6-1,构建模型的3种方法

可以使用以下3种方式构建模型：使用Sequential按层顺序构建模型，使用函数式API构建任意结构模型，继承Model基类构建自定义模型。

对于顺序结构的模型，优先使用Sequential方法构建。

如果模型有多输入或者多输出，或者模型需要共享权重，或者模型具有残差连接等非顺序结构，推荐使用函数式API进行创建。

如果无特定必要，尽可能避免使用Model子类化的方式构建模型，这种方式提供了极大的灵活性，但也有更大的概率出错。

下面以IMDB电影评论的分类问题为例，演示3种创建模型的方法。

  1. import numpy as np 
  2. import pandas as pd 
  3. import tensorflow as tf
  4. from tqdm import tqdm 
  5. from tensorflow.keras import *
  8. train_token_path = "./data/imdb/train_token.csv"
  9. test_token_path = "./data/imdb/test_token.csv"
  11. MAX_WORDS = 10000  # We will only consider the top 10,000 words in the dataset
  12. MAX_LEN = 200  # We will cut reviews after 200 words
  13. BATCH_SIZE = 20 
  15. # 构建管道
  16. def parse_line(line):
  17.     t = tf.strings.split(line,"\t")
  18.     label = tf.reshape(tf.cast(tf.strings.to_number(t[0]),tf.int32),(-1,))
  19.     features = tf.cast(tf.strings.to_number(tf.strings.split(t[1]," ")),tf.int32)
  20.     return (features,label)
  22. ds_train=  tf.data.TextLineDataset(filenames = [train_token_path]) \
  23.    .map(parse_line,num_parallel_calls = tf.data.experimental.AUTOTUNE) \
  24.    .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
  25.    .prefetch(tf.data.experimental.AUTOTUNE)
  27. ds_test=  tf.data.TextLineDataset(filenames = [test_token_path]) \
  28.    .map(parse_line,num_parallel_calls = tf.data.experimental.AUTOTUNE) \
  29.    .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
  30.    .prefetch(tf.data.experimental.AUTOTUNE)

一，Sequential按层顺序创建模型

  1. tf.keras.backend.clear_session()
  3. model = models.Sequential()
  5. model.add(layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN))
  6. model.add(layers.Conv1D(filters = 64,kernel_size = 5,activation = "relu"))
  7. model.add(layers.MaxPool1D(2))
  8. model.add(layers.Conv1D(filters = 32,kernel_size = 3,activation = "relu"))
  9. model.add(layers.MaxPool1D(2))
  10. model.add(layers.Flatten())
  11. model.add(layers.Dense(1,activation = "sigmoid"))
  13. model.compile(optimizer='Nadam',
  14.             loss='binary_crossentropy',
  15.             metrics=['accuracy',"AUC"])
  17. model.summary()

6-1,构建模型的3种方法 - 图1

  1. import datetime
  2. baselogger = callbacks.BaseLogger(stateful_metrics=["AUC"])
  3. logdir = "./data/keras_model/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
  4. tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)
  5. history = model.fit(ds_train,validation_data = ds_test,
  6.         epochs = 6,callbacks=[baselogger,tensorboard_callback])
  1. %matplotlib inline
  2. %config InlineBackend.figure_format = 'svg'
  4. import matplotlib.pyplot as plt
  6. def plot_metric(history, metric):
  7.     train_metrics = history.history[metric]
  8.     val_metrics = history.history['val_'+metric]
  9.     epochs = range(1, len(train_metrics) + 1)
  10.     plt.plot(epochs, train_metrics, 'bo--')
  11.     plt.plot(epochs, val_metrics, 'ro-')
  12.     plt.title('Training and validation '+ metric)
  13.     plt.xlabel("Epochs")
  14.     plt.ylabel(metric)
  15.     plt.legend(["train_"+metric, 'val_'+metric])
  16.     plt.show()
  1. plot_metric(history,"AUC")

6-1,构建模型的3种方法 - 图2

二，函数式API创建任意结构模型

  1. tf.keras.backend.clear_session()
  3. inputs = layers.Input(shape=[MAX_LEN])
  4. x  = layers.Embedding(MAX_WORDS,7)(inputs)
  6. branch1 = layers.SeparableConv1D(64,3,activation="relu")(x)
  7. branch1 = layers.MaxPool1D(3)(branch1)
  8. branch1 = layers.SeparableConv1D(32,3,activation="relu")(branch1)
  9. branch1 = layers.GlobalMaxPool1D()(branch1)
  11. branch2 = layers.SeparableConv1D(64,5,activation="relu")(x)
  12. branch2 = layers.MaxPool1D(5)(branch2)
  13. branch2 = layers.SeparableConv1D(32,5,activation="relu")(branch2)
  14. branch2 = layers.GlobalMaxPool1D()(branch2)
  16. branch3 = layers.SeparableConv1D(64,7,activation="relu")(x)
  17. branch3 = layers.MaxPool1D(7)(branch3)
  18. branch3 = layers.SeparableConv1D(32,7,activation="relu")(branch3)
  19. branch3 = layers.GlobalMaxPool1D()(branch3)
  21. concat = layers.Concatenate()([branch1,branch2,branch3])
  22. outputs = layers.Dense(1,activation = "sigmoid")(concat)
  24. model = models.Model(inputs = inputs,outputs = outputs)
  26. model.compile(optimizer='Nadam',
  27.             loss='binary_crossentropy',
  28.             metrics=['accuracy',"AUC"])
  30. model.summary()
  1. Model: "model"
  2. __________________________________________________________________________________________________
  3. Layer (type)                    Output Shape         Param #     Connected to                     
  4. ==================================================================================================
  5. input_1 (InputLayer)            [(None, 200)]        0                                            
  6. __________________________________________________________________________________________________
  7. embedding (Embedding)           (None, 200, 7)       70000       input_1[0][0]                    
  8. __________________________________________________________________________________________________
  9. separable_conv1d (SeparableConv (None, 198, 64)      533         embedding[0][0]                  
  10. __________________________________________________________________________________________________
  11. separable_conv1d_2 (SeparableCo (None, 196, 64)      547         embedding[0][0]                  
  12. __________________________________________________________________________________________________
  13. separable_conv1d_4 (SeparableCo (None, 194, 64)      561         embedding[0][0]                  
  14. __________________________________________________________________________________________________
  15. max_pooling1d (MaxPooling1D)    (None, 66, 64)       0           separable_conv1d[0][0]           
  16. __________________________________________________________________________________________________
  17. max_pooling1d_1 (MaxPooling1D)  (None, 39, 64)       0           separable_conv1d_2[0][0]         
  18. __________________________________________________________________________________________________
  19. max_pooling1d_2 (MaxPooling1D)  (None, 27, 64)       0           separable_conv1d_4[0][0]         
  20. __________________________________________________________________________________________________
  21. separable_conv1d_1 (SeparableCo (None, 64, 32)       2272        max_pooling1d[0][0]              
  22. __________________________________________________________________________________________________
  23. separable_conv1d_3 (SeparableCo (None, 35, 32)       2400        max_pooling1d_1[0][0]            
  24. __________________________________________________________________________________________________
  25. separable_conv1d_5 (SeparableCo (None, 21, 32)       2528        max_pooling1d_2[0][0]            
  26. __________________________________________________________________________________________________
  27. global_max_pooling1d (GlobalMax (None, 32)           0           separable_conv1d_1[0][0]         
  28. __________________________________________________________________________________________________
  29. global_max_pooling1d_1 (GlobalM (None, 32)           0           separable_conv1d_3[0][0]         
  30. __________________________________________________________________________________________________
  31. global_max_pooling1d_2 (GlobalM (None, 32)           0           separable_conv1d_5[0][0]         
  32. __________________________________________________________________________________________________
  33. concatenate (Concatenate)       (None, 96)           0           global_max_pooling1d[0][0]       
  34.                                                                  global_max_pooling1d_1[0][0]     
  35.                                                                  global_max_pooling1d_2[0][0]     
  36. __________________________________________________________________________________________________
  37. dense (Dense)                   (None, 1)            97          concatenate[0][0]                
  38. ==================================================================================================
  39. Total params: 78,938
  40. Trainable params: 78,938
  41. Non-trainable params: 0
  42. __________________________________________________________________________________________________

6-1,构建模型的3种方法 - 图3

  1. import datetime
  2. logdir = "./data/keras_model/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
  3. tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)
  4. history = model.fit(ds_train,validation_data = ds_test,epochs = 6,callbacks=[tensorboard_callback])
  1. Epoch 1/6
  2. 1000/1000 [==============================] - 32s 32ms/step - loss: 0.5527 - accuracy: 0.6758 - AUC: 0.7731 - val_loss: 0.3646 - val_accuracy: 0.8426 - val_AUC: 0.9192
  3. Epoch 2/6
  4. 1000/1000 [==============================] - 24s 24ms/step - loss: 0.3024 - accuracy: 0.8737 - AUC: 0.9444 - val_loss: 0.3281 - val_accuracy: 0.8644 - val_AUC: 0.9350
  5. Epoch 3/6
  6. 1000/1000 [==============================] - 24s 24ms/step - loss: 0.2158 - accuracy: 0.9159 - AUC: 0.9715 - val_loss: 0.3461 - val_accuracy: 0.8666 - val_AUC: 0.9363
  7. Epoch 4/6
  8. 1000/1000 [==============================] - 24s 24ms/step - loss: 0.1492 - accuracy: 0.9464 - AUC: 0.9859 - val_loss: 0.4017 - val_accuracy: 0.8568 - val_AUC: 0.9311
  9. Epoch 5/6
  10. 1000/1000 [==============================] - 24s 24ms/step - loss: 0.0944 - accuracy: 0.9696 - AUC: 0.9939 - val_loss: 0.4998 - val_accuracy: 0.8550 - val_AUC: 0.9233
  11. Epoch 6/6
  12. 1000/1000 [==============================] - 26s 26ms/step - loss: 0.0526 - accuracy: 0.9865 - AUC: 0.9977 - val_loss: 0.6463 - val_accuracy: 0.8462 - val_AUC: 0.9138
  1. plot_metric(history,"AUC")

6-1,构建模型的3种方法 - 图4

三，Model子类化创建自定义模型

  1. # 先自定义一个残差模块，为自定义Layer
  3. class ResBlock(layers.Layer):
  4.     def __init__(self, kernel_size, **kwargs):
  5.         super(ResBlock, self).__init__(**kwargs)
  6.         self.kernel_size = kernel_size
  8.     def build(self,input_shape):
  9.         self.conv1 = layers.Conv1D(filters=64,kernel_size=self.kernel_size,
  10.                                    activation = "relu",padding="same")
  11.         self.conv2 = layers.Conv1D(filters=32,kernel_size=self.kernel_size,
  12.                                    activation = "relu",padding="same")
  13.         self.conv3 = layers.Conv1D(filters=input_shape[-1],
  14.                                    kernel_size=self.kernel_size,activation = "relu",padding="same")
  15.         self.maxpool = layers.MaxPool1D(2)
  16.         super(ResBlock,self).build(input_shape) # 相当于设置self.built = True
  18.     def call(self, inputs):
  19.         x = self.conv1(inputs)
  20.         x = self.conv2(x)
  21.         x = self.conv3(x)
  22.         x = layers.Add()([inputs,x])
  23.         x = self.maxpool(x)
  24.         return x
  26.     #如果要让自定义的Layer通过Functional API 组合成模型时可以序列化，需要自定义get_config方法。
  27.     def get_config(self):  
  28.         config = super(ResBlock, self).get_config()
  29.         config.update({'kernel_size': self.kernel_size})
  30.         return config
  1. # 测试ResBlock
  2. resblock = ResBlock(kernel_size = 3)
  3. resblock.build(input_shape = (None,200,7))
  4. resblock.compute_output_shape(input_shape=(None,200,7))
  1. TensorShape([None, 100, 7])
  1. # 自定义模型，实际上也可以使用Sequential或者FunctionalAPI
  3. class ImdbModel(models.Model):
  4.     def __init__(self):
  5.         super(ImdbModel, self).__init__()
  7.     def build(self,input_shape):
  8.         self.embedding = layers.Embedding(MAX_WORDS,7)
  9.         self.block1 = ResBlock(7)
  10.         self.block2 = ResBlock(5)
  11.         self.dense = layers.Dense(1,activation = "sigmoid")
  12.         super(ImdbModel,self).build(input_shape)
  14.     def call(self, x):
  15.         x = self.embedding(x)
  16.         x = self.block1(x)
  17.         x = self.block2(x)
  18.         x = layers.Flatten()(x)
  19.         x = self.dense(x)
  20.         return(x)
  1. tf.keras.backend.clear_session()
  3. model = ImdbModel()
  4. model.build(input_shape =(None,200))
  5. model.summary()
  7. model.compile(optimizer='Nadam',
  8.             loss='binary_crossentropy',
  9.             metrics=['accuracy',"AUC"])
  1. Model: "imdb_model"
  2. _________________________________________________________________
  3. Layer (type)                 Output Shape              Param #   
  4. =================================================================
  5. embedding (Embedding)        multiple                  70000     
  6. _________________________________________________________________
  7. res_block (ResBlock)         multiple                  19143     
  8. _________________________________________________________________
  9. res_block_1 (ResBlock)       multiple                  13703     
  10. _________________________________________________________________
  11. dense (Dense)                multiple                  351       
  12. =================================================================
  13. Total params: 103,197
  14. Trainable params: 103,197
  15. Non-trainable params: 0
  16. _________________________________________________________________

6-1,构建模型的3种方法 - 图5

  1. import datetime
  3. logdir = "./tflogs/keras_model/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
  4. tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)
  5. history = model.fit(ds_train,validation_data = ds_test,
  6.                     epochs = 6,callbacks=[tensorboard_callback])
  1. Epoch 1/6
  2. 1000/1000 [==============================] - 47s 47ms/step - loss: 0.5629 - accuracy: 0.6618 - AUC: 0.7548 - val_loss: 0.3422 - val_accuracy: 0.8510 - val_AUC: 0.9286
  3. Epoch 2/6
  4. 1000/1000 [==============================] - 43s 43ms/step - loss: 0.2648 - accuracy: 0.8903 - AUC: 0.9576 - val_loss: 0.3276 - val_accuracy: 0.8650 - val_AUC: 0.9410
  5. Epoch 3/6
  6. 1000/1000 [==============================] - 42s 42ms/step - loss: 0.1573 - accuracy: 0.9439 - AUC: 0.9846 - val_loss: 0.3861 - val_accuracy: 0.8682 - val_AUC: 0.9390
  7. Epoch 4/6
  8. 1000/1000 [==============================] - 42s 42ms/step - loss: 0.0849 - accuracy: 0.9706 - AUC: 0.9950 - val_loss: 0.5324 - val_accuracy: 0.8616 - val_AUC: 0.9292
  9. Epoch 5/6
  10. 1000/1000 [==============================] - 43s 43ms/step - loss: 0.0393 - accuracy: 0.9876 - AUC: 0.9986 - val_loss: 0.7693 - val_accuracy: 0.8566 - val_AUC: 0.9132
  11. Epoch 6/6
  12. 1000/1000 [==============================] - 44s 44ms/step - loss: 0.0222 - accuracy: 0.9926 - AUC: 0.9994 - val_loss: 0.9328 - val_accuracy: 0.8584 - val_AUC: 0.9052
  1. plot_metric(history,"AUC")

6-1,构建模型的3种方法 - 图6

如果对本书内容理解上有需要进一步和作者交流的地方，欢迎在公众号”Python与算法之美”下留言。作者时间和精力有限，会酌情予以回复。

也可以在公众号后台回复关键字：加群，加入读者交流群和大家讨论。

image.png