DAY66. Tensorflow Keras model (2)Overfitting solution

karas mnist history

* keras mnist batch참고 
History : 훈련과 검증과정에서 발생하는 손실값/정확도 결과 기억 기능 

from tensorflow.keras.datasets import mnist #mnist load 
from tensorflow.keras.utils import to_categorical #Y변수 : encoding 
from tensorflow.keras import Sequential #keras model 생성 
from tensorflow.keras.layers import Dense #DNN layer 구축
import matplotlib.pyplot as plt #시각화 도구

keras 내부 w,b변수 seed 적용

import tensorflow as tf
import numpy as np 
import random as rd

tf.random.set_seed(123)
np.random.seed(123)
rd.seed(123)
import time #학습 소요 시간 측정

1. mnist dataset load 

(x_train, y_train), (x_val, y_val) = mnist.load_data() #(images, labels)

images : X변수 

x_train.shape #(60000, 28, 28) - (size, h, w) : 2d 제공 
x_val.shape #(10000, 28, 28)

x_train[0] #0~255
x_train.max() #255

 

labels : y변수 

y_train.shape #(60000,)
y_train[0] #5

2. X,y변수 전처리 
1) X변수 : 정규화 & reshape(2d -> 1d)

x_train = x_train / 255. #정규화 
x_val = x_val / 255.

x_train[0]

reshape(2d -> 1d)

x_train = x_train.reshape(-1, 784) #(60000, 28*28)
x_val = x_val.reshape(-1, 784) #(10000, 28*28)

2) y변수 : class(10진수) -> one-hot encoding(2진수)

y_train = to_categorical(y_train)
y_val = to_categorical(y_val)

전처리 확인 

x_train.shape #(60000, 784)
y_train[0] #[0., 0., 0., 0., 0., 1., 0., 0., 0., 0.] - 5
y_train.shape #(60000, 10)

start_time = time.time() #소요 시간 체트

3. keras model

model = Sequential()

4. DNN model layer 구축 
hidden layer1 : w[784, 128]

model.add(Dense(units=128, input_shape=(784,), activation='relu')) #1층

hidden layer2 : w[128, 64]

model.add(Dense(units=64, activation='relu')) #2층

hidden layer3 : w[64, 32]

model.add(Dense(units=32, activation='relu')) #3층

output layer : w[32, 10]

model.add(Dense(units=10, activation='softmax')) #4층

model layer 확인 

model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_8 (Dense)              (None, 128)               100480=784x128+128    
_________________________________________________________________
dense_9 (Dense)              (None, 64)                8256      
_________________________________________________________________
dense_10 (Dense)             (None, 32)                2080      
_________________________________________________________________
dense_11 (Dense)             (None, 10)                330       
=================================================================
Total params: 111,146

 

5. model compile : 학습과정 설정(다항분류기) 

model.compile(optimizer='adam', #default : learning_rate=0.001
              loss='categorical_crossentropy', 
              metrics=['accuracy'])

6. [수정] model training : train(70) vs val(30)

model_fit = model.fit(x=x_train, y=y_train, #훈련셋 
          epochs=15, #반복학습 횟수 : 60000 * 10 = 600,000 -> full batch
          batch_size=100, # 1epoch(100 * 600) * 10 = 600,000 -> mini batch 
          verbose=1, #출력여부 
          validation_data= (x_val, y_val)) #검증셋

stop_time = time.time() - start_time 

print('소요시간 : ', stop_time)

full batch 
accuracy: 0.9923 - val_loss: 0.0873 - val_accuracy: 0.9793
소요시간 :  25.037985801696777

mini batch 
accuracy: 0.9933 - val_loss: 0.0916 - val_accuracy: 0.9739
소요시간 :  11.333117961883545



7. model evaluation : val dataset 

print('model evaluation')
model.evaluate(x=x_val, y=y_val)
#loss: 0.0916 - accuracy: 0.9739

 

8. model history 

print(model_fit.history.keys()) #dict_keys(['loss', 'accuracy', 'val_loss', 'val_accuracy'])

loss vs val_loss : overfitting 시작점 : epoch=2

plt.plot(model_fit.history['loss'], 'y', label='train loss')
plt.plot(model_fit.history['val_loss'], 'r', label='val loss')
plt.xlabel('epochs')
plt.ylabel('loss')
plt.legend(loc = 'best')
plt.show()

accuracy vs val_accuracy : overfitting 시작점 : epoch=2

plt.plot(model_fit.history['accuracy'], 'y', label='train accuracy')
plt.plot(model_fit.history['val_accuracy'], 'r', label='val accuracy')
plt.xlabel('epochs')
plt.ylabel('accuracy')
plt.legend(loc = 'best')
plt.show()

 

 

 

 

 

 

karas mnist dropout

Dropout : 무작위 네트워크 삭제 -> 과적화 최소화  
* karas_mnist_history 참고

from tensorflow.keras.datasets import mnist #mnist load 
from tensorflow.keras.utils import to_categorical #Y변수 : encoding 
from tensorflow.keras import Sequential #keras model 생성 
from tensorflow.keras.layers import Dense, Dropout #[추가] DNN layer 구축
import matplotlib.pyplot as plt #시각화 도구

keras 내부 w,b변수 seed 적용 

import tensorflow as tf
import numpy as np 
import random as rd

tf.random.set_seed(123)
np.random.seed(123)
rd.seed(123)
import time #학습 소요 시간 측정

1. mnist dataset load 

(x_train, y_train), (x_val, y_val) = mnist.load_data() #(images, labels)

images : X변수 

x_train.shape #(60000, 28, 28) - (size, h, w) : 2d 제공 
x_val.shape #(10000, 28, 28)

x_train[0] #0~255
x_train.max() #255

labels : y변수 

y_train.shape #(60000,)
y_train[0] #5

 

 

2. X,y변수 전처리 
1) X변수 : 정규화 & reshape(2d -> 1d)

x_train = x_train / 255. #정규화 
x_val = x_val / 255.

x_train[0]

reshape(2d -> 1d)

x_train = x_train.reshape(-1, 784) #(60000, 28*28)
x_val = x_val.reshape(-1, 784) #(10000, 28*28)

2) y변수 : class(10진수) -> one-hot encoding(2진수)

y_train = to_categorical(y_train)
y_val = to_categorical(y_val)

전처리 확인 

x_train.shape # (60000, 784)
y_train[0] #[0., 0., 0., 0., 0., 1., 0., 0., 0., 0.] - 5
y_train.shape #(60000, 10)

start_time = time.time() #소요 시간 체트

3. keras model

model = Sequential()

4. DNN model layer 구축 
hidden layer1 : w[784, 128]

model.add(Dense(units=128, input_shape=(784,), activation='relu'))#1층 
model.add(Dropout(rate=0.3)) #[추가]

hidden layer2 : w[128, 64]

model.add(Dense(units=64, activation='relu')) #2층 
model.add(Dropout(rate=0.1)) #[추가]

hidden layer3 : w[64, 32]

model.add(Dense(units=32, activation='relu')) #3층
model.add(Dropout(rate=0.1)) #[추가]

output layer : w[32, 10]

model.add(Dense(units=10, activation='softmax')) #4층

model layer 확인 

model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_8 (Dense)              (None, 128)               100480=784x128+128    
_________________________________________________________________
dense_9 (Dense)              (None, 64)                8256      
_________________________________________________________________
dense_10 (Dense)             (None, 32)                2080      
_________________________________________________________________
dense_11 (Dense)             (None, 10)                330       
=================================================================
Total params: 111,146

 

 

5. model compile : 학습과정 설정(다항분류기) 

model.compile(optimizer='adam', #default : learning_rate=0.001
              loss='categorical_crossentropy', 
              metrics=['accuracy'])

6. model training : train(70) vs val(30)

model_fit = model.fit(x=x_train, y=y_train, #훈련셋 
          epochs=15, #반복학습 횟수 : 60000 * 10 = 600,000 -> full batch
          batch_size=100, #1epoch(100 * 600) * 10 = 600,000 -> mini batch 
          verbose=1, #출력여부 
          validation_data= (x_val, y_val)) #검증셋

stop_time = time.time() - start_time 

print('소요시간 : ', stop_time)

full batch 
accuracy: 0.9923 - val_loss: 0.0873 - val_accuracy: 0.9793
소요시간 :  25.037985801696777

mini batch 
accuracy: 0.9933 - val_loss: 0.0916 - val_accuracy: 0.9739
소요시간 :  11.333117961883545



7. model evaluation : val dataset 

print('model evaluation')
model.evaluate(x=x_val, y=y_val) #loss: 0.0916 - accuracy: 0.9739

8. model history 

print(model_fit.history.keys()) #dict_keys(['loss', 'accuracy', 'val_loss', 'val_accuracy'])

loss vs val_loss : overfitting 시작점 : epoch=2

plt.plot(model_fit.history['loss'], 'y', label='train loss')
plt.plot(model_fit.history['val_loss'], 'r', label='val loss')
plt.xlabel('epochs')
plt.ylabel('loss')
plt.legend(loc = 'best')
plt.show()

accuracy vs val_accuracy : overfitting 시작점 : epoch=2

plt.plot(model_fit.history['accuracy'], 'y', label='train accuracy')
plt.plot(model_fit.history['val_accuracy'], 'r', label='val accuracy')
plt.xlabel('epochs')
plt.ylabel('accuracy')
plt.legend(loc = 'best')
plt.show()

 

 

 

 

 

karas mnist earlyStopping

* step02_karas_mnist_dropout 참고 

 

1. Dropout : 무작위 네트워크 삭제 -> 과적화 최소화  
2. EarlyStopping 
- val loss value에 변화가 없는 경우 특정 시점에서 학습 조기 종료 
 

from tensorflow.keras.datasets import mnist #mnist load 
from tensorflow.keras.utils import to_categorical # Y변수 : encoding 
from tensorflow.keras import Sequential #keras model 생성 
from tensorflow.keras.layers import Dense, Dropout #DNN layer 구축
from tensorflow.keras.callbacks import EarlyStopping #[추가]

import matplotlib.pyplot as plt #시각화 도구

keras 내부 w,b변수 seed 적용

import tensorflow as tf
import numpy as np 
import random as rd

tf.random.set_seed(123)
np.random.seed(123)
rd.seed(123)
import time #학습 소요 시간 측정

1. mnist dataset load 

(x_train, y_train), (x_val, y_val) = mnist.load_data() #(images, labels)

images : X변수 

x_train.shape #(60000, 28, 28) - (size, h, w) : 2d 제공 
x_val.shape #(10000, 28, 28)

x_train[0] #0~255
x_train.max() #255

labels : y변수 

y_train.shape #(60000,)
y_train[0] #5

2. X,y변수 전처리 
1) X변수 : 정규화 & reshape(2d -> 1d)

x_train = x_train / 255. #정규화 
x_val = x_val / 255.

x_train[0]

reshape(2d -> 1d)

x_train = x_train.reshape(-1, 784) #(60000, 28*28)
x_val = x_val.reshape(-1, 784) #(10000, 28*28)

2) y변수 : class(10진수) -> one-hot encoding(2진수)

y_train = to_categorical(y_train)
y_val = to_categorical(y_val)

전처리 확인 

x_train.shape #(60000, 784)
y_train[0] #[0., 0., 0., 0., 0., 1., 0., 0., 0., 0.] - 5
y_train.shape #(60000, 10)

start_time = time.time() # 요 시간 체트

3. keras model

model = Sequential()

4. DNN model layer 구축 
hidden layer1 : w[784, 128]

model.add(Dense(units=128, input_shape=(784,), activation='relu')) #1층 
model.add(Dropout(rate=0.3))

hidden layer2 : w[128, 64]

model.add(Dense(units=64, activation='relu')) #2층 
model.add(Dropout(rate=0.1))

hidden layer3 : w[64, 32]

model.add(Dense(units=32, activation='relu')) #3층
model.add(Dropout(rate=0.1))

output layer : w[32, 10]

model.add(Dense(units=10, activation='softmax')) #4층

model layer 확인 

model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_8 (Dense)              (None, 128)               100480=784x128+128    
_________________________________________________________________
dense_9 (Dense)              (None, 64)                8256      
_________________________________________________________________
dense_10 (Dense)             (None, 32)                2080      
_________________________________________________________________
dense_11 (Dense)             (None, 10)                330       
=================================================================
Total params: 111,146



5. model compile : 학습과정 설정(다항분류기) 

model.compile(optimizer='adam', #default : learning_rate=0.001
              loss='categorical_crossentropy', 
              metrics=['accuracy'])

6. model training : train(70) vs val(30)
[추가] epoch=5 이후 검증 손실이 개선되지 않으면 조기종료 

callback = EarlyStopping(monitor='val_loss', patience=5)

model_fit = model.fit(x=x_train, y=y_train, #훈련셋 
          epochs=30, #[수정] 반복학습 횟수 
          batch_size=100, # 1epoch(100 * 600) * 10 = 600,000 -> mini batch 
          verbose=1, #출력여부 
          validation_data= (x_val, y_val), # 검증셋
          callbacks = callback) #[추가] 조기종료 

stop_time = time.time() - start_time 

print('소요시간 : ', stop_time)

full batch 
accuracy: 0.9923 - val_loss: 0.0873 - val_accuracy: 0.9793
소요시간 :  25.037985801696777

mini batch 
accuracy: 0.9933 - val_loss: 0.0916 - val_accuracy: 0.9739
소요시간 :  11.333117961883545



7. model evaluation : val dataset 

print('model evaluation')
model.evaluate(x=x_val, y=y_val) #loss: 0.0916 - accuracy: 0.9739

8. model history 

print(model_fit.history.keys()) #dict_keys(['loss', 'accuracy', 'val_loss', 'val_accuracy'])

loss vs val_loss : overfitting 시작점 : epoch=2

plt.plot(model_fit.history['loss'], 'y', label='train loss')
plt.plot(model_fit.history['val_loss'], 'r', label='val loss')
plt.xlabel('epochs')
plt.ylabel('loss')
plt.legend(loc = 'best')
plt.show()

accuracy vs val_accuracy : overfitting 시작점 : epoch=2

plt.plot(model_fit.history['accuracy'], 'y', label='train accuracy')
plt.plot(model_fit.history['val_accuracy'], 'r', label='val accuracy')
plt.xlabel('epochs')
plt.ylabel('accuracy')
plt.legend(loc = 'best')
plt.show()