Method gradients_check

test/python/test_operation.py:348–383

(self,
                        func,
                        param,
                        autograds,
                        h=0.0005,
                        df=1,
                        dev=cpu_dev)

        self._batchnorm2d_helper(gpu_dev)

    def gradients_check(self,
                        func,
                        param,
                        autograds,
                        h=0.0005,
                        df=1,
                        dev=cpu_dev):
        # param: PyTensor
        # autograds: numpy_tensor
        p = tensor.to_numpy(param)
        it = np.nditer(p, flags=['multi_index'], op_flags=['readwrite'])
        while not it.finished:
            idx = it.multi_index
            diff = np.zeros_like(p)
            diff[idx] += h
            diff = tensor.from_numpy(diff)
            diff.to_device(dev)

            param += diff
            pos = func()
            pos = tensor.to_numpy(pos)

            param -= diff
            param -= diff
            neg = func()
            neg = tensor.to_numpy(neg)

            numerical_grad = np.sum((pos - neg) * df) / (2 * h)
            #print((autograds[idx] - numerical_grad)/numerical_grad)
            # threshold set as -5% to +5%
            #self.assertAlmostEqual((autograds[idx] - numerical_grad)/(numerical_grad+0.0000001), 0., places=1)
            self.assertAlmostEqual(autograds[idx] - numerical_grad,
                                   0.,
                                   places=2)

            it.iternext()

    def _vanillaRNN_gpu_tiny_ops_shape_check_helper(self, dev):
        # gradients shape check.