resize with averaging or rebin a numpy 2d array

Question

I am trying to reimplement in python an IDL function:

http://star.pst.qub.ac.uk/idl/REBIN.html

which downsizes by an integer factor a 2d array by averaging.

For example:

>>> a=np.arange(24).reshape((4,6)) >>> a array([[ 0,  1,  2,  3,  4,  5],        [ 6,  7,  8,  9, 10, 11],        [12, 13, 14, 15, 16, 17],        [18, 19, 20, 21, 22, 23]]) 

I would like to resize it to (2,3) by taking the mean of the relevant samples, the expected output would be:

>>> b = rebin(a, (2, 3)) >>> b array([[  3.5,   5.5,  7.5],        [ 15.5, 17.5,  19.5]]) 

i.e. b[0,0] = np.mean(a[:2,:2]), b[0,1] = np.mean(a[:2,2:4]) and so on.

I believe I should reshape to a 4 dimensional array and then take the mean on the correct slice, but could not figure out the algorithm. Would you have any hint?

Answer 1

Here's an example based on the answer you've linked (for clarity):

>>> import numpy as np >>> a = np.arange(24).reshape((4,6)) >>> a array([[ 0,  1,  2,  3,  4,  5],        [ 6,  7,  8,  9, 10, 11],        [12, 13, 14, 15, 16, 17],        [18, 19, 20, 21, 22, 23]]) >>> a.reshape((2,a.shape[0]//2,3,-1)).mean(axis=3).mean(1) array([[  3.5,   5.5,   7.5],        [ 15.5,  17.5,  19.5]]) 

As a function:

def rebin(a, shape):     sh = shape[0],a.shape[0]//shape[0],shape[1],a.shape[1]//shape[1]     return a.reshape(sh).mean(-1).mean(1) 

Answer 2

J.F. Sebastian has a great answer for 2D binning. Here is a version of his "rebin" function that works for N dimensions:

def bin_ndarray(ndarray, new_shape, operation='sum'):     """     Bins an ndarray in all axes based on the target shape, by summing or         averaging.      Number of output dimensions must match number of input dimensions and          new axes must divide old ones.      Example     -------     >>> m = np.arange(0,100,1).reshape((10,10))     >>> n = bin_ndarray(m, new_shape=(5,5), operation='sum')     >>> print(n)      [[ 22  30  38  46  54]      [102 110 118 126 134]      [182 190 198 206 214]      [262 270 278 286 294]      [342 350 358 366 374]]      """     operation = operation.lower()     if not operation in ['sum', 'mean']:         raise ValueError("Operation not supported.")     if ndarray.ndim != len(new_shape):         raise ValueError("Shape mismatch: {} -> {}".format(ndarray.shape,                                                            new_shape))     compression_pairs = [(d, c//d) for d,c in zip(new_shape,                                                   ndarray.shape)]     flattened = [l for p in compression_pairs for l in p]     ndarray = ndarray.reshape(flattened)     for i in range(len(new_shape)):         op = getattr(ndarray, operation)         ndarray = op(-1*(i+1))     return ndarray