An error occurred while calculating the square of ndarray by using gpu context

[shitou112]

1. When I use a**2 to compute square, I encounter an error.

Input: a = nd.array([-1], mx.gpu())
       a ** 2

Output: [ nan]
        <NDArray 1 @gpu(0)>

2. However, when using ctx=mx.cpu(), the result is right.

Input: a = nd.array([-1], mx.cpu())
       a ** 2

Output: [ 1.]
        <NDArray 1 @cpu(0)>

3. when I use a*a to compute square, the result is also right.

Input: a = nd.array([-1], mx.gpu())
       a * a

Output: [ 1.]
        <NDArray 1 @gpu(0)>

[reminisce]

Since running operators in Python is async with backend, you need to wait until the computation of pow is finished. Try this: (a**2).asnumpy().

[shitou112]

I try asnumpy() function. But the result is wrong.

[reminisce]

That's very strange. I couldn't reproduce the problem you encountered. Does it happen to other versions of MXNet on your machine?

[shitou112]

Thank you a lot. I upgrade mxnet and the result is right.

[vdantu]

@sandeep-krishnamurthy : please label this "Question" and close this. This seems to have been resolved.