Replace NaN's in NumPy array with closest non-

问题:

I have a NumPy array a like the following:

>>> str(a)
'[        nan         nan         nan  1.44955726  1.44628034  1.44409573\n  1.4408188   1.43657094  1.43171624  1.42649744  1.42200684  1.42117704\n  1.42040255  1.41922908         nan         nan         nan         nan\n         nan         nan]'

I want to replace each NaN with the closest non-NaN value, so that all of the NaN's at the beginning get set to 1.449... and all of the NaN's at the end get set to 1.419....

I can see how to do this for specific cases like this, but I need to be able to do it generally for any length of array, with any length of NaN's at the beginning and end of the array (there will be no NaN's in the middle of the numbers). Any ideas?

I can find the NaN's easily enough with np.isnan(), but I can't work out how to get the closest value to each NaN.

回答1:

I want to replace each NaN with the closest non-NaN value... there will be no NaN's in the middle of the numbers

The following will do it:

ind = np.where(~np.isnan(a))[0]
first, last = ind[0], ind[-1]
a[:first] = a[first]
a[last + 1:] = a[last]

This is a straight numpy solution requiring no Python loops, no recursion, no list comprehensions etc.

回答2:

As an alternate solution (this will linearly interpolate for arrays NaNs in the middle, as well):

import numpy as np

# Generate data...
data = np.random.random(10)
data[:2] = np.nan
data[-1] = np.nan
data[4:6] = np.nan

print data

# Fill in NaN's...
mask = np.isnan(data)
data[mask] = np.interp(np.flatnonzero(mask), np.flatnonzero(~mask), data[~mask])

print data

This yields:

[        nan         nan  0.31619306  0.25818765         nan         nan
  0.27410025  0.23347532  0.02418698         nan]

[ 0.31619306  0.31619306  0.31619306  0.25818765  0.26349185  0.26879605
  0.27410025  0.23347532  0.02418698  0.02418698]

回答3:

NaNs have the interesting property of comparing different from themselves, thus we can quickly find the indexes of the non-nan elements:

idx = np.nonzero(a==a)[0]

it's now easy to replace the nans with the desired value:

for i in range(0, idx[0]):
    a[i]=a[idx[0]]
for i in range(idx[-1]+1, a.size)
    a[i]=a[idx[-1]]

Finally, we can put this in a function:

import numpy as np

def FixNaNs(arr):
    if len(arr.shape)>1:
        raise Exception("Only 1D arrays are supported.")
    idxs=np.nonzero(arr==arr)[0]

    if len(idxs)==0:
        return None

    ret=arr

    for i in range(0, idxs[0]):
        ret[i]=ret[idxs[0]]

    for i in range(idxs[-1]+1, ret.size):
        ret[i]=ret[idxs[-1]]

    return ret

edit

Ouch, coming from C++ I always forget about list ranges... @aix's solution is way more elegant and efficient than my C++ish loops, use that instead of mine.

回答4:

A recursive solution!

def replace_leading_NaN(a, offset=0):
    if a[offset].isNaN():
        new_value = replace_leading_NaN(a, offset + 1)
        a[offset] = new_value
        return new_value
    else:
        return a[offset]

def replace_trailing_NaN(a, offset=-1):
    if a[offset].isNaN():
        new_value = replace_trailing_NaN(a, offset - 1)
        a[offset] = new_value
        return new_value
    else:
        return a[offset]

回答5:

I came across the problem and had to find a custom solution for scattered NaNs. The function below replaces any NaN by the first number occurrence to the right, if none exists, it replaces it by the first number occurrence to the left. Further manipulation can be done to replace it with the mean of boundary occurrences.

import numpy as np

Data = np.array([np.nan,1.3,np.nan,1.4,np.nan,np.nan])

nansIndx = np.where(np.isnan(Data))[0]
isanIndx = np.where(~np.isnan(Data))[0]
for nan in nansIndx:
    replacementCandidates = np.where(isanIndx>nan)[0]
    if replacementCandidates.size != 0:
        replacement = Data[isanIndx[replacementCandidates[0]]]
    else:
        replacement = Data[isanIndx[np.where(isanIndx<nan)[0][-1]]]
    Data[nan] = replacement

Result is:

>>> Data
array([ 1.3,  1.3,  1.4,  1.4,  1.4,  1.4])

回答6:

I got something like this

i = [i for i in range(len(a)) if not np.isnan(a[i])]
a = [a[i[0]] if x < i[0] else (a[i[-1]] if x > i[-1] else a[x]) for x in range(len(a))]

It's a bit clunky though given it's split up in two lines with nested inline if's in one of them.

回答7:

Here is a solution using simple python iterators. They are actually more efficient here than numpy.where, especially with big arrays! See comparison of similar code here.

import numpy as np

a = np.array([np.NAN, np.NAN, np.NAN, 1.44955726, 1.44628034, 1.44409573, 1.4408188, 1.43657094, 1.43171624,  1.42649744, 1.42200684, 1.42117704, 1.42040255, 1.41922908, np.NAN, np.NAN, np.NAN, np.NAN, np.NAN, np.NAN])

mask = np.isfinite(a)

# get first value in list
for i in range(len(mask)):
    if mask[i]:
        first = i
        break

# get last vaue in list
for i in range(len(mask)-1, -1, -1):
    if mask[i]:
        last = i
        break

# fill NaN with near known value on the edges
a = np.copy(a)
a[:first] = a[first]
a[last + 1:] = a[last]

print(a)

Output:

[1.44955726 1.44955726 1.44955726 1.44955726 1.44628034 1.44409573
 1.4408188  1.43657094 1.43171624 1.42649744 1.42200684 1.42117704
 1.42040255 1.41922908 1.41922908 1.41922908 1.41922908 1.41922908
 1.41922908 1.41922908]

It replaces only the first and last NaNs like requested here.