[Python] Functional API

Functional API

---

• Sequential API는 여러층을 공유하거나 다양한 종류의 입력과 출력을 사용하는 등의 복잡한 모델을 만드는 일에는 한계가 있어더욱 복잡한 모델을 생성할 수 있는 방식인 Functional API(함수형 API)를 사용.
• 각 레이어를 함수로서 정의.
• 반드시 입력층을 정의.

 

 

 

 

 

▶ 기본 활용문

from tensorflow.keras.layers import Input, Dense, concatenate
from tensorflow.keras.models import Model, load_model

inputs = Input(shape=(10,)) # 입력층
hidden1 = Dense(64, activation='relu')(inputs) # 히든레이어1을 입력층과 연결
hidden2 = Dense(64, activation='relu')(hidden1) # 히든레이어2를 히든레이어1과 연결
output = Dense(1, activation='sigmoid')(hidden2) # 출력층을 히든레이어2와 연결
model = Model(inputs=inputs, outputs=output) # 입력층과 출력층을 저장

• Input() 함수에 입력의 크기를 정의합니다.
• 이전층을 다음층 함수의 입력으로 사용하고, 변수에 할당합니다.
• Model() 함수에 입력과 출력을 정의합니다.

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
model.fit(X, y, epochs=5000, verbose=0)
model.predict(X)

 

 

 

 

 

▶ 모델 구현

inputs_1 = Input(shape=(2,))
hidden1_1 = Dense(4, activation='elu')(inputs_1)
output_1 = Dense(2, activation='sigmoid')(hidden1_1)
model_1 = Model(inputs=inputs_1, outputs=output_1)

model_1.summary()
# Model: "model_9"
# _________________________________________________________________
#  Layer (type)                Output Shape              Param #   
# =================================================================
#  input_10 (InputLayer)       [(None, 2)]               0         
                                                                 
#  dense_21 (Dense)            (None, 4)                 12        
                                                                 
#  dense_22 (Dense)            (None, 2)                 10        
                                                                 
# =================================================================
# Total params: 22
# Trainable params: 22
# Non-trainable params: 0
# _________________________________________________________________

inputs_2 = Input(shape=(4,))
hidden2_1 = Dense(4, activation='elu')(inputs_2)
hidden2_2 = Dense(2, activation='elu')(hidden2_1)
output_2 = Dense(2, activation='sigmoid')(hidden2_2)
model_2 = Model(inputs=inputs_2, outputs=output_2)

model_2.summary()
# Model: "model_10"
# _________________________________________________________________
#  Layer (type)                Output Shape              Param #   
# =================================================================
#  input_11 (InputLayer)       [(None, 4)]               0         
                                                                 
#  dense_23 (Dense)            (None, 4)                 20        
                                                                 
#  dense_24 (Dense)            (None, 2)                 10        
                                                                 
#  dense_25 (Dense)            (None, 2)                 6         
                                                                 
# =================================================================
# Total params: 36
# Trainable params: 36
# Non-trainable params: 0
# _________________________________________________________________

 

 

▷ 다중 입력 모델(model that accepts multiple inputs)

inputs_1 = Input(shape=(2,))
hidden1_1 = Dense(4, activation='elu')(inputs_1)
output_1 = Dense(2, activation='sigmoid')(hidden1_1)
model_1 = Model(inputs=inputs_1, outputs=output_1)

inputs_2 = Input(shape=(4,))
hidden2_1 = Dense(4, activation='elu')(inputs_2)
hidden2_2 = Dense(2, activation='elu')(hidden2_1)
output_2 = Dense(2, activation='sigmoid')(hidden2_2)
model_2 = Model(inputs=inputs_2, outputs=output_2)

# 두개의 인공 신경망의 출력을 연결(concatenate)
result = concatenate([model_1.output, model_2.output])

z = Dense(2, activation="relu")(result)
z = Dense(1, activation="linear")(z)

model_A = Model(inputs=[model_1.input, model_2.input], outputs=z)

 

 

 

 

 

▶ save_weigths, load_weights

model.save_weights('파일경로')

model.load_weights('파일경로')

 

 

 

 

 

 

▶ plot_model

!pip install pydot

from tensorflow.keras.utils import plot_model

plot_model(model_2)

plot_model(model_A)

 

 

 

 

 

 

 

 

▶ xavier initialization 

• 2015년까지 사용된 일반적인 딥러닝 프레임 초기화 기법.
• 활성화 함수가 선형인 것을 전제로 이끈 결과.
• simoid, tanh 함수는 좌우 대칭임으로 중앙 부근이 선형함수라고 볼 수 있고 Xavier 초깃값이 적당하다.
• ReLU를 사용시 He 초깃값 사용을 권장한다.

 

w = np.random.randn(node_num, node_num) / np.sqrt(node_ num)

 

▷ no xavier

xor = {'x1':[0,0,1,1], 'x2':[0,1,0,1], 'y':[0,1,1,0]}
XOR = pd.DataFrame(xor)
X = XOR.drop('y', axis=1)
y = XOR.y

ip = Input(shape=(2,))a
n = Dense(2, activation='sigmoid')(ip)
n = Dense(1, activation='linear')(n)
model = Model(inputs=ip, outputs=n)

model.compile(loss='mse', optimizer='rmsprop')
model.fit(X, y, epochs=1400, verbose=0)
model.predict(X)

 

 

▷ with xavier initialization

xor = {'x1':[0,0,1,1], 'x2':[0,1,0,1], 'y':[0,1,1,0]}
XOR = pd.DataFrame(xor)
X = XOR.drop('y', axis=1)
y = XOR.y

ip = Input(shape=(2,))
n = Dense(2, activation='sigmoid')(ip)
n = Dense(1, activation='linear')(n)
model = Model(inputs=ip, outputs=n)

weights = model.get_weights()
xavier_w12 = np.random.randn(2, 2) / np.sqrt(2)
xavier_w21 = np.random.randn(2, 1) / np.sqrt(2)
xavier_weights = weights
xavier_weights[0] = xavier_w12
xavier_weights[2] = xavier_w21
model.set_weights(xavier_weights)


model.compile(loss='mse', optimizer='rmsprop')
model.fit(X, y, epochs=1400, verbose=0)
model.predict(X)

 

 

 

 

 

▷ with xavier initialization & tanh

xor = {'x1':[0,0,1,1], 'x2':[0,1,0,1], 'y':[0,1,1,0]}
XOR = pd.DataFrame(xor)
X = XOR.drop('y', axis=1)
y = XOR.y

ip = Input(shape=(2,))
n = Dense(2, activation='tanh')(ip)
n = Dense(1, activation='linear')(n)
model = Model(inputs=ip, outputs=n)

weights = model.get_weights()
xavier_w12 = np.random.randn(2, 2) / np.sqrt(2)
xavier_w21 = np.random.randn(2, 1) / np.sqrt(2)
xavier_weights = weights
xavier_weights[0] = xavier_w12
xavier_weights[2] = xavier_w21
model.set_weights(xavier_weights)


model.compile(loss='mse', optimizer='rmsprop')
model.fit(X, y, epochs=1400, verbose=0)model.predict(X)

 

 

 

▷ with xavier initialization & relu

xor = {'x1':[0,0,1,1], 'x2':[0,1,0,1], 'y':[0,1,1,0]}
XOR = pd.DataFrame(xor)
X = XOR.drop('y', axis=1)
y = XOR.y

ip = Input(shape=(2,))
n = Dense(2, activation='relu')(ip)
n = Dense(1, activation='linear')(n)
model = Model(inputs=ip, outputs=n)

weights = model.get_weights()
xavier_w12 = np.random.randn(2, 2) / np.sqrt(2)
xavier_w21 = np.random.randn(2, 1) / np.sqrt(2)
xavier_weights = weights
xavier_weights[0] = xavier_w12
xavier_weights[2] = xavier_w21
model.set_weights(xavier_weights)


model.compile(loss='mse', optimizer='rmsprop')
model.fit(X, y, epochs=1400, verbose=0)
model.predict(X)

 

 

▷ with xavier initialization & selu

xor = {'x1':[0,0,1,1], 'x2':[0,1,0,1], 'y':[0,1,1,0]}
XOR = pd.DataFrame(xor)
X = XOR.drop('y', axis=1)
y = XOR.y

ip = Input(shape=(2,))
n = Dense(2, activation='selu')(ip)
n = Dense(1, activation='linear')(n)
model = Model(inputs=ip, outputs=n)

weights = model.get_weights()
xavier_w12 = np.random.randn(2, 2) / np.sqrt(2)
xavier_w21 = np.random.randn(2, 1) / np.sqrt(2)
xavier_weights = weights
xavier_weights[0] = xavier_w12
xavier_weights[2] = xavier_w21
model.set_weights(xavier_weights)


model.compile(loss='mse', optimizer='rmsprop')
model.fit(X, y, epochs=1400, verbose=0)
model.predict(X)