RuntimeError: one of the variables needed for gradient computation has been modified by an inplace

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [2, 1218]], which is output 0 of TanhBackward0, is at version 1; expected version 0 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

我出现这个错误是，通过一个变量计算两种损失时，计算完第一种损失，对这个变量的值进行修改再计算第二种损失时就会出现这个错误。出现这个错误还有很多种情况，其他帖子有。

        generate_fake_data = G(z_final)  # 通过噪声生成序列。G（优化Z）
        recons_loss_1 = torch.mean(
            torch.sum(torch.pow(want_label - generate_fake_data, 2), dim=-1))  # 长码误差1

        generate_fake_data_2 = generate_fake_data  # 修改变量
        generate_fake_data_2 = generate_fake_data_2.view(2, 58, 21)
        index = torch.tensor([0, 1, 2, 3, 4, 5, 7, 11, 14, 15, 17, 18, 19, 20, 21, 22,
                              24, 25, 28, 29, 31, 32, 33, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46,
                              47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57]).cuda()
        generate_fake_data_2.index_fill_(1, index, 0)
        generate_fake_data_2 = generate_fake_data_2.view(2, 1218)  # 2*1218

        recons_loss_2 = torch.mean(
            torch.sum(torch.pow(want_label - generate_fake_data_2, 2), dim=-1))  # 长码误差2
        all_loss = recons_loss_1 + recons_loss_2
        all_loss.backward()
        G_optimizer.step()

报错！

上面的代码只是示意。我用生成的变量 generate_fake_data，计算了第一种损失 recons_loss_1后，我将变量修改了，再计算 recons_loss_2.这时候就会报错。

解决办法：深拷贝这个变量。

generate_fake_data_2 = generate_fake_data.clone()

        generate_fake_data = G(z_final)  # 通过噪声生成序列。G（优化Z）
        recons_loss_1 = torch.mean(
            torch.sum(torch.pow(want_label - generate_fake_data, 2), dim=-1))  # 长码误差1

        generate_fake_data_2 = generate_fake_data.clone()  #修改变量
        generate_fake_data_2 = generate_fake_data_2.view(2, 58, 21)
        index = torch.tensor([0, 1, 2, 3, 4, 5, 7, 11, 14, 15, 17, 18, 19, 20, 21, 22,
                              24, 25, 28, 29, 31, 32, 33, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46,
                              47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57]).cuda()
        generate_fake_data_2.index_fill_(1, index, 0)
        generate_fake_data_2 = generate_fake_data_2.view(2, 1218)  # 2*1218

        recons_loss_2 = torch.mean(
            torch.sum(torch.pow(want_label - generate_fake_data_2, 2), dim=-1))  # 长码误差2
        all_loss = recons_loss_1 + recons_loss_2
        all_loss.backward()
        G_optimizer.step()

正常运行！

好像涉及了计算图的东西，不太懂。个人理解，欢迎指正！