[Pytorch1.12] RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation

1

I have tried many methods but nothing seems to work. As a newbie, I really don’t understand how to modify the code to make it work. Did I modify it wrong?

same problem

torch.autograd.set_detect_anomaly(True)
        
        # 1st path update D_xvs
        d_x_loss = x_bar_bar_loss_s - x_loss_s + 10. * gp_loss + v_loss_x
        
        D_xvs_solver.zero_grad()
        
        d_x_loss.backward()
        
        D_xvs_solver.step()
        
        # 2nd path, update G_xvz
        x_bar_bar_loss_v, x_bar_bar_loss_s = D_xvs(x_bar_bar)  # 
        x_bar_bar_loss_s = x_bar_bar_loss_s.mean()
        
        if reconstruct_fake:
            x_l1_loss = L1_loss(x_bar_bar, x_bar)
            v_loss_x_bar_bar = crossEntropyLoss(x_bar_bar_loss_v, vv1)  # ACGAN loss of x_bar_bar(v1)
        else:
            x_l1_loss = L1_loss(x_bar_bar, x2)
            v_loss_x_bar_bar = crossEntropyLoss(x_bar_bar_loss_v, vv2)  # ACGAN loss of x_bar_bar(v2)
        
        # id loss
        v_loss_x = crossEntropyLoss(v_bar, vv1)
        
        if not reconstruct_fake:
            l_g_rec = ID_weight * cri_rec(
                fake_fea, real_fea,
                torch.ones((real_fea.shape[0]), device=device)  # 
            )
        
            symmetry_128_loss = args.symmetry_loss_weight * Sym_loss(x_bar_bar)
        
            g_loss = -x_bar_bar_loss_s + 4 * x_l1_loss + v_loss_x_bar_bar + 0.01 * v_loss_x + l_g_rec + symmetry_128_loss
        else:
            g_loss = -x_bar_bar_loss_s + 4 * x_l1_loss + v_loss_x_bar_bar + 0.01 * v_loss_x

        G_vzx_solver.zero_grad()
        G_xvz_solver.zero_grad()

        g_loss.backward()

        if not reconstruct_fake:
            
            G_vzx_solver.step()
        
        G_xvz_solver.step()