resample becomes non-deterministic, depending on DateTimeIndex values

[haeusser]

Minimal Example

import datetime as dt
import numpy as np
import pandas as pd


def np_to_df(data, start_time):
    index = pd.DatetimeIndex(
        [start_time + dt.timedelta(milliseconds=t) for t in range(len(data))])
    df = pd.DataFrame(data, index=index)
    return df


# generate sample data
data = np.sin(np.arange(1000) / 30)

# create DataFrames with DateTimeIndices
df_1 = np_to_df(data, dt.datetime(2019, 9, 30, 9, 41))
df_2 = np_to_df(data, dt.datetime(2019, 9, 30, 9, 42))

# print difference before resampling
print("error_1-2:", np.mean(np.abs(df_1.values - df_2.values)))

# resample
df_1 = df_1.resample("19ms").mean()
df_2 = df_2.resample("19ms").mean()

# print difference after resampling
print("error_1-2:", np.mean(np.abs(df_1.values - df_2.values)))

Output:

error_1-2: 0.0
error_1-2: 0.04119868246404099

Problem description

When you give the exact same data to the resample function, it becomes non-deterministic if the DateTimeIndex has differing values - even though the frequency is the same.

Expected Output

The values of the two DataFrames should be exactly the same.

Output of pd.show_versions()

• commit : None
• python : 3.6.8.final.0
• python-bits : 64
• OS : Linux
• OS-release : 4.15.0-51-generic
• machine : x86_64
• processor : x86_64
• byteorder : little
• LC_ALL : None
• LANG : C.UTF-8
• LOCALE : en_US.UTF-8
• pandas : 0.25.1
• numpy : 1.17.2
• pytz : 2019.2
• dateutil : 2.8.0
• pip : 9.0.1
• setuptools : 41.0.1
• Cython : None
• pytest : 4.4.0
• hypothesis : None
• sphinx : None
• blosc : None
• feather : None
• xlsxwriter : None
• lxml.etree : 4.3.3
• html5lib : 0.999999999
• pymysql : None
• psycopg2 : 2.7.7 (dt dec pq3 ext lo64)
• jinja2 : 2.10.1
• IPython : 7.1.1
• pandas_datareader: None
• bs4 : None
• bottleneck : None
• fastparquet : None
• gcsfs : None
• lxml.etree : 4.3.3
• matplotlib : 3.1.1
• numexpr : None
• odfpy : None
• openpyxl : None
• pandas_gbq : None
• pyarrow : None
• pytables : None
• s3fs : None
• scipy : 1.3.1
• sqlalchemy : 1.3.7
• tables : None
• xarray : None
• xlrd : None
• xlwt : None
• xlsxwriter : None

Happy about any help, @jreback ?

[haeusser]

PUSH - this issue is causing serious problems for me. Happy about any feedback.

[jreback]

PUSH - this issue is causing serious problems for me. Happy about any feedback.

your issue has been open for 3 days

we have 3000, when / if someone has a chance they can look

[nrebena]

If you look at df_1 and df_2 after resampling, you'll see that one start at 2019-09-30 09:40:59.984 and the other at 2019-09-30 09:41:59.986. Resampling does not start at the first value of the serie, but at the start of the day. So when resampling, you are not necessarly grouping the data as you think you would (it would work for resampling in a dividor of on minute in your specific case).

Maybe what you wanna do is binning the values every 19 samples. This would give what your expect. Or look at the base option of resampling

But resampling is pretty deterministic.

[haeusser]

Thank you @nrebena for your response. What is the reason for this behavior? It does not make any sense to me. The same data should be binned the exact same way, no matter what the start time is - in my opinion. Cheers, Philip

On Fri 4. Oct 2019 at 00:56, nrebena ***@***.***> wrote: If you look at df_1 and df_2 after resampling, you'll see that one start at 2019-09-30 09:40:59.984 and the other at 2019-09-30 09:41:59.986. Resampling does not start at the first value of the serie, but at the start of the day. So when resampling, you are not necessarly grouping the data as you think you would (it would work for resampling in a dividor of on minute in your specific case). Maybe what you wanna do is binning the values every 19 samples. This would give what your expect. Or look at the base option of resampling But resampling is pretty deterministic. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#28675?email_source=notifications&email_token=ABNSDGEGIYW2WIJ5GYZG5UTQMZMBVA5CNFSM4I3UWOYKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEAJSFNI#issuecomment-538124981>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABNSDGGEXUMOPTUYIOL7JBTQMZMBVANCNFSM4I3UWOYA> .

--

…

____________________________________________________ Dr. Philip Häusser ph@enterbrainment.tv +49 172 5247343 http://enterbrainment.tv

[nrebena]

The data are definitely binned in the same way, no matter the first value. You could consider the bin as fixed discretization of the axis, they do not depend on the values, as they should.

Depending on what you really wanna do, you could also look at pandas.cut, and define your one bin with a DatetimeIndex or something.

[hasB4K]

#31809 should help fix this, I have reproduced your code using the new origin argument:

import numpy as np
import pandas as pd
import datetime as dt

def np_to_df(data, start_time):
    index = pd.DatetimeIndex(
        [start_time + dt.timedelta(milliseconds=t) for t in range(len(data))])
    df = pd.DataFrame(data, index=index)
    return df

data = np.sin(np.arange(1000) / 30)

df_1 = np_to_df(data, dt.datetime(2019, 9, 30, 9, 41))
df_2 = np_to_df(data, dt.datetime(2019, 9, 30, 9, 42))

print("error_1-2:", np.mean(np.abs(df_1.values - df_2.values)))

df_1_resampled = df_1.resample("19ms", origin="start").mean()
df_2_resampled = df_2.resample("19ms", origin="start").mean()

print("error_1-2:", np.mean(np.abs(df_1_resampled.values - df_2_resampled.values)))

Output:

error_1-2: 0.0
error_1-2: 0.0

---

EDIT: use start option on origin.