pytorch int32 to int64 conversion

Question

I'm trying to convert a simple image mask to int64

image = np.array([[1, 2], [3, 4]], dtype='int32')

transform = Compose([
    torch.from_numpy, 
    ConvertImageDtype(torch.int64)
])

However, transform(image) yields

tensor([[ 4294967296,  8589934592],
        [12884901888, 17179869184]])

Is there something wrong, or am I fundamentally misunderstanding something about how the conversion should work?

Answer 1

If you skip torch's conversion, the image is transformed correctly.

image = np.array([[1, 2], [3, 4]], dtype='int64')

transform = Compose([
    torch.from_numpy
])
transform(image)
# tensor([[1, 2],
#        [3, 4]])

Answer 2

I'm late but just in case…

The ConvertImageDtype docstring states:

Convert a tensor image to the given dtype and scale the values accordingly

In the source code, we can see this function calls a F.convert_image_dtype function which then calls a F_t.convert_image_dtype function where we can understand how the scaling is done:

...
input_max = _max_value(image.dtype)
...
output_max = _max_value(dtype)
...
factor = int((output_max + 1) // (input_max + 1))
...
return image * factor

For the int32 representation input_max == 2147483647 == 2**(32-1) -1.
For the int64 representation output_max == 9223372036854775807 == 2**(64-1) -1.

(The -1 in 32-1 and 64-1 is because one byte is used for the sign.
The final -1… I don't know, but we see that factor uses a +1 to "correct" it.)

So we get factor == 4294967296 which is used to scale your image…

Now, in the source code we can also see:

    if image.is_floating_point():
        ...
        # float to int  <= actually written in the code
        ...
    else:
        ...
        # int to float  <= actually written in the code
        ...

So basically the code assumes one won't try anything except "float to int" and "int to float" conversion, but does not strictly check it.