“AttributeError: 'matrix' object has no attribute 'isocalendar'” in numpy python

Question

I have a numpy matrix as the following:

dates = np.matrix([['09/01/70'], ['10/01/70'], ['11/01/70']])

And I want to convert this matrix to a matrix which involves week numbers instead of the dates.

Thanks to "isocalendar()" function in python, you can find the week number of the date.

But when I used the function as the following;

weeks = dates.isocalendar()[1],

I get this error:

AttributeError: 'matrix' object has no attribute 'isocalendar'

What is the right way of converting the date matrix to datenumber matrix in numpy python?

Answer 1

I would use pandas in this case, converting the matrix into a DataFrame with:

df = pd.DataFrame(dates)

Then I'd convert the dates into datetime objects:

df[0] = pd.to_datetime(df[0])

And then finally apply the isocalendar() function:

df = df[0].apply(lambda x: x.isocalendar()[1])

This returns:

0    36
1    40
2    44
Name: 0, dtype: int64

Then you can get an array of this with df.values or df.as_matrix() . Hope that helps even if it needs the use of a new package.

Answer 2

There are 2 issues: 1) converting strings to datetime objects, 2) iterating over elements of a matrix .

You've given us a matrix that contains strings:

dates = np.matrix([['09/01/70'], ['10/01/70'], ['11/01/70']])

In [286]: dates
Out[286]: 
matrix([['09/01/70'],
        ['10/01/70'],
        ['11/01/70']], 
       dtype='|S8')

Neither the matrix nor the strings have datetime functionality. Those strings have to be converted individually.

def getweek(dstr):
    return datetime.datetime.strptime(dstr,"%d/%m/%y").isocalendar()[1]

In [288]: getweek("15/10/15")
Out[288]: 42

Applying it to the matrix is a bit of a bother since I have to index both row and column.

Just be clear about what is going on, lets iterate over the matrix and return both the date string and the week.

In [285]: [(x[0,0],getweek(x[0,0])) for x in dates]                    
Out[285]: [('09/01/70', 2), ('10/01/70', 2), ('11/01/70', 2)]

getweek could be tweaked to work along with apply_along_axis as suggested in the answer to your other date question.

Are you using np.matrix for a special reason? Why not np.array ? A problem with matrix is that it keeps returning another matrix - another 2d object, even if it only has 1 element.

There is a nice way of turning it into a 1d array:

In [322]: dates.A1
Out[322]: 
array(['09/01/70', '10/01/70', '11/01/70'], 
      dtype='|S8')

With a 1d array:

In [293]: dates = np.array(['09/01/70', '10/01/70', '11/01/70'])
In [294]: [(x,getweek(x)) for x in dates]
Out[294]: [('09/01/70', 2), ('10/01/70', 2), ('11/01/70', 2)]
In [295]: np.array([getweek(x) for x in dates])
Out[295]: array([2, 2, 2])

Usually date strings like this come from a csv file. We can use genfromtxt and getweek to load them

Simulate the file with a text lines, and define converters for genfromtxt .

In [339]: txt=b"""09/01/70
   .....: 10/02/70
   .....: 11/03/70
   .....: """

In [340]: np.genfromtxt(txt.splitlines(),dtype=None,converters={0:getweek})
Out[340]: array([ 2,  7, 11])

---

Here's a way of using np.datetime64 dtype to get weeks

In [350]: txt=b"""1970-09-01
1970-10-15
1970-11-25
1970-01-10"""

In [351]: d=np.genfromtxt(txt.splitlines(),dtype='datetime64[D]')
# load the dates as Days; yyyy-mm-dd is the default format

In [352]: d
Out[352]: array(['1970-09-01', '1970-10-15', '1970-11-25', '1970-01-10'], dtype='datetime64[D]')

In [353]: d.astype('datetime64[W]')-d.astype('datetime64[Y]')    
Out[353]: array([34, 41, 46,  1], dtype='timedelta64[W]')

d.astype('datetime64[W]') converts day units to weeks - though it displays as the first day of the week

d.astype('datetime64[Y]' - is year units, but displays as the start of the year.

Their difference is the number of weeks (0 is a day in the 1st week of the year). There may be a more compact way of deriving this, but I find this approach instructive.