Find Unique Dates in Numpy Datetime Array

Question

I have timeseries data (epoch, values) which i have transformed into (datetime, values), which is stored in Numpy arrays. Now i wish to find the indexes of the first row corresponding to a given day. Thus, only a single index per day is needed.

Following is a purely Python function which is very slow.

def day_wise_datetime(datetimes,dataseries):
    unique_dates=[]
    unique_indices=[]
    for i in range(len(datetimes)):
        if datetimes[i].day not in unique_dates:
           unique_dates.append(datetimes[i])
           unique_indices.append(i)
    return [unique_dates,unique_indices]

Numpy provides a unique method, but it says that it cannot sort datetime. So what Numpy based technique can be used for the same.

I know that Pandas is recommended, but while i am learning it, would like to know if some NumPy/SciPy solution might suffice.

EDIT The value in datetimes variable are like. I have just sliced the first five elements.

[datetime.datetime(2011, 4, 18, 18, 52, 9),
datetime.datetime(2011, 4, 18, 18, 52, 10),
datetime.datetime(2011, 4, 18, 18, 52, 11),
datetime.datetime(2011, 4, 18, 18, 52, 12),
datetime.datetime(2011, 4, 18, 18, 52, 13)]

Answer 1

pandas's DataFrame provides drop_duplictes which can easily achieve your goal:

In [121]: arr1 = np.array([dt.datetime(2013, 1, 1), dt.datetime(2013, 1, 1), dt.datetime(2013, 1, 2)]) 

In [122]: arr2 = np.array([1, 2, 3]) 

In [123]: df = pd.DataFrame({'date': arr1, 'value': arr2})

In [124]: df
Out[124]: 
                 date  value
0 2013-01-01 00:00:00      1   
1 2013-01-01 00:00:00      2   
2 2013-01-02 00:00:00      3   

In [125]: df.drop_duplicates('date')
Out[125]: 
                 date  value
0 2013-01-01 00:00:00      1   
2 2013-01-02 00:00:00      3 

EDIT

I misunderstood your problem in the very beginning. Please try following one:

Seems sorting is one of your mainly problem, I create the example as a reversed datetime list:

In [74]: now = dt.datetime.utcnow()
In [75]: datetimes = [now - dt.timedelta(hours=6) * i for i in range(10)]

In [76]: datetimes
Out[76]:
[datetime.datetime(2013, 5, 8, 16, 47, 32, 60500),
 datetime.datetime(2013, 5, 8, 10, 47, 32, 60500),
 datetime.datetime(2013, 5, 8, 4, 47, 32, 60500),
 datetime.datetime(2013, 5, 7, 22, 47, 32, 60500),
 datetime.datetime(2013, 5, 7, 16, 47, 32, 60500),
 datetime.datetime(2013, 5, 7, 10, 47, 32, 60500),
 datetime.datetime(2013, 5, 7, 4, 47, 32, 60500),
 datetime.datetime(2013, 5, 6, 22, 47, 32, 60500),
 datetime.datetime(2013, 5, 6, 16, 47, 32, 60500),
 datetime.datetime(2013, 5, 6, 10, 47, 32, 60500)]

Create a DataFrame by datetimes and set the column name as date:

In [81]: df = pd.DataFrame(datetimes, columns=['date'])

In [82]: df
Out[82]:
                        date
0 2013-05-08 16:47:32.060500
1 2013-05-08 10:47:32.060500
2 2013-05-08 04:47:32.060500
3 2013-05-07 22:47:32.060500
4 2013-05-07 16:47:32.060500
5 2013-05-07 10:47:32.060500
6 2013-05-07 04:47:32.060500
7 2013-05-06 22:47:32.060500
8 2013-05-06 16:47:32.060500
9 2013-05-06 10:47:32.060500

Next, sort your DataFrame by the date column:

In [83]: df = df.sort('date')

And then append a new columns for index:

In [85]: df['index'] = df['date'].apply(lambda x:x.day)

In [86]: df
Out[86]:
                        date  index
9 2013-05-06 10:47:32.060500      6
8 2013-05-06 16:47:32.060500      6
7 2013-05-06 22:47:32.060500      6
6 2013-05-07 04:47:32.060500      7
5 2013-05-07 10:47:32.060500      7
4 2013-05-07 16:47:32.060500      7
3 2013-05-07 22:47:32.060500      7
2 2013-05-08 04:47:32.060500      8
1 2013-05-08 10:47:32.060500      8
0 2013-05-08 16:47:32.060500      8

Then group your data by index, and then get the first one for each group. If you are familiar with SQL, it just like SELECT FIRST(*) FROM table GROUP BY table.index:

In [87]: df = df.groupby('index').first()
In [88]: df
Out[88]: 
                            date
index                           
6     2013-05-06 10:47:32.060500
7     2013-05-07 04:47:32.060500
8     2013-05-08 04:47:32.060500

Now you can get the unique indices:

In [91]: df.index.values
Out[91]: array([6, 7, 8])

And get the unique dates:

In [92]: df['date'].values
Out[92]: 
array(['2013-05-06T18:47:32.060500000+0800',
   '2013-05-07T12:47:32.060500000+0800',
   '2013-05-08T12:47:32.060500000+0800'], dtype='datetime64[ns]')