How to generate pandas dataframe timedelta column grouped by id and date (YYYY-MM-DD)?

Suppose I have a dataframe with id and datetime columns:

df = pd.DataFrame({"id": ["a1", "a1", "a1", "a1", "a2", "a2", "a2", "a2", "a3", "a3", "a3", "a3"],
                   "datetime": ["2016-01-01 00:01:00.156",
                                "2016-01-01 12:00:00.425",
                                "2016-01-02 00:59:00.123",
                                "2016-01-02 14:16:00.548",
                                "2016-01-01 12:00:00.147",
                                "2016-01-01 13:59:00.123",
                                "2016-01-02 08:01:00.147",
                                "2016-01-02 18:49:00.123",
                                "2016-02-01 12:00:00.147",
                                "2016-02-01 13:59:00.123",
                                "2016-02-02 08:01:00.147",
                                "2016-02-02 18:49:00.123"]})
df["datetime"] = pd.to_datetime(df["datetime"])
df

Here is the dataframe:

    id	datetime
0	a1	2016-01-01 00:01:00.156
1	a1	2016-01-01 12:00:00.425
2	a1	2016-01-02 00:59:00.123
3	a1	2016-01-02 14:16:00.548
4	a2	2016-01-01 12:00:00.147
5	a2	2016-01-01 13:59:00.123
6	a2	2016-01-02 08:01:00.147
7	a2	2016-01-02 18:49:00.123
8	a3	2016-02-01 12:00:00.147
9	a3	2016-02-01 13:59:00.123
10	a3	2016-02-02 08:01:00.147
11	a3	2016-02-02 18:49:00.123

I want to generate column timedelta that has a timedelta value. This is the output I expect to get:

    id	datetime	            datetime_baseline	    timedelta
0	a1	2016-01-01 00:01:00.156	2016-01-01 00:01:00.156	0
1	a1	2016-01-01 12:00:00.425	2016-01-01 00:01:00.156	719
2	a1	2016-01-02 00:59:00.123	2016-01-02 00:59:00.123	0
3	a1	2016-01-02 14:16:00.548	2016-01-02 00:59:00.123	797
4	a2	2016-01-01 12:00:00.147	2016-01-01 12:00:00.147	0
5	a2	2016-01-01 13:59:00.123	2016-01-01 12:00:00.147	119
6	a2	2016-01-02 08:01:00.147	2016-01-02 08:01:00.147	0
7	a2	2016-01-02 18:49:00.123	2016-01-02 08:01:00.147	648
8	a3	2016-02-01 12:00:00.147	2016-02-01 12:00:00.147	0
9	a3	2016-02-01 13:59:00.123	2016-02-01 12:00:00.147	119
10	a3	2016-02-02 08:01:00.147	2016-02-02 08:01:00.147	0
11	a3	2016-02-02 18:49:00.123	2016-02-02 08:01:00.147	648

Here is how the timedelta values should be calculated: 1) the code needs to identify the FIRST datetime within the same id and date ('YYYY-MM-DD'), and 2) use it as baseline (datetime_baseline) to compute the timedelta (in minutes) w.r.t. other datetimes within same id and same date. For id='a1' and date='2016-01-01', the datetime_baseline='2016-01-01 00:01:00.156'. So, at index=0, timedelta has value=0 because '2016-01-01 00:01:00.156' - datetime_baseline=0. Meanwhile, at index=1, timedelta has value=719 because '2016-01-01 12:00:00.425' - datetime_baseline=719 (minutes). At index=2, id is the same as before but date is now '2016-01-02', so a new baseline will be used: '2016-01-02 00:59:00.123'. timedelta='2016-01-02 00:59:00.123' - datetime_baseline=0. At index=3, timedelta='2016-01-02 14:16:00.548' - datetime_baseline=797.

Although I see how the timedelta values should be calculated (timedelta=datetime-datetime_baseline), I don't know how to have the baseline values identified (i.e. how to generate datetime_baseline column). Please, let me know if you need any further explanation.

ps> the actual dataframe has +500 thousand rows.

With GroupBy.transform to make the baseline :

df["datetime_baseline"] = (df.groupby(["id", df["datetime"].dt.date])
["datetime"].transform("first"))

And dt.total_seconds to compute the timedelta :

df["timedelta"] = ((df["datetime"].sub(df["datetime_baseline"]))
.dt.total_seconds().div(60).round(0).astype(int))

​
Output :

print(df)
id                datetime       datetime_baseline  timedelta
0   a1 2016-01-01 00:01:00.156 2016-01-01 00:01:00.156          0
1   a1 2016-01-01 12:00:00.425 2016-01-01 00:01:00.156        719
2   a1 2016-01-02 00:59:00.123 2016-01-02 00:59:00.123          0
3   a1 2016-01-02 14:16:00.548 2016-01-02 00:59:00.123        797
4   a2 2016-01-01 12:00:00.147 2016-01-01 12:00:00.147          0
5   a2 2016-01-01 13:59:00.123 2016-01-01 12:00:00.147        119
6   a2 2016-01-02 08:01:00.147 2016-01-02 08:01:00.147          0
7   a2 2016-01-02 18:49:00.123 2016-01-02 08:01:00.147        648
8   a3 2016-02-01 12:00:00.147 2016-02-01 12:00:00.147          0
9   a3 2016-02-01 13:59:00.123 2016-02-01 12:00:00.147        119
10  a3 2016-02-02 08:01:00.147 2016-02-02 08:01:00.147          0
11  a3 2016-02-02 18:49:00.123 2016-02-02 08:01:00.147        648

Try:

df['datetime_baseline'] = df.groupby(['id', df['datetime'].dt.date])["datetime"].transform('min')
df['timedelta'] = np.round((df['datetime'] - df['datetime_baseline']).dt.seconds / 60)

print(df)

Prints:

    id                datetime       datetime_baseline  timedelta
0   a1 2016-01-01 00:01:00.156 2016-01-01 00:01:00.156        0.0
1   a1 2016-01-01 12:00:00.425 2016-01-01 00:01:00.156      719.0
2   a1 2016-01-02 00:59:00.123 2016-01-02 00:59:00.123        0.0
3   a1 2016-01-02 14:16:00.548 2016-01-02 00:59:00.123      797.0
4   a2 2016-01-01 12:00:00.147 2016-01-01 12:00:00.147        0.0
5   a2 2016-01-01 13:59:00.123 2016-01-01 12:00:00.147      119.0
6   a2 2016-01-02 08:01:00.147 2016-01-02 08:01:00.147        0.0
7   a2 2016-01-02 18:49:00.123 2016-01-02 08:01:00.147      648.0
8   a3 2016-02-01 12:00:00.147 2016-02-01 12:00:00.147        0.0
9   a3 2016-02-01 13:59:00.123 2016-02-01 12:00:00.147      119.0
10  a3 2016-02-02 08:01:00.147 2016-02-02 08:01:00.147        0.0
11  a3 2016-02-02 18:49:00.123 2016-02-02 08:01:00.147      648.0