Hey, I am working to convert a torch work to paddle. I need to use paddle.autograd.grad to update gradients against some variables. However, it is not working while it is working for another group of vectors.

The code snippet is as following,

for i in range(option['g_l_steps']):
                gen_res = model(images, z, depths)
                g_log_lkhd = 1.0 / (2.0 * option['g_llhd_sigma'] * option['g_llhd_sigma']) * mse_loss(
                    F.sigmoid(gen_res[0]), gts)
                en = ebm_model(z)
                #z_grad_g = g_log_lkhd.grad[0]
                z_grad_g = g_log_lkhd.grad(z)
                z_grad_g = paddle.autograd.grad(g_log_lkhd, z)[0]
                z_grad_e = paddle.autograd.grad(en.sum(), z)[0]

                z.data = z.data - 0.5 * option['g_l_step_size'] * option['g_l_step_size'] * (
                            z_grad_g + z_grad_e + 1.0 / (option['e_prior_sig'] * option['e_prior_sig']) * z.data)
                z.data += option['g_l_step_size'] * fluid.layers.gaussian_random_batch_size_like(z, z.shape).data

May you help？
Thanks,