import torch
import paddle
import numpy

device = torch.device("cuda:0")
torch.set_default_device(device)

def init_input(numpy_tensor):
    paddle_x = paddle.to_tensor(numpy_tensor)
    torch_x = torch.tensor(numpy_tensor, requires_grad=True)
    paddle_x.stop_gradient = False

    numpy.testing.assert_allclose(
        paddle_x.numpy(),
        torch_x.cpu().detach().numpy(),
        1e-10,
        1e-10,
        err_msg='intput diff'
    )
    return paddle_x, torch_x



# paddle.diff(Tensor([2, 760567127],"float32"), n=1, axis=0, prepend=Tensor([3, 760567127],"float32"), append=None, )
input_tensor = (numpy.ones([1, 1, 7, 7])).astype("float32") 
paddle_input, torch_input = init_input(input_tensor)
paddle_out = paddle.nn.functional.adaptive_avg_pool2d(paddle_input, output_size=(2, 5,), data_format="NCHW", name=None, )
torch_out = torch.nn.functional.adaptive_avg_pool2d(torch_input, output_size=(2, 5,))

print("input:", paddle_input[0][0])
print("torch shape:", torch_out.shape)
print("paddle shape:", paddle_out.shape)

print("paddle out:", paddle_out[0][0])
print("torch out:", torch_out[0][0])

numpy.testing.assert_allclose(
    torch_out.cpu().detach().numpy(),
    paddle_out.numpy(),
    1e-2,
    1e-2,
    err_msg='output diff'
)

numpy_tensor = (numpy.ones([1, 1, 2, 5])).astype("float32")
paddle_grad, torch_grad = init_input(numpy_tensor)
torch_x_grad = torch.autograd.grad([torch_out], [torch_input], grad_outputs=torch_grad)
paddle_x_grad = paddle.grad([paddle_out], [paddle_input], grad_outputs=paddle_grad, allow_unused=True)

print(torch_x_grad[0], torch.sum(torch_x_grad[0]))
print(paddle_x_grad[0], paddle.sum(paddle_x_grad[0]))
numpy.testing.assert_allclose(
    torch_x_grad[0].cpu().detach().numpy(),
    paddle_x_grad[0].numpy(),
    1e-2,
    1e-2,
    err_msg='output diff'
)