Count number of non-nan values in array

Question

I have a list of floats/nan values, that looks like this:

a = [(9.62, np.nan, 0.063), (np.nan, np.nan, np.nan), (np.nan, 0.34, np.nan), (9.50, 2.65, 5.85), (np.nan, np.nan, np.nan), (8.9423173497260166e-06, np.nan, np.nan), (np.nan, np.nan, np.nan), (10.53906499271581, np.nan, 3.4981897643207153e-08), (27.945228892337656, np.nan, np.nan), (np.nan, np.nan, np.nan), (0.00015676098048248007, 428.53224664333368, 15.597030989617416), (3.219339103511719e-08, np.nan, np.nan), (351.3486881626871, 118.79412856376891, 96.925698744436318), (np.nan, np.nan, np.nan), (np.nan, np.nan, np.nan), (0.038185812702743384, 0.011979539923543838, 1.4161404311887908e-05), (6.5891883211951452, np.nan, np.nan), (np.nan, np.nan, np.nan), (np.nan, np.nan, np.nan), (np.nan, np.nan, np.nan), (0.01992113565158183, 1.0858887135978378e-08, 6.949483102803238e-08), (np.nan, np.nan, np.nan), (0.0053471054969118897, 32.364223190908589, 0.29950485126829518), (0.022687094833899225, np.nan, 3.3927513616780456e-05), (0.0065459356887503, 5.0304474154655309e-06, 6.1755309734841293e-06), (1.2854278279876815e-07, 110.94572059986106, 2.0737305081677166e-06), (2.8909153747692473, np.nan, np.nan), (np.nan, np.nan, np.nan), (0.00085244354118369653, np.nan, 547.28608997823414), (0.21609437779080298, 2.9772785752782283e-08, 0.024868855470372788), (np.nan, 1.0571674432090431e-08, np.nan), (np.nan, 0.00042711039439664552, np.nan), (np.nan, 3.7576842775630178e-09, np.nan), (np.nan, 1.2436122988008544e-08, np.nan), (np.nan, 0.008772060008242254, np.nan), (np.nan, 2.9731267579988852, np.nan), (np.nan, 152.69348161610276, np.nan), (np.nan, 1.7976907012194907, np.nan), (np.nan, 0.0006232073677262973, np.nan), (np.nan, 1.3468250342036237e-08, np.nan), (np.nan, 6.9699321813542907e-05, np.nan), (np.nan, 5.2001506649804148e-05, np.nan), (np.nan, np.nan, np.nan)]

i.e.: made up of N sub-lists, each one containing the same number of elements M (in this case 3, but it could change), where each of those elements is either a float or a np.nan value (my actual list has much larger N and M values).

I need to efficiently count the number of non np.nan values in each sublist. If the number is zero (all np.nan values), a np.nan value should be stored.

The final list/array would look like (using a above):

count = [2, nan, 1, 3, ...]

I tried with np.count_nonzero() but it counts np.nan as non-zeros, so it returns all counts as 3.

Answer 1

You can use numpy.isnan to create a boolean array, and then count the Trues with sum for each row (axis=1):

import numpy as np
# count the non-nan values
non_nans = (~np.isnan(a)).sum(1)

# replace 0 count with np.nan
np.where(non_nans == 0, np.nan, non_nans)
# array([  2.,  nan,   1.,   3.,  nan, ...])