()

array([[0.3733],
       [0.1232],
       [0.23  ]], dtype=float16)

pd.DataFrame(np.array([0.37324, 0.12321, 0.23]).astype(np.float16), columns=['A']).round({'A': 4}).to_numpy()

array([[0.3733],
       [0.1232],
       [0.23  ]], dtype=float16)

[Genarito]

Thank you! When I save in CSV format it was 4 decimals truncated as expected. But now I found a use case where it's not working:

df = pd.DataFrame(np.array([0.37324, 0.12321, 0.23]), columns=['A'])
df.astype(np.float16).round(4).to_dict('records')

Outputs [{'A': 0.373291015625}, {'A': 0.12322998046875}, {'A': 0.22998046875}]. Shouldn't the truncation be preserved during the conversion?

Also important: the same problem occurs with float32. The float64 and float128 types don't present the problem.

Please, consider re-open the issue as the examples I gave are the commonly used way to work with Pandas and PyMongo. We could iterate over all the dict, but that's not the idea.

Thanks in advance

[erfannariman]

Can confirm the problem stated above occurs on master as well.

@jreback would that be sufficient to re-open this issue?

[erfannariman]

Maybe important to note that to_numpy still gives the correct result, to_dict does not:

>>> df.astype(np.float16).round(4).to_numpy()
array([[0.3733],
       [0.1232],
       [0.23  ]], dtype=float16)

[Genarito]

@jreback Is there any update on this? If you like i would try making a PR when I have some free time.

[jreback]

you can make a PR but really sure what you are attempting to solve

[Genarito]

Ok, I'll try in a few weeks

[mickaelandrieu]

Hi,

The issue is still present on the latest release of pandas with float32.

Regards,

[jreback]

you would have to open a new issue with a reproducible example

[mickaelandrieu]

Hi Jeff,

I don't know ! This issue was reproducible, clear and ... closed with no action.

The best I can do here is to inform you, then cast my series to float64 and move on.

No hard feelings here, I'm glad this bug was reported because I know why it's not working as I expected.

Regards,

[jreback]

this is about float16

[Genarito]

Now it's about float16 and float32 :'D

[ghost]

I found the root cause: #50125 (comment)