可以压缩和分类的自动编码器 - 2个问题

Question

我正在尝试训练一个自动编码器（共享权重），它也能够对MNIST数字进行分类。

我设置了两个输出和两个损失函数。

tied_ae_model = keras.Model(inputs=inputs, outputs=[classification_neurons, outputs], name="mnist")
tied_ae_model.compile(loss=["mean_squared_error", "binary_crossentropy"], optimizer="adam", metrics=["accuracy"])

我还需要不同的拟合函数的Y值。虽然看起来对Y有用，但看起来对验证数据没用。到底要如何传递验证数据呢，？

history = tied_ae_model.fit(x_train_norm, [y_train, x_train_norm], epochs=2, 
validation_data=[(x_train_norm, y_train), (x_train_norm, x_train_norm)])

Epoch 1/5
1875/1875 [==============================] - 4s 2ms/step - loss: 0.1978 - classification_loss: 0.0183 - reconstructions_loss: 0.1795 - classification_accuracy: 0.8828 - reconstructions_accuracy: 0.8032
val_loss: 0.0000e+00 - val_classification_loss: 0.0000e+00 - val_reconstructions_loss: 0.0000e+00 - val_classification_accuracy: 0.0000e+00 - val_reconstructions_accuracy: 0.0000e+00
Epoch 2/5
1875/1875 [==============================] - 4s 2ms/step - loss: 0.1310 - classification_loss: 0.0083 - reconstructions_loss: 0.1227 - classification_accuracy: 0.9485 - reconstructions_accuracy: 0.8106
val_loss: 0.0000e+00 - val_classification_loss: 0.0000e+00 - val_reconstructions_loss: 0.0000e+00 - val_classification_accuracy: 0.0000e+00 - val_reconstructions_accuracy: 0.0000e+00

我也很奇怪，为什么解码器的分数不能超过8182？即使我设置隐藏神经元==输入大小。

代码片段。

# Custom layer to tie the weights of the encoder and decoder
class DenseTranspose(keras.layers.Layer):
    def __init__(self, dense, activation=None, **kwargs):
        self.dense = dense
        self.activation = keras.activations.get(activation)
        super().__init__(**kwargs)

    def build(self, batch_input_shape):
        self.biases = self.add_weight(name="bias", shape=[self.dense.input_shape[-1]], initializer="zeros")
        super().build(batch_input_shape)

    def call(self, inputs):
        z = tf.matmul(inputs, self.dense.weights[0], transpose_b=True)
        return self.activation(z + self.biases)


# Model: Stacked Autoencoder with tied (shared) weights.
# Encoder
inputs = keras.Input(shape=(28,28,1))
x = layers.Flatten()(inputs)
dense = layers.Dense(200, activation="sigmoid", name='encoder')
x = dense(x)

# Latent space: contains classification neurons and latent space neurons
classification_neurons = layers.Dense(10, activation="softmax", name='classification')(x)
latent_space_neurons = layers.Dense(190, activation="sigmoid", name='latent_space')(x)
# Merge classification neurons and latent space neurons into a single vector via concatenation
x = layers.concatenate([classification_neurons, latent_space_neurons])

# Decoder
x = DenseTranspose(dense, activation="sigmoid", name="decoder")(x)
outputs = layers.Reshape([28,28,1], name='reconstructions')(x)

# Build & Show the model
tied_ae_model = keras.Model(inputs=inputs, outputs=[classification_neurons, outputs], name="mnist")
tied_ae_model.summary()
keras.utils.plot_model(tied_ae_model, "my_first_model_with_shape_info.png", show_shapes=True)

# Compile & Train
tied_ae_model.compile(loss=["mean_squared_error", "binary_crossentropy"], optimizer="adam", metrics=["accuracy"])
history = tied_ae_model.fit(x_train_norm, [y_train, x_train_norm], epochs=2, validation_data=[(x_train_norm, y_train), (x_train_norm, x_train_norm)])