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.amax(Tensor([3, 38028357, 4, 5],"float32"), axis=-1, keepdim=True, )
input_tensor = (numpy.random.random([3, 38028357, 4, 5]) - 0.5).astype("float32")
paddle_x, torch_x = init_input(input_tensor)
paddle_out = paddle.amax(paddle_x, axis=-1, keepdim=True)
torch_out = torch.amax(torch_x, dim=-1, keepdim=True)
print(paddle_out.shape)
print(torch_out.shape)
numpy_tensor = (numpy.random.random([3, 38028357, 4, 1]) - 0.5).astype("float32")
paddle_grad, torch_grad = init_input(numpy_tensor)
torch_x_grad = torch.autograd.grad([torch_out], [torch_x], grad_outputs=torch_grad)
paddle_x_grad = paddle.grad([paddle_out], [paddle_x], grad_outputs=paddle_grad, allow_unused=True)
p = paddle_x_grad[0].numpy()
t = torch_x_grad[0].cpu().detach().numpy()
for i in range(3):
for j in range(38028357):
for k in range(4):
for m in range(5):
if p[i][j][k][m] != t[i][j][k][m]:
print("i = {}".format(i), "j = {}".format(j), "k = {}".format(k), "m = {}".format(m), "paddle = {}".format(p[i][j][k][m]), "torch = {}".format(t[i][j][k][m]), "numpy_tensor = {}".format(numpy_tensor[i][j][k][0]), "paddle_out = {}".format(paddle_out[i][j][k][0]), "torch_out = {}".format(torch_out[i][j][k][0]))
print("input_tensor = ", "{}".format(input_tensor[i][j][k][0]), "{}".format(input_tensor[i][j][k][1]), "{}".format(input_tensor[i][j][k][2]), "{}".format(input_tensor[i][j][k][3]), "{}".format(input_tensor[i][j][k][4]))
print("p = ", "{}".format(p[i][j][k][0]), "{}".format(p[i][j][k][1]), "{}".format(p[i][j][k][2]), "{}".format(p[i][j][k][3]), "{}".format(p[i][j][k][4]))
print("t = ", "{}".format(t[i][j][k][0]), "{}".format(t[i][j][k][1]), "{}".format(t[i][j][k][2]), "{}".format(t[i][j][k][3]), "{}".format(t[i][j][k][4]))
numpy.testing.assert_allclose(
paddle_out.numpy(),
torch_out.cpu().detach().numpy(),
1e-2,
1e-2,
err_msg='output diff'
)
numpy.testing.assert_allclose(
paddle_x_grad[0].numpy(),
torch_x_grad[0].cpu().detach().numpy(),
1e-2,
1e-2,
err_msg='output diff'
)