解决one of the variables needed for gradient computation has been modified by an inplace operation

一、项目背景

HiDDen的jpeg()的训练，要求的torch版本是1.0，我的是1.11，考虑过降低版本号，但如果要降版本的话还要改python版本、cuda版本，而且刚开始复现代码，希望能提高自己读代码和纠正错误的能力，所以决定就在这个基础上直接改。

    def train_on_batch(self, batch: list):
        """
        Trains the network on a single batch consisting of images and messages
        :param batch: batch of training data, in the form [images, messages]
        :return: dictionary of error metrics from Encoder, Decoder, and Discriminator on the current batch
        """
        images, messages = batch

        batch_size = images.shape[0]
        self.encoder_decoder.train()
        self.discriminator.train()
        with (torch.enable_grad()):
            # ---------------- Train the discriminator -----------------------------
            self.optimizer_discrim.zero_grad()
            # train on cover
            d_target_label_cover = torch.full((batch_size, 1), self.cover_label, device=self.device).float()
            d_target_label_encoded = torch.full((batch_size, 1), self.encoded_label, device=self.device).float()
            g_target_label_encoded = torch.full((batch_size, 1), self.cover_label, device=self.device).float()

            d_on_cover = self.discriminator(images)
            d_loss_on_cover = self.bce_with_logits_loss(d_on_cover, d_target_label_cover)
            d_loss_on_cover.backward()

            # train on fake
            encoded_images, noised_images, decoded_messages = self.encoder_decoder(images, messages)
            d_on_encoded = self.discriminator(encoded_images.detach())
            d_loss_on_encoded = self.bce_with_logits_loss(d_on_encoded, d_target_label_encoded)

            d_loss_on_encoded.backward()
            self.optimizer_discrim.step()

            # --------------Train the generator (encoder-decoder) ---------------------
            self.optimizer_enc_dec.zero_grad()
            # target label for encoded images should be 'cover', because we want to fool the discriminator
            d_on_encoded_for_enc = self.discriminator(encoded_images)
            g_loss_adv = self.bce_with_logits_loss(d_on_encoded_for_enc, g_target_label_encoded)

            if self.vgg_loss == None:
                g_loss_enc = self.mse_loss(encoded_images, images)
            else:
                vgg_on_cov = self.vgg_loss(images)
                vgg_on_enc = self.vgg_loss(encoded_images)
                g_loss_enc = self.mse_loss(vgg_on_cov, vgg_on_enc)

            g_loss_dec = self.mse_loss(decoded_messages, messages)
            g_loss = (self.config.adversarial_loss * g_loss_adv
                      + self.config.encoder_loss * g_loss_enc
                      + self.config.decoder_loss * g_loss_dec)

            g_loss.backward()
            self.optimizer_enc_dec.step()

        decoded_rounded = decoded_messages.detach().cpu().numpy().round().clip(0, 1)
        bitwise_avg_err = np.sum(np.abs(decoded_rounded - messages.detach().cpu().numpy())) / (
                batch_size * messages.shape[1])

        losses = {
            'loss           ': g_loss.item(),
            'encoder_mse    ': g_loss_enc.item(),
            'dec_mse        ': g_loss_dec.item(),
            'bitwise-error  ': bitwise_avg_err,
            'adversarial_bce': g_loss_adv.item(),
            'discr_cover_bce': d_loss_on_cover.item(),
            'discr_encod_bce': d_loss_on_encoded.item()
        }
        return losses, (encoded_images, noised_images, decoded_messages)

二、报错内容

在运行的时候出现了如下报错：

Traceback (most recent call last):
  
  File "K:\python_project\test_HiDDeN\HiDDeN\main.py", line 150, in <module>
    main()
  File "K:\python_project\test_HiDDeN\HiDDeN\main.py", line 146, in main
    train(model, device, hidden_config, train_options, this_run_folder, tb_logger)
  File "K:\python_project\test_HiDDeN\HiDDeN\train.py", line 52, in train
    losses, _ = model.train_on_batch([image, message])
  File "K:\python_project\test_HiDDeN\HiDDeN\model\hidden.py", line 101, in train_on_batch
    g_loss.backward()
  File "K:\python_project\test_HiDDeN\HiDDeN\venv\lib\site-packages\torch\_tensor.py", line 363, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "K:\python_project\test_HiDDeN\HiDDeN\venv\lib\site-packages\torch\autograd\__init__.py", line 173, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
    
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: 
[torch.cuda.FloatTensor [48, 1, 128, 128]], which is output 0 of AsStridedBackward0, is at version 3; expected version 2 instead. 
Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

报错说：在g_loss.backward()【梯度计算】的时候，所需的一个变量已被就地操作修改。
给的建议是：使用torch.autograd.set_detect_anomaly(True)启用异常检测，以查找未能计算其梯度的操作。

三、解决：

3.1 跟着报错建议：

在代码报错行附近加了 torch.autograd.set_detect_anomaly(True) ，但是运行之后还是指到了原来的位置，并没有进一笔指出异常的地方，看起来这个方法不适合我的情况。

3.2 找教程参考：

1. 参考1 [未解决]

看到问题描述是一样的，没多想就套用了，但我还是报错。
回头想想我连人家因为什么报错都不知道，这么鲁莽，肯定会出问题。

后面他又说检查是否有inplace=True的设置，检查了一下确定没有，所以放弃了。

2. 参考2 [未解决]

提到了retain_graph参数：是否在计算完成后释放空间。说的是在存在多个要求导对象的情况。虽然最后没有用到，但是觉得这篇博客之后会派上用场的。

3. 参考3 [未解决]

把更新梯度的步骤调后放在一起

尝试了按照他的格式改代码，但是改了改还是没用，同样会报错。

4. 自己摸索 [ 已解决 ]

知道学艺不精，甚至还没入门呢，打算学李沐老师的动手机器学习熟悉熟悉操作，反而被里面一句if isinstance(updater, torch.optim.Optimizer) 给绕住了。

还是打算回头看看。这时候像是突然长了脑子一样，很有灵感：
他说一个求梯度过程中一个必要的变量出错了，出错又在g_loss，那肯定是g_loss的一个参数有问题。找到了g_loss：

g_loss = (self.config.adversarial_loss * g_loss_adv
		 + self.config.encoder_loss * g_loss_enc
         + self.config.decoder_loss * g_loss_dec)       

一个一个试，把变量换成常数，结果发现把g_loss_dec换成常数就可以了，说明它在计算过程中发生了变化。追溯到函数本身。

g_loss_dec = self.mse_loss(decoded_messages, messages)

一开始没直接想到，但是就是觉得可以试试detach操作一下，以为g_loss_adv和g_loss_enc的函数会导致g_loss_dec函数参数发生变化，就给他们加了.detach() ，但是没用。又单独给decoded_messages 和 messages加了.detach() ，还是没用。过了一中午，想着要不给g_loss_dec加个.detach() 试试看？本来都不抱希望，觉得不可能就这么简单，结果真的跑出来了！

所以解决办法就是：

1. 找到报错行 [ 求g_loss.backward()]
2. 找g_loss中哪个变量出错了 [ g_loss_dec ]
3. 对出错变量进行排查，既然说是被原地修改了，那就尝试在函数中用该变量的.detach() 版本，而不直接使用变量本身 [ 加上g_loss_dec.detach() 操作]

修改后的代码片段如下：

            self.optimizer_enc_dec.zero_grad()
            # target label for encoded images should be 'cover', because we want to fool the discriminator
            d_on_encoded_for_enc = self.discriminator(encoded_images)
            g_loss_adv = self.bce_with_logits_loss(d_on_encoded_for_enc, g_target_label_encoded)

            if self.vgg_loss == None:
                g_loss_enc = self.mse_loss(encoded_images, images)
            else:
                vgg_on_cov = self.vgg_loss(images)
                vgg_on_enc = self.vgg_loss(encoded_images)
                g_loss_enc = self.mse_loss(vgg_on_cov, vgg_on_enc)

            g_loss_dec = self.mse_loss(decoded_messages, messages)
            g_loss_dec_detach = g_loss_dec.detach()
            g_loss = (self.config.adversarial_loss * g_loss_adv
                      + self.config.encoder_loss * g_loss_enc
                      + self.config.decoder_loss * g_loss_dec_detach)

            g_loss.backward()
            self.optimizer_enc_dec.step()