If partial is your original data, and replace is an array of the same shape containing averaged values then this code will use the value from partial if one exists.

Complete= np.where(np.isnan(partial),replace,partial)

Using masked arrays

The standard way to do this using only numpy would be to use the masked array module.

Scipy is a pretty heavy package which relies on external libraries, so it's worth having a numpy-only method. This borrows from @DonaldHobson's answer.

Edit: np.nanmean is now a numpy function. However, it doesn't handle all-nan columns...

Suppose you have an array a:

>>> a
array([[  0.,  nan,  10.,  nan],
       [  1.,   6.,  nan,  nan],
       [  2.,   7.,  12.,  nan],
       [  3.,   8.,  nan,  nan],
       [ nan,   9.,  14.,  nan]])

>>> import numpy.ma as ma
>>> np.where(np.isnan(a), ma.array(a, mask=np.isnan(a)).mean(axis=0), a)    
array([[  0. ,   7.5,  10. ,   0. ],
       [  1. ,   6. ,  12. ,   0. ],
       [  2. ,   7. ,  12. ,   0. ],
       [  3. ,   8. ,  12. ,   0. ],
       [  1.5,   9. ,  14. ,   0. ]])

Note that the masked array's mean does not need to be the same shape as a, because we're taking advantage of the implicit broadcasting over rows.

Also note how the all-nan column is nicely handled. The mean is zero since you're taking the mean of zero elements. The method using nanmean doesn't handle all-nan columns:

>>> col_mean = np.nanmean(a, axis=0)
/home/praveen/.virtualenvs/numpy3-mkl/lib/python3.4/site-packages/numpy/lib/nanfunctions.py:675: RuntimeWarning: Mean of empty slice
  warnings.warn("Mean of empty slice", RuntimeWarning)
>>> inds = np.where(np.isnan(a))
>>> a[inds] = np.take(col_mean, inds[1])
>>> a
array([[  0. ,   7.5,  10. ,   nan],
       [  1. ,   6. ,  12. ,   nan],
       [  2. ,   7. ,  12. ,   nan],
       [  3. ,   8. ,  12. ,   nan],
       [  1.5,   9. ,  14. ,   nan]])

Explanation

Converting a into a masked array gives you

>>> ma.array(a, mask=np.isnan(a))
masked_array(data =
 [[0.0 --  10.0 --]
  [1.0 6.0 --   --]
  [2.0 7.0 12.0 --]
  [3.0 8.0 --   --]
  [--  9.0 14.0 --]],
             mask =
 [[False  True False  True]
 [False False  True  True]
 [False False False  True]
 [False False  True  True]
 [ True False False  True]],
       fill_value = 1e+20)

And taking the mean over columns gives you the correct answer, normalizing only over the non-masked values:

>>> ma.array(a, mask=np.isnan(a)).mean(axis=0)
masked_array(data = [1.5 7.5 12.0 --],
             mask = [False False False  True],
       fill_value = 1e+20)

Further, note how the mask nicely handles the column which is all-nan!

Finally, np.where does the job of replacement.

Row-wise mean

To replace nan values with row-wise mean instead of column-wise mean requires a tiny change for broadcasting to take effect nicely:

>>> a
array([[  0.,   1.,   2.,   3.,  nan],
       [ nan,   6.,   7.,   8.,   9.],
       [ 10.,  nan,  12.,  nan,  14.],
       [ nan,  nan,  nan,  nan,  nan]])

>>> np.where(np.isnan(a), ma.array(a, mask=np.isnan(a)).mean(axis=1), a)
ValueError: operands could not be broadcast together with shapes (4,5) (4,) (4,5)

>>> np.where(np.isnan(a), ma.array(a, mask=np.isnan(a)).mean(axis=1)[:, np.newaxis], a)
array([[  0. ,   1. ,   2. ,   3. ,   1.5],
       [  7.5,   6. ,   7. ,   8. ,   9. ],
       [ 10. ,  12. ,  12. ,  12. ,  14. ],
       [  0. ,   0. ,   0. ,   0. ,   0. ]])

This isn't very clean but I can't think of a way to do it other than iterating

#example
a = np.arange(16, dtype = float).reshape(4,4)
a[2,2] = np.nan
a[3,3] = np.nan

indices = np.where(np.isnan(a)) #returns an array of rows and column indices
for row, col in zip(*indices):
    a[row,col] = np.mean(a[~np.isnan(a[:,col]), col])

Alternative: Replacing NaNs with interpolation of columns.

def interpolate_nans(X):
    """Overwrite NaNs with column value interpolations."""
    for j in range(X.shape[1]):
        mask_j = np.isnan(X[:,j])
        X[mask_j,j] = np.interp(np.flatnonzero(mask_j), np.flatnonzero(~mask_j), X[~mask_j,j])
    return X

Example use:

X_incomplete = np.array([[10,     20,     30    ],
                         [np.nan, 30,     np.nan],
                         [np.nan, np.nan, 50    ],
                         [40,     50,     np.nan    ]])

X_complete = interpolate_nans(X_incomplete)

print X_complete
[[10,     20,     30    ],
 [20,     30,     40    ],
 [30,     40,     50    ],
 [40,     50,     50    ]]

I use this bit of code for time series data in particular, where columns are attributes and rows are time-ordered samples.

To extend Donald's Answer I provide a minimal example. Let's say a is an ndarray and we want to replace its zero values with the mean of the column.

In [231]: a
Out[231]: 
array([[0, 3, 6],
       [2, 0, 0]])


In [232]: col_mean = np.nanmean(a, axis=0)
Out[232]: array([ 1. ,  1.5,  3. ])

In [228]: np.where(np.equal(a, 0), col_mean, a)
Out[228]: 
array([[ 1. ,  3. ,  6. ],
       [ 2. ,  1.5,  3. ]])

If you don't have access to a numpy version with numpy.nanmean, this is a compact way to do it, by indexing into the means.

m = numpy.isnan(a)
a[m] = numpy.mean(numpy.nan_to_num(a), axis=0)[numpy.where(m)[0]]

import numpy as np
import pandas as pd

def replace_nan_with_mean(matrix_in):
    for i in np.arange(matrix_in.shape[1]):
        mean=np.mean(matrix_in[~pd.isnull(matrix_in[:,i]),i],axis=0) # caculate the mean, ignoring all the na values
        matrix_in[pd.isnull(matrix_in[:,i]),i]=mean; # assign all nan value with mean
    return matrix_in    

you might want to try this built-in function:

x = np.array([np.inf, -np.inf, np.nan, -128, 128])
np.nan_to_num(x)
array([  1.79769313e+308,  -1.79769313e+308,   0.00000000e+000,
-1.28000000e+002,   1.28000000e+002])