小编yom*_*bob的帖子

我正在尝试使用 Keras 实现回归序列模型，但得到了非常奇怪的结果。我的代码如下。我使用 tf.data 数据集作为输入数据集。我的损失最初下降，但随后开始波动。发生这种情况是有原因的吗？

    #initialize model
    model = keras.models.Sequential()
    
    model.add(Dense(1+(len(feature_names)-1), input_shape=((len(feature_names)-1),)))
    model.add(LeakyReLU())
    model.add(Dropout(0.5))
    
    model.add(Dense(10))
    model.add(LeakyReLU())
    
    #output layer, 1 unit
    model.add(Dense(1))
    
    loss = 'mean_squared_error'
    optimizer = tf.keras.optimizers.Nadam(learning_rate=0.0001)
    
    #compile the model
    m = tf.keras.metrics.RootMeanSquaredError()
    model.compile(optimizer=optimizer, loss=loss, metrics=[m,'mae'])
    
#    early_stop = keras.callbacks.EarlyStopping(monitor='val_loss', patience=4)
    tb = TensorBoard(histogram_freq=1)
    
    ##try learning rates, different noise, dropout, using only nasadem
    
    #run the model
    history = model.fit(x=final_train_dataset,epochs=count,steps_per_epoch=steps_per_epoch,validation_data=final_valid_dataset,callbacks=[tb])

5468/5468 [==============================] - 19s 4ms/step - loss: 19.8461 - root_mean_squared_error: 4.4549 - mae: 3.1814 - val_loss: 13.2963 - val_root_mean_squared_error: 3.6464 - …