Created by: Aurelius84

PR types

Bug fixes

PR changes

APIs

Describe

fix unused var with zero gradient

Example code: x1, x2, x3 are output vars from same op, but only x1 is used to calculate gradient, so x2@GRAD、x3@GRAD should be zero but not before in this PR. They are not initialized as expected.

class TestGradientWithPrune(unittest.TestCase):
    def test_prune(self):
        x = fluid.data(name='x', shape=[3], dtype='float32')
        x.stop_gradient = False
        x1, x2, x3 = fluid.layers.split(x, dim=0, num_or_sections=3)
        y = x1 * 2
        x1_grad = fluid.gradients(y, x)

        exe = fluid.Executor(fluid.CPUPlace())
        main = fluid.default_main_program()
        exe.run(fluid.default_startup_program())
        out = exe.run(main,
                      feed={'x': np.ones([3]).astype('float32')},
                      fetch_list=[x1_grad])
        self.assertTrue(np.array_equal(out[0], [2., 0., 0.]))