3.3 keras模型構建的三種方式

阿新 • • 發佈：2020-12-15

技術標籤：視覺

3.3 `keras`模型構建的三種方式

1. 使用`tf.keras.Sequential`按層順序構建模型，程式碼示例：

model = Sequential()

#卷積層conv_1_1
model.add(Cov2D(input_shape = (64, 64, 3），filters = 32, 
			kernel_size = 3, activation = 'relu', kernel_initializer = 'he_uniform', name = 'conv_1_1)

#卷積層 conv_1_2
model.add(Cov2D(filters = 32 
, kernel_size = 3, 
			activation = 'relu', kernel_initializer = 'he_uniform', name = 'conv_1_2)

#池化層max_pool_1
model.add(MaxPool2D(pool_size = 32, name = 'max_pool_1))

#展平層
model.add(Flatten(name = 'flatten'))

#全連線層
model.add(Dense(unit = 6, activation = 'softmax', name = 'logit'))

#設定損失函式loss、優化器optimizer、評價指標metrics 

model.compile(loss="categorical_crossentropy", 
			  optimizer = tf.keras.optimizers.SGD(learning_rate = 0.001), 
			  metrics = ["accuracy"])

或者：

model = Sequential([
	Conv2D(input_shape = (64, 64, 3), filters = 32,
			kernel_size = 3, activation = 'relu', name = 'conv_1_1'),
	Conv2D( 
filters = 32, kernel_size = 3, activation = 'relu', name = 'conv_1_2'),
	MaxPool2D(poo_size = 2, anme = 'max_pool_1'),
	Flatten(name = 'flatten'),
	Dense(units = 6, activation = "softmax", name = 'logit')])

#設定損失函式loss、優化器optimizer、評價標準metrics
model.compile(loss="categorical_crossentropy", 
			  optimizer = tf.keras.optimizers.SGD(learning_rate = 0.001), 
			  metrics = ["accuracy"])

適用場合：對於順序結構的模型（沒有多個輸入輸出，也沒有分支），優先使用Sequential方法構建。

缺點：不能建立以下模型結構

共享層
模型分支
多個輸入分支
多個輸出分支

2. `Keras`函式式API建立模型，程式碼示例：

#輸入層input
input = input(shape = (64, 64, 3), name = 'input')

#卷積層conv_1_2
x = Conv2D(filters = 32, kernel_size = 3, activation = 'relu', name = 'conv_1_1')(input)

#卷積層con_1_2
x = Conv2D(filters = 32, kernel_size = 3, activation = 'relu', name = 'conv_1_2)(x)

#池化層max_pool_1
x = MaxPool2D(pool_size = 2, name = 'max_pool_1)(x)

#展平層
x = Flatten(name = 'flatten')(x)

#全連線層
output = Dense(units = 6, activation = "softmax", name = 'logit')(x)

model = Model(inputs = input, outputs = output)

#設定損失函式loss、優化器optimizer、評價標準metrics
model.compile(loss="categorical_crossentropy", 
			  optimizer = tf.keras.optimizers.SGD(learning_rate = 0.001), 
			  metrics = ["accuracy"])

適用場合：如果模型有多輸入或者多輸出，或者模型需要共享權重，或者模型具有分支連線、迴圈連線等非順序結構，推薦使用函式式API進行建立。

3. `Keras Model Subclassing`方式，程式碼示例：

#定義一個子類來搭建模型
class ConvModel(Model):
	def __init__(self):
		#父類初始化
		super(ConvModel, self).__init__()

		#卷積層conv_1_1
		self.conv_1_1 = Conv2D(input_shape = (64, 64, 3),
							  filters = 32, kernel_size = 3, activation = 'relu', name = 'con_1_1')
		
		#卷積層conv_1_2
		self.conv_1_2 = Conv2D(filters = 32, kernel_size = 3,
								activation = 'relu', name = 'conv_1_2')
		
		#池化層max_pool_1
		self.max_pool_1 = MaxPool2D(pool.size = 2, name = 'max_pool_1')

		#展平層flatten
		self.dense =  Dense(units = 6, activation = "softmax", name = 'logit')

	def call(selfm, x):
		x = self.conv_1_1(x)
		x = self.conv_1_2(x)
		x = self.max_pool_1(x)
		x = self.conv_2_1(x)
		x = self.conv_2_2(x)
		x = self.max_pool_2(x)
		x = self.flatten(x)
		x =  self.dense(x)
		return x

#類例項化
model = ConvModel()