For testing purposes, I'd like to create a M by N numpy array with c randomly placed NaNs
import numpy as np
M = 10;
N = 5;
c = 15;
A = np.random.randn(M,N)
A[mask] = np.nan
I am having problems in creating a mask with c true elements, or maybe this can be done with indices directly?
You can use np.random.choice with the optional replace=False for random selection without replacement and use those on a flattened version of A (done with .ravel()), like so -
A.ravel()[np.random.choice(A.size, c, replace=False)] = np.nan
Sample run -
In [100]: A
Out[100]:
array([[-0.35365726, 0.26754527, -0.44985524, -1.29520237, 2.01505444],
[ 0.01319146, 0.65150356, -2.32054478, 0.40924753, 0.24761671],
[ 0.3014714 , -0.80688589, -2.61431163, 0.07787956, 1.23381951],
[-1.70725777, 0.07856845, -1.04354202, -0.68904925, 1.07161002],
[-1.08061614, 1.17728247, -1.5913516 , -1.87601976, 1.14655867],
[ 1.12542853, -0.26290025, -1.0371326 , 0.53019033, -1.20766258],
[ 1.00692277, 0.171661 , -0.89646634, 1.87619114, -1.04900026],
[ 0.22238353, -0.6523747 , -0.38951426, 0.78449948, -1.14698869],
[ 0.58023183, 1.99987331, -0.85938155, 1.4211672 , -0.43369898],
[-2.15682219, -0.6872121 , -1.28073816, -0.97523148, -2.27967001]])
In [101]: A.ravel()[np.random.choice(A.size, c, replace=False)] = np.nan
In [102]: A
Out[102]:
array([[ nan, 0.26754527, -0.44985524, nan, 2.01505444],
[ 0.01319146, 0.65150356, -2.32054478, nan, 0.24761671],
[ nan, -0.80688589, nan, nan, 1.23381951],
[ nan, nan, -1.04354202, -0.68904925, 1.07161002],
[-1.08061614, 1.17728247, -1.5913516 , nan, 1.14655867],
[ 1.12542853, nan, -1.0371326 , 0.53019033, -1.20766258],
[ nan, 0.171661 , -0.89646634, nan, nan],
[ 0.22238353, -0.6523747 , -0.38951426, 0.78449948, -1.14698869],
[ 0.58023183, 1.99987331, -0.85938155, nan, -0.43369898],
[-2.15682219, -0.6872121 , -1.28073816, -0.97523148, nan]])
np.random.choice with np.random.randint(0,high=A.size,size=c) for my application (if replacement does not really matter). However, why the array does not stay flat after ravel()?
np.random.randint might give you repeated indices, so I don't think that would work in your case. Regarding the .ravel() thing, it's a view only, so it's not exactly flattening in memory. So, the "flattened view" is indexed and set as NaNs, while being kept as a 2D array.ravel() say A copy is made only if needed.. Can it happen that I get a flattened A?You could use np.random.shuffle on a new array to create your mask:
import numpy as np
M = 10;
N = 5;
c = 15;
A = np.random.randn(M,N)
mask=np.zeros(M*N,dtype=bool)
mask[:c] = True
np.random.shuffle(mask)
mask=mask.reshape(M,N)
A[mask] = np.nan
Which gives:
[[ 0.98244168 0.72121195 0.99291217 0.17035834 0.46987918]
[ 0.76919975 0.53102064 nan 0.78776918 nan]
[ 0.50931304 0.91826809 0.52717345 nan nan]
[ 0.35445471 0.28048106 0.91922292 0.76091783 0.43256409]
[ 0.69981284 0.0620876 0.92502572 nan nan]
[ nan nan nan 0.24466688 0.70259211]
[ 0.4916004 nan nan 0.94945378 0.73983538]
[ 0.89057404 0.4542628 nan 0.95547377 nan]
[ 0.4071912 0.36066797 0.73169132 0.48217226 0.62607888]
[ 0.30341337 nan 0.75608859 0.31497997 nan]]
random_choice had that optional replace argument, just worked! :)