1. 實現簡單的邏輯迴歸

from keras import Input
from keras import layers
from keras.models import Model
x = Input(shape=(32, ))
	# TensorShape([Dimension(None), Dimension(32)])
y = layers.Dense(16, activation='softmax')(x)
model = Model(x, y) 
	# model.output_shape
	# model.summary()

2. 函數語言程式設計的模型構建

• 序列化模型構建

  seq_model = Sequential()
  seq_model.add(layers.Dense(32, activation='relu', input_shape=(64, )))
  	# 首層必須指定 input_shape 或 batch_input_shape
  seq_model.add(layers.Dense(32, activation='relu'))
  seq_model.add(layers.Dense(10, activation='softmax'))
  

• 函式式API 下的模型構建

  input_tensor = Input(shape=(64, ))
  z = layers.Dense(32, activation='relu')(input_tensor)
  z = layers.Dense(32, activation='relu')(z)
  y = layers.Dense(10, activation='softmax')(z)
  from keras.models import Model
  model = Model(input_tensor, y)
   
  

• 兩種方式對比：

  seq_model.summary()
  model.summary()
  

3. layers.concatenate 與 layers.add

• layers.add：執行 tensor 的相加操作，要求輸入必須同維度（或者經過 broadcast 之後是同維度）；

  x_1 = Input(shape=(32, ))
  x_2 = Input(shape=(32, ))
  >> layers.add([x_1, x_2])
  <tf.Tensor 'add_3/add:0' shape=(?, 32) dtype=float32>
  
  x_1 = Input(shape=(32, 64))
  x_2 = Input(shape=(64,))
  >> layers.add([x_1, x_2])
  <tf.Tensor 'add_4/add:0' shape=(?, 32, 64) dtype=float32>
   
  

• layers.concatente：執行的是 tensor 的拼接操作（維度會拉長）

  x_1 = Input(shape=(32, ))
  x_2 = Input(shape=(64, ))
  >> layers.concatenate([x_1, x_2])
  <tf.Tensor 'concatenate_3/concat:0' shape=(?, 96) dtype=float32>