Question

我正在尝试在VAE中输出kl损失。我的代码基于以下链接中的代码：https://colab.research.google.com/github/tensorflow/probability/blob/master/tensorflow_probability/examples/jupyter_notebooks/Probabilistic_Layers_VAE.ipynb

他们使用

tfpl.MultivariateNormalTriL(
        encoded_size,
        activity_regularizer=tfpl.KLDivergenceRegularizer(prior, weight=1.0)),
])

为模型损失添加KL散度正则化器。我更改了解码器的输出，现在有几个输出，所有输出都来自编码器的输出（作为解码器的输入），带有一些密集的lstm层，然后像下面那样编译模型

vae.compile(optimizer=optimizer,
                                loss=['categorical_crossentropy', 'categorical_crossentropy',
                                      'sparse_categorical_crossentropy',
                                      'sparse_categorical_crossentropy',
                                      'sparse_categorical_crossentropy'
                                      ],
                                metrics=[[keras.metrics.CategoricalAccuracy()],
                                         [keras.metrics.CategoricalAccuracy()],
                                         [keras.metrics.SparseCategoricalAccuracy()],
                                         [keras.metrics.SparseCategoricalAccuracy()],
                                         [keras.metrics.SparseCategoricalAccuracy()],
                                         ],
                                )

vae.metrics_names是

[['loss',
 'decoder_loss',
 'decoder_1_loss',
 'decoder_2_loss',
 'decoder_3_loss',
 'decoder_4_loss',
 'decoder_categorical_accuracy',
 'decoder_1_categorical_accuracy',
 'decoder_2_sparse_categorical_accuracy',
 'decoder_3_sparse_categorical_accuracy',
 'decoder_4_sparse_categorical_accuracy']]

编译后。而encoder.losses是

[<tf.Tensor 'encoder_output/ActivityRegularizer/truediv:0' shape=() dtype=float32>,
 <tf.Tensor 'encoder_output/ActivityRegularizer/truediv_1:0' shape=() dtype=float32>]

解码器损耗为[]。

vae.losses是

[<tf.Tensor 'encoder_output/ActivityRegularizer/truediv:0' shape=() dtype=float32>,
 <tf.Tensor 'encoder_output/ActivityRegularizer/truediv_1:0' shape=() dtype=float32>,
 <tf.Tensor 'encoder/encoder_output/ActivityRegularizer/truediv:0' shape=() dtype=float32>,
 <tf.Tensor 'encoder/encoder_output/ActivityRegularizer/truediv_1:0' shape=() dtype=float32>,
 <tf.Tensor 'encoder_1/encoder_output/ActivityRegularizer/truediv:0' shape=() dtype=float32>,
 <tf.Tensor 'encoder_1/encoder_output/ActivityRegularizer/truediv_1:0' shape=() dtype=float32>]

vae.metrics是

[<tensorflow.python.keras.metrics.CategoricalAccuracy at 0x7f475d0209e8>,
 <tensorflow.python.keras.metrics.CategoricalAccuracy at 0x7f475d020b70>,
 <tensorflow.python.keras.metrics.SparseCategoricalAccuracy at 0x7f475d020d68>,
 <tensorflow.python.keras.metrics.SparseCategoricalAccuracy at 0x7f475cd17cf8>,
 <tensorflow.python.keras.metrics.SparseCategoricalAccuracy at 0x7f475cd17da0>]

我想记录所有5个交叉熵损失以及kl损失。如何在此处记录KL损失？为什么vae.losss有6个活动调节器损失，而编码器只有2个？这里是第一个损失输出的kl损失吗？

在制作模型之后并在编译之前，我还可以更改KL损失权重吗？我想尝试一些减肥减肥法。

如何打印调节器损耗并更改调节器重量？

0 个答案: