Tensorflow 和 Keras 中的多重损失

Question

在我的简单变分自动编码器代码中，我想在模型运行ning 时查看重建和 KL 散度损失值。我使用了 Keras 文档页面中的 Example of VAE on MNIST dataset using MLP，并将损失部分修改如下：

reconstruction_loss = original_dim*binary_crossentropy(inputs,outputs)                                              
kl_loss = -0.5*K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
vae_loss = K.mean(reconstruction_loss + kl_loss)



losses = {  'recon_loss': reconstruction_loss,
             'latent_loss': kl_loss,
             'total_loss': vae_loss,}


vae.add_loss(losses)
vae.compile(optimizer='adam')
history = vae.fit(x_train, epochs=epochs, batch_size=batch_size,
                validation_data=(x_test, None))

但是当我运行模型时，编译命令returns这个错误信息：

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-21-b668ab0f8437> in <module>
----> 1 vae.compile(optimizer='adam')
      2 

~\Anaconda3\lib\site-packages\keras\engine\training.py in compile(self, optimizer, loss, metrics, loss_weights, sample_weight_mode, weighted_metrics, target_tensors, **kwargs)
    349                     total_loss += loss_weight * output_loss
    350             if total_loss is None:
--> 351                 if not self.losses:
    352                     raise ValueError('The model cannot be compiled '
    353                                      'because it has no loss to optimize.')

~\Anaconda3\lib\site-packages\keras\engine\network.py in losses(self)
    423 
    424         unique_tensors = list(
--> 425             set(x for x in losses if not isinstance(x, (float, int))))
    426         non_tensors = [x for x in losses if isinstance(x, (float, int))]
    427         return unique_tensors + non_tensors

TypeError: unhashable type: 'dict'

我看到了一些关于 TypeError: unhashable type: 'dict' 的其他问题，但我无法解决我的问题问题。有没有想法在 Keras 中采用 Compile 命令并处理多个损失？

Answer 1

请在您的损失字典中检查每项损失的类型。

print (type(losses['recon_loss']))

Answer 2

我修改了损失函数并在编译函数中使用了指标。

def recon_loss(inputs,outputs):
    reconstruction_loss = original_dim*binary_crossentropy(inputs,outputs)                                              
    return K.mean(reconstruction_loss)

def latent_loss(inputs,outputs):
    kl_loss =  -0.5*K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)    
    return K.mean(kl_loss)

def total_loss(inputs,outputs):
    reconstruction_loss = original_dim*binary_crossentropy(inputs,outputs)                                              
    kl_loss =  -0.5*K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)   
    return K.mean(reconstruction_loss + kl_loss)
vae.compile(optimizer='adam',loss=total_loss,metrics=[recon_loss, latent_loss])

现在，模型 returns 训练和验证数据集的重建、潜在损失和总损失。

Tensorflow 和 Keras 中的多重损失

Multiple losses in Tensorflow and Keras

python

keras

tensorflow

loss-function