无法修复: RuntimeError:梯度计算所需的变量之一已被内部操作修改

Question

目前，我正在尝试复制一个DeblurGanV2网络。目前，我正在进行训练。以下是我目前的培训计划：

from torch.autograd import Variable
torch.autograd.set_detect_anomaly(mode=True)
total_generator_loss = 0
total_discriminator_loss = 0
psnr_score = 0.0
used_loss_function = 'wgan_gp_loss'
for epoch in range(n_epochs):

      #set to train mode
      generator.train(); discriminator.train()
      tqdm_bar = tqdm(train_loader, desc=f'Training Epoch {epoch} ', total=int(len(train_loader)))
      for batch_idx, imgs in enumerate(tqdm_bar):
        
        #load imgs to cpu
        blurred_images = imgs["blurred"].cuda()
        sharped_images = imgs["sharp"].cuda()
        
        # generator output
        deblurred_img = generator(blurred_images)
    
        # denormalize
        with torch.no_grad():
          denormalized_blurred = denormalize(blurred_images)
          denormalized_sharp = denormalize(sharped_images)
          denormalized_deblurred = denormalize(deblurred_img)
    
        # get D's output
        sharp_discriminator_out = discriminator(sharped_images)
        deblurred_discriminator_out = discriminator(deblurred_img)
    
        # set critic_updates
        if used_loss_function== 'wgan_gp_loss':
          critic_updates = 5
        else:
            critic_updates = 1
    
        #train discriminator
        discriminator_loss = 0
        for i in range(critic_updates):
          discriminator_optimizer.zero_grad()
          # train discriminator on real and fake
          if used_loss_function== 'wgan_gp_loss':
            gp_lambda = 10
            alpha = random.random()
            interpolates = alpha * sharped_images + (1 - alpha) * deblurred_img
            interpolates_discriminator_out = discriminator(interpolates)
            kwargs = {'gp_lambda': gp_lambda,
                       'interpolates': interpolates,
                       'interpolates_discriminator_out': interpolates_discriminator_out,
                       'sharp_discriminator_out': sharp_discriminator_out,
                       'deblurred_discriminator_out': deblurred_discriminator_out
                        }
            wgan_loss_d, gp_d = wgan_gp_loss('D', **kwargs)
            discriminator_loss_per_update = wgan_loss_d + gp_d
    
          discriminator_loss_per_update.backward(retain_graph=True)
          discriminator_optimizer.step()
          discriminator_loss += discriminator_loss_per_update.item()

但是，当我运行此代码时，会收到以下错误消息：

RuntimeError:梯度计算所需的变量之一已被内部操作修改：[torch.cuda.FloatTensor 1,512，4，4]位于版本2；预期的版本1改为。提示:上面的回溯跟踪进一步显示了未能计算其梯度的操作。所讨论的变量在那里或以后的任何地方都进行了更改。祝好运!discriminator_loss_per_update.backward(retain_graph=True)跟踪(最近一次调用) in () 62 # discriminator_loss_per_update = gan_loss_d 63 -> 64 -> 65 discriminator_optimizer.step() 66 discriminator_loss += discriminator_loss_per_update.item()

1帧/usr/local/lib/python3.7/dist-packages/torch/tensor.py在后向(自，梯度，retain_graph，create_graph，输入) 243 create_graph=create_graph，244 inputs=inputs)→245 torch.autograd.backward(自我，梯度，retain_graph，create_graph，inputs=inputs) 246247个def register_hook(自，钩子)：

/usr/local/lib/python3.7/dist-packages/torch/autograd/init.py在后向(张量，grad_tensors，retain_graph，create_graph，grad_variables，输入) 145 Variable.execution_engine.run_backward( 146个张量，grad_tensors，retain_graph，create_graph，inputs，→147 allow_unreachable=True，accumulate_grad=True) # allow_unreachable标志148 149

不幸的是，我无法真正跟踪可能导致此错误的就地操作。有人可能有什么主意或建议给我吗？如果有任何输入:slight_smile，我将不胜感激：

Answer 1

尝试将最后一行更改为：

discriminator_loss = discriminator_loss + discriminator_loss_per_update.item()