numpy数组缩放未返回正确的值

Question

I have a numpy array that I want to alter by scaling all of the columns (eg all the values in a column are divided by the maximum value in that column so that all values are <1).

A sample output of the array is

[ 2. 0. 367.877 ..., -0.358 51.547 -32.633]

[ 2. 0. 339.824 ..., -0.33 52.562 -27.581]

[ 3. 0. 371.438 ..., -0.406 55.108 -35.573]

I've tried scaling the array (data_in) by the following code:

#normalize the data_in array 
data_in_normalized = data_in / data_in.max(axis=0)

However, the output of data_in_normalized is:

[ 0.5 0. 0.95437199 0.89363654 0.80751792 ]

[ 0.46931238 0.50660904 0.5003812 0.91250444 0.625 ]

[ 0.96229214 0.89483109 0.86989432 0.86491407 0.71287646 ]

[ -23.90909091 0.34346373 1.25110652 0. 0.8537859 1. 1.]

Clearly, it didn't normalize--there are multiple areas where the maximum value is >1. Is there a better way to scale the data, or am I using the max() function incorrectly (eg is the max() value being shared between columns?)

Answer 1

IIUC, it's not that the maximum value is shared between columns, it's that you probably want to divide by the maximum absolute value instead, because you have elements of both signs. 1 > -100, after all, and so if you divide by the maximum value of a column with [1, -100], nothing would change.

For example:

>>> data_in = np.array([[-3,-2],[2,1]])
>>> data_in
array([[-3, -2],
       [ 2,  1]])
>>> data_in.max(axis=0)
array([2, 1])
>>> data_in / data_in.max(axis=0)
array([[-1.5, -2. ],
       [ 1. ,  1. ]])

but

>>> data_in / np.abs(data_in).max(axis=0)
array([[-1.        , -1.        ],
       [ 0.66666667,  0.5       ]])