DAY63. Tensorflow LinearRegression (3)keras dnn

Linear Regression 신경망

 

 

Keras DNN model(Linear Regression 신경망)

Hidden layer : 2개(뉴런 2개)

model = Sequential()

#hidden layer1 : shape = [2, 2]

model.add(Dense(2, input_shape=(2,), activation = 'relu')) # 1층(hidden1)

#hidden layer2 : shape = [2, 2]

model.add(Dense(2, activation = 'relu')) # 2층(hidden2)

#output layer : shape = [2, 1]

model.add(Dense(units=1))

#출력층(output)

 

 

 

 

 

keras model

High Level API : Keras - 인간어 수준 언어  
Low Level API : 기계어 수준 언어 

dataset & 모델 평가

from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_squared_error, r2_score

keras model

import tensorflow as tf
from tensorflow.keras import Sequential #keras model
from tensorflow.keras.layers import Dense #DNN layer

1. dataset load

X, y = load_iris(return_X_y = True)
X.shape # (150, 4)
y #0~1

 

2. train_test_split

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size = 0.3, random_state=123)

* 조절변수(w, b) 정의 없음 : model에서 자동 생성



3. DNN model 생성

model = Sequential() #keras model
print(model) #Sequential class에서 만들어진 object

 

 

 

 

 

model에 layer추가
hidden layer 2개 : unit = 12, unit=6

hidden layer1 : unit = 12 -> w[4, 12], b=12

model.add(Dense(units=12, input_shape=(4,), activation='relu')) #1층

hidden layer2 : unit = 6 -> w[12, 6], b=6

model.add(Dense(units=6, activation='relu')) #2층

* 6개의 뉴런에 의해 2층 레이어 생성. 1층에서 받아 온 자료가 input되므로, input_shape=(4,)조건은 필요치 않다.

output layer : unit = 1 -> w[6, 1], b=1

model.add(Dense(units=1)) #3층

model layer 구조 확인

print(model.summary())

_________________________________________________________________
Layer (type)                 Output Shape              Param -> (in*out)+b   
=================================================================
dense (Dense)              (None, 12)                60    -> (4*12)+12
_________________________________________________________________
dense_1 (Dense)              (None, 6)                 78    -> (12*6)+6
_________________________________________________________________
dense_2 (Dense)              (None, 1)                 7     -> (6*1)+1
=================================================================
Total params: 145
Trainable params: 145
Non-trainable params: 0
_________________________________________________________________



4. model compile : 학습과정 설정

model.compile(optimizer='adam', loss='mse', metrics=['mae'])

* 어느정도 정형화 되어있는 학습과정
optimizer : 최적화 알고리즘 ('adam' or 'sgd')
loss : 손실함수(mes : 평균제곱오차)
metrics : 평가 방법 ('mae' or 'mse')



5. model training : model 학습

model.fit(x=X_train, y=y_train, #훈련셋
          epochs=100, #반복학습횟수 (사용자가 조절가능)
          verbose=1, #학습결과 출력여부 Y
          validation_data=(X_test, y_test)) #검증셋

Epoch 1/100
loss: 16.4911 - mae: 3.7187 - val_loss: 14.8984 - val_mae: 3.4226
:
Epoch 100/100
loss: 0.1761 - mae: 0.3791 - val_loss: 0.1872 - val_mae: 0.3978

* val_ 은 검증결과

 

 

 

 

 

keras_history

keras history : model 학습과정과 검증과정의 손실값 등을 기억하는 기능 

 

dataset & 모델 평가 

from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_squared_error, r2_score

#keras model 
import tensorflow as tf 
from tensorflow.keras import Sequential #keras model 
from tensorflow.keras.layers import Dense #DNN layer 

#keras seed값 지정 
import numpy as np 
import random as rd

 

keras 내부 가중치 seed 적용 

tf.random.set_seed(123) #global seed
np.random.seed(123) #numpy seed
rd.seed(123) #random seed

1. dataset load 

X, y = load_iris(return_X_y=True)
X.shape  #(150, 4)
y #0~2

2. train_test_split : 훈련셋 vs 검증셋 

X_train, X_val, y_train, y_val = train_test_split(
    X, y, test_size=0.3, random_state=123)

조절변수(w, b) 정의 없음 : model에서 자동 생성 

 

3. DNN model 생성 

model = Sequential() #keras model 
print(model) #Sequential object

hidden layer 2개 : unit=12, unit=6
hidden layer1 : unit = 12 -> w[4, 12], b=12

model.add(Dense(units=12, input_shape=(4,), activation='relu')) #1층

hidden layer2 : unit = 6 -> w[12, 6], b=6 

model.add(Dense(units=6, activation='relu')) #2층

output layer : unit = 1 -> w[6, 1], b=1

model.add(Dense(units=1)) #3층

model layer 확인 

print(model.summary())

_________________________________________________________________
Layer (type)                 Output Shape              Param #(in*out)+b   
=================================================================
dense (Dense)                (None, 12)                60=(4*12)+12        
_________________________________________________________________
dense_1 (Dense)              (None, 6)                 78=(12*6)+6        
_________________________________________________________________
dense_2 (Dense)              (None, 1)                 7=(6*1)+1         
=================================================================
Total params: 145
Trainable params: 145
Non-trainable params: 0
_________________________________________________________________


4. model compile : 학습과정 설정 

model.compile(optimizer='adam', loss='mse', metrics=['mae'])

optimizer : 최적화 알고리즘('adam', 'sgd')
loss : 손실함수(mse) 
metrics : 평가방법('mae', 'mse')

 

[수정] 5. model training : model 학습 

model_fit = model.fit(x=X_train, y=y_train, #훈련셋
          epochs=100, #학습횟수 
          verbose=1, #출력여부 
          validation_data=(X_val, y_val)) #검증셋

6. model evaluation : model 평가 

loss, mae = model.evaluate(X_val, y_val) 
print('loss value =', loss)
print('mae =', mae )

training data : 70% 훈련셋 
validation data : 30% 검증셋 
test data : 실제 업무용 dataset 


[추가] 7. model history

print(model_fit.history.keys()) #key 확인 
#dict_keys(['loss', 'mae', 'val_loss', 'val_mae'])
model_fit.history['loss'] #훈련셋 손실값 
model_fit.history['val_loss'] #검증셋 손실값

loss vs val loss 

import matplotlib.pyplot as plt 
plt.plot(model_fit.history['loss'], 'y', label='training loss value')
plt.plot(model_fit.history['val_loss'], 'r', label='val loss value')
plt.xlabel('epochs')
plt.ylabel('loss value')
plt.legend(loc = 'best')
plt.show()

model 평가 : mae vs val mae 

plt.plot(model_fit.history['mae'], 'y', label='training mae value')
plt.plot(model_fit.history['val_mae'], 'r', label='val mae value')
plt.xlabel('epochs')
plt.ylabel('mae')
plt.legend(loc = 'best')
plt.show()