Ignoring negative values when using np.log(array)

You can control this behavior with np.seterr. Here's an example.

First, tell numpy to ignore invalid values:

In [4]: old = np.seterr(invalid='ignore')

Now log(-1) doesn't generate a warning:

In [5]: x = np.array([-1.,1])

In [6]: np.log(x)
Out[6]: array([ nan,   0.])

Restore the previous settings:

In [7]: np.seterr(**old)
Out[7]: {'divide': 'warn', 'invalid': 'ignore', 'over': 'warn', 'under': 'ignore'}

And now we get the warning:

In [8]: np.log(x)
/Users/warren/anaconda/bin/ipython:1: RuntimeWarning: invalid value encountered in log
  #!/Users/warren/anaconda/python.app/Contents/MacOS/python
Out[8]: array([ nan,   0.])

There is also a context manager, np.errstate. For example,

In [10]: with np.errstate(invalid='ignore'):
   ....:     y = np.log(x)
   ....:     

In [11]: y
Out[11]: array([ nan,   0.])

You can also use a masked array and NumPy will automatically apply a mask for the invalid values after you perform the np.log() calculation:

a = np.array([1,2,3,0,4,-1,-2])
b = np.log(np.ma.array(a))

print(b.sum())
# 3.17805383035

Where np.ma.array(a) is creating a masked array with no masked elements. It works because NumPy automatically masks elements that are inf (or any invalid value) in calculations with masked arrays.

Alternatively, you could have created the mask yourself (which I recommend) like:

a = np.ma.array(a, mask=(a<=0))

One hack is to limit the values from being negative in the first place. np.clip to the rescue.

positive_array = np.clip(array, some_small_positive_value, None) to avoid negative values in your array. Though I am not sure if bringing the values close to zero serve your purpose.