fix gradient(nan) when two inputs are equal

[zhangting2020]

PR types

Bug fixes

PR changes

OPs

Describe

fix gradient(nan) when two inputs are equal

test code

mport torch
import paddle
import numpy as np

x_i = np.random.random((4, 4)).astype("float32")

x = torch.tensor(x_i, requires_grad=True)
out = torch.dist(x, x, 2)
out.backward(gradient=torch.tensor(1.0), retain_graph=True)
print(x.grad)

x = paddle.to_tensor(x_i)
x.stop_gradient = False
out = paddle.dist(x, x, 2)
dx = paddle.grad(out, x)
print(dx)

Before

• torch

tensor([[0., 0., 0., 0.],
        [0., 0., 0., 0.],
        [0., 0., 0., 0.],
        [0., 0., 0., 0.]])

• paddle

[Tensor(shape=[4, 4], dtype=float32, place=CUDAPlace(0), stop_gradient=True,
       [[nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan]])]

After

paddle

[Tensor(shape=[4, 4], dtype=float32, place=CUDAPlace(0), stop_gradient=True,
       [[0., 0., 0., 0.],
        [0., 0., 0., 0.],
        [0., 0., 0., 0.],
        [0., 0., 0., 0.]])]

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[Xreki]

LGTM