stopGradient on PyTorch

[enpasos]

One of the enhancements in 0.10.0 was "Adds the NDArray stopGradient and scaleGradient functions (#548)".
It works great on MXNet. Thanks for the feature.

To benefit from one of the 0.10.0 key features "Upgrades PyTorch to 1.7.1" I wired up the stopGradient for PyTorch which makes up the line of change in this pull request.

It works for me ... I have tested it with

    @Test
    void testStopGradient() {

        try (NDManager manager = NDManager.newBaseManager()) {

            NDArray x = manager.create(new float[] {1.0f}, new Shape(1));
            x.attachGradient();
            try (GradientCollector gc = Engine.getInstance().newGradientCollector()) {
                NDArray y = x.mul(x);
                gc.backward(y);
                NDArray grad = x.getGradient();
                System.out.println("grad: " + grad);
                assertEquals(2f, grad.getFloat(0));
            }

            x = manager.create(new float[] {1.0f}, new Shape(1));
            x.attachGradient();
            try (GradientCollector gc = Engine.getInstance().newGradientCollector()) {
                NDArray z = x.mul(x.stopGradient());
                gc.backward(z);
                NDArray grad = x.getGradient();
                System.out.println("grad: " + grad);
                assertEquals(1f, grad.getFloat(0));
            }

        }
    }

[zachgk]

Thanks for your contribution @enpasos! Since you wrote the test for it, do you want to add that to our test suite as well? You can add it in NDArrayOtherOpTest

[stu1130]

Thanks for your contribution @enpasos! Since you wrote the test for it, do you want to add that to our test suite as well? You can add it in NDArrayOtherOpTest

Sorry I didn't see your comment. @enpasos can you raise another PR. Thanks