首页 > 解决方案 > Keras:训练性能在完全相同的数据和架构下是不同的。唯一的区别是使用 .Sequential() 或 .Model()

问题描述

下面的模型来自Keras 网站,它的行为完全符合预期。它用 定义keras.models.Sequential()。我想将其转换为定义,keras.models.Model()以使其更灵活以供我将来使用。但是在我转换之后,性能直线下降。

您可以在Keras网站上找到原始模型:

def build_model():
  model = Sequential([
    layers.Dense(64, activation=tf.nn.relu, input_shape=[len(train_dataset.keys())]),
    layers.Dense(64, activation=tf.nn.relu),
    layers.Dense(1)
  ])

  optimizer = keras.optimizers.Adam()
  model.compile(loss='mean_squared_error',
                optimizer=optimizer,
                metrics=['mean_absolute_error', 'mean_squared_error'])
  return model

model = build_model()
model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_22 (Dense)             (None, 64)                640       
_________________________________________________________________
dense_23 (Dense)             (None, 64)                4160      
_________________________________________________________________
dense_24 (Dense)             (None, 1)                 65        
=================================================================
Total params: 4,865
Trainable params: 4,865
Non-trainable params: 0
_________________________________________________________________

以下代码是我的转换:

def build_model_base():
  input = Input(shape=[len(train_dataset.keys())])
  x = Dense(64, activation='relu', name="dense1")(input)   
  x = Dense(64, activation='relu', name="dense2")(x)
  output = Dense(1, activation='sigmoid', name='output')(x)
  model = Model(input=[input], output=[output])
  optimizer = keras.optimizers.Adam()

  model.compile(loss='mean_squared_error', 
                optimizer=optimizer,
                metrics=['mean_absolute_error', 'mean_squared_error'])
  return model

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_18 (InputLayer)        (None, 9)                 0         
_________________________________________________________________
dense1 (Dense)               (None, 64)                640       
_________________________________________________________________
dense2 (Dense)               (None, 64)                4160      
_________________________________________________________________
output (Dense)               (None, 1)                 65        
=================================================================
Total params: 4,865
Trainable params: 4,865
Non-trainable params: 0

我能看到的唯一区别是计数时.Sequential不计数,但我不相信它们会使模型结构不同。但是,性能是:input layer.Model.Sequential

在此处输入图像描述

.Model()我转换的性能是:

在此处输入图像描述

谁能告诉我我做错了什么?

其他一些上下文:

我读过这篇文章,但我的代码都在 Google Colab 的 CPU 上运行

print(keras.__version__) # 2.0.4

print(tf.__version__) #1.14.0-rc1

绘制损失的代码:

def plot_history(history):
  hist = pd.DataFrame(history.history)
  hist['epoch'] = history.epoch

  plt.figure()
  plt.xlabel('Epoch')
  plt.ylabel('Mean Abs Error [MPG]')
  plt.plot(hist['epoch'], hist['mean_absolute_error'],
           label='Train Error')
  plt.plot(hist['epoch'], hist['val_mean_absolute_error'],
           label = 'Val Error')
  y_max = max(hist['val_mean_absolute_error'])
  plt.ylim([0,y_max])
  plt.legend()

  plt.figure()
  plt.xlabel('Epoch')
  plt.ylabel('Mean Square Error [$MPG^2$]')
  plt.plot(hist['epoch'], hist['mean_squared_error'],
           label='Train Error')
  plt.plot(hist['epoch'], hist['val_mean_squared_error'],
           label = 'Val Error')
  y_max = max(hist['val_mean_squared_error'])
  plt.ylim([0,y_max])
  plt.legend()
  plt.show() 

训练模型的代码(两种模型完全相同):

his_seq = model.fit(normed_train_data.values, train_labels.values,
          batch_size=128,
          validation_split = 0.1,
          epochs = 100,
          verbose=0)
plot_history(his_seq)

任何建议表示赞赏!

标签: pythontensorflowmachine-learningkerasdeep-learning

解决方案


在 keras 的密集层以及您构建的顺序模型的输出层中默认使用“线性”激活函数。

但是您在转换中将激活函数指定为“sigmoid”,这可能会有所不同。

下面是 Keras 提供的默认激活函数的说明:

activation: Activation function to use (see activations). If you don't specify anything, no activation is applied (ie. "linear" activation: a(x) = x).

推荐阅读