CODEWITHSUNDEEP
pythonnumpymultidimensional-array
Godfray Qiu
Godfray Qiu

Reputation: 55

Convert string elements in numpy 2D-array to array and yield a 3D-array

I have a numpy 2D-array of string with shape (3,2):

ar_2d = array([['123', '456'],
               ['789', '0ab'],
               ['cde', 'fgh']],
              dtype='<U3')

To make it easier, I am sure the length of each string is equal.

And I have a function, i.e. namely split(), to make string '123' to python list ['1','2','3']

Now I would like to produce a 3D-array with '123' to an array array(['1', '2', '3']) and finally I can get a 3D-array with shape (3,2,3):

ar_3d = array([[['1', '2', '3'],
                ['4', '5', '6']],

                [['7', '8', '9'],
                 ['0', 'a', 'b']],

                [['c', 'd', 'e'],
                 ['f', 'g', 'h']]],
               dtype='<U1')

I have an idea that splitting the string to list first and write to file with numpy's format. Then, I shall read the array from the file.

If the elements are integer numbers, would it be easier? i.e. number 123 to list [1,2,3]

So here is the problem, is there an elegant method to do the convertion?

Thanks in advance!

Upvotes: 2

Views: 814

Answers (1)

Divakar
Divakar

Reputation: 221664

Use ndarray.view to view as U1 and then reshape into 3D -

In [15]: a
Out[15]: 
array([['123', '456'],
       ['789', '0ab'],
       ['cde', 'fgh']], 
      dtype='<U3')

In [16]: a.view('U1').reshape(a.shape + (-1,))
Out[16]: 
array([[['1', '2', '3'],
        ['4', '5', '6']],

       [['7', '8', '9'],
        ['0', 'a', 'b']],

       [['c', 'd', 'e'],
        ['f', 'g', 'h']]], 
      dtype='<U1')

The output would simply be a view of the input and hence this would be memory-efficient. As such, the runtime would be constant (irrespective of array shape) -

In [20]: %timeit a.view('U1').reshape(a.shape + (-1,))
1000000 loops, best of 3: 828 ns per loop

In [21]: a_big = np.tile(a,10000)

In [22]: %timeit a_big.view('U1').reshape(a_big.shape + (-1,))
1000000 loops, best of 3: 851 ns per loop

Upvotes: 2

Related Questions