I have a dataset called MEL of shape (94824,) wherein most instances have shape (99, 13) but some have smaller shapes. It consists of (float) MEL frequencies. I'm trying to put all the values in an empty numpy matrix of shape (94824, 99 , 13). So some instances are left empty. Any suggestions?
MEL type = numpy.ndarray
for i in MEL type(i) = <class 'numpy.ndarray'>
for j in i type (j) = <class 'numpy.ndarray'>
Since your MEL array is not of homogeneous shape, first we need to filter out the arrays whose shape is common (i.e. (99, 13)). For this, we could use:
filtered = []
for arr in MEL:
if arr.shape == (99, 13):
filtered.append(arr)
else:
continue
Then we can initialize an array to hold the results. And then we can iterate over this filtered list of arrays and calculate the mean over axis 1 like:
averaged_arr = np.zeros((len(filtered), 99))
for idx, arr in enumerate(filtered):
averaged_arr[idx] = np.mean(arr, axis=1)
This should compute the desired matrix.
Here is a demo to reproduce your setup, assuming all arrays of the same shape:
# inputs
In [20]: MEL = np.empty(94824, dtype=np.object)
In [21]: for idx in range(94824):
...: MEL[idx] = np.random.randn(99, 13)
# shape of the array of arrays
In [13]: MEL.shape
Out[13]: (94824,)
# shape of each array
In [15]: MEL[0].shape
Out[15]: (99, 13)
# to hold results
In [17]: averaged_arr = np.zeros((94824, 99))
# compute average
In [18]: for idx, arr in enumerate(MEL):
...: averaged_arr[idx] = np.mean(arr, axis=1)
# check the shape of resultant array
In [19]: averaged_arr.shape
Out[19]: (94824, 99)
(99, 13)? I think then the dataset would be small, but that's the only approach that comes to my mind atm