Cannot cast array data from dtype('float64') to dtype('int32') according to the rule 'safe'

Question

I have a numpy array like

result = np.array([[[289, 354, 331],
                    [291, 206,  66],
                    [242,  70, 256]],

                   [[210, 389, 342],
                    [273, 454, 218],
                    [255,  87, 256]],

                   [[127, 342, 173],
                    [450, 395, 147],
                    [223, 228, 401]]])

and i'm trying to mask the array if an element is greater than 255. ie i assume it to be of range 0-1024 and divide my value by 4

result = np.putmask(result, result > 255, result/4)

NOTE: result is the previous 3D array.And i'm getting this error

TypeError: Cannot cast array data from dtype('float64') to dtype('int32') according to the rule 'safe'

What am i doing wrong? Thanks in advance

Answer 1

Explanation of Error:

This is illustrative of an interesting property of numpy arrays: all elements of a numpy array must be of the same type

For instance if you have the following array:

>>> array1 = np.array([[23, 632, 634],[23.5, 67, 123.6]])
>>> array1
array([[  23. ,  632. ,  634. ],
   [  23.5,   67. ,  123.6]])
>>> type(array1[0][0])
<class 'numpy.float64'>

We notice that even though all the elements in the list [23, 632, 634] were all of the type int (specifically 'numpy.int64' ), all of the elements in array1 were converted to floats because of the element 123.6 in the second row (notice the decimal points in the array print out).

Similarily, if we include even one string anywhere in the array, all the elements of the array are cast to strings:

>>> array2 = np.array([[23, 632, 'foo'],[23.5, 67, 123.6]])
>>> type(array2[0][0])
<class 'numpy.str_'>

Conclusion:

Your original result array contains elements of type 'numpy.int64' , but the result/4 operation returns an array of elements of type 'numpy.float64' (since 82 / 4 = 20.5 , etc.). Thus when you try and replace the values in result , it is not 'safe' as you are inadvertently trying to place floats into an array of ints.

Answer 2

The problem is that when you divide by 4, you are creating float values, which don't want to go into the array of int s.

If you want to use putmask , and avoid the problem of trying to convert to float, then you can use floor division ( // ) in order to change your values to int :

np.putmask(result, result>255, result//4)

>>> result
array([[[ 72,  88,  82],
        [ 72, 206,  66],
        [242,  70,  64]],

       [[210,  97,  85],
        [ 68, 113, 218],
        [255,  87,  64]],

       [[127,  85, 173],
        [112,  98, 147],
        [223, 228, 100]]])

Alternative #1:

Convert your result array to a float dtype , and use your original putmask :

result = result.astype(float)

np.putmask(result, result > 255, result/4)

>>> result
array([[[ 72.25,  88.5 ,  82.75],
        [ 72.75, 206.  ,  66.  ],
        [242.  ,  70.  ,  64.  ]],

       [[210.  ,  97.25,  85.5 ],
        [ 68.25, 113.5 , 218.  ],
        [255.  ,  87.  ,  64.  ]],

       [[127.  ,  85.5 , 173.  ],
        [112.5 ,  98.75, 147.  ],
        [223.  , 228.  , 100.25]]])

You can even convert back to int after if desired:

result = result.astype(int)

array([[[ 72,  88,  82],
        [ 72, 206,  66],
        [242,  70,  64]],

       [[210,  97,  85],
        [ 68, 113, 218],
        [255,  87,  64]],

       [[127,  85, 173],
        [112,  98, 147],
        [223, 228, 100]]])

Alternative #2:

Do away with putmask altogether, and you can get your desired results like this:

result[result > 255] = result[result > 255] / 4

>>> result
array([[[ 72,  88,  82],
        [ 72, 206,  66],
        [242,  70,  64]],

       [[210,  97,  85],
        [ 68, 113, 218],
        [255,  87,  64]],

       [[127,  85, 173],
        [112,  98, 147],
        [223, 228, 100]]])