DAY67. Tensorflow CNN model

mnist cnnbasic

MNIST + CNN basic 
합성곱층 : 특징맵(feature map) - 이미지 특징 추출 
폴링층 : 픽셀 축소(다운샘플링) - 이미지 특징 강조 

import tensorflow as tf 
from tensorflow.keras.datasets import mnist # image dataset 
import matplotlib.pyplot as plt # image 시각화 
import numpy as np

1. image dataset load 

(x_train, y_train), (x_test,y_test) = mnist.load_data()

x_train.shape #(60000, 28, 28) - (size, h, w)

print(x_train[0])

2. 정수형 -> 실수형 변환 

x_train = x_train.astype('float32')

3. 정규화 규정화 

x_train = x_train/255.
x_train[0]

4. first image 특징 추출 -> 특징 강조 

img = x_train[0]

plt.imshow(X=img, cmap='gray')
plt.show()

img.shape # (28, 28)

1) 입력 이미지 : 4차원 모양 변경 

firstImg = img.reshape(1,28,28,1) #[size, h, w, color]

2) 필터(Filter) 정의 : w변수 역할(수정 대상)

Filter = tf.Variable(tf.random.normal(shape=[3,3,1,5])) #[h,w,color,map_size]

3) 합성곱층 : 이미지 특징 추출 

conv2d = tf.nn.conv2d(input=firstImg, filters=Filter, 
                      strides=[1,1,1,1], padding='SAME')

 

 

 

합성곱 연산 결과  시각화 

print(conv2d.shape) #(1, 28, 28, 5) 
conv2d_img = np.swapaxes(conv2d, 0, 3) #(5, 28, 28, 1) - 축교환 

for i, img in enumerate(conv2d_img) : #index, image 넘김 
    plt.subplot(1, 5, i+1) #1행5열 격자 : 1 -> 5 이미지 배열  
    plt.imshow(img.reshape(28, 28), cmap='gray')  
plt.show()

폴링 연산 결과 시각화

print(pool.shape) # (1, 14, 14, 5)
pool_img = np.swapaxes(pool, 0, 3) # (5, 14, 14, 1) - 축교환 

for i, img in enumerate(pool_img) : # index, image 넘김 
    plt.subplot(1,5, i+1) # 1행5열 격자 : 1 -> 5 이미지 배열
    plt.imshow(img.reshape(14, 14), cmap='gray') 
plt.show()

 

 

 

 

 

keras cnn cifar10

Keras CNN model 생성 
1. image dataset load
2. image dataset 전처리 
3. CNN model 생성 
4. CNN model layer 구축 
5. model compile : 학습환경 
6. model training : 학습 
7. model 평가 
8. model history 

from tensorflow.keras.datasets import cifar10 #color image dataset
from tensorflow.keras.utils import to_categorical #y변수 encoding
from tensorflow.keras import Sequential #keras model
from tensorflow.keras.layers import Conv2D, MaxPool2D #Conv layer
from tensorflow.keras.layers import Flatten, Dense, Dropout #DNN layer
import matplotlib.pyplot as plt #image show

1. image dataset load

(x_train, y_train), (x_val, y_val) = cifar10.load_data()

x_train.shape #(50000, 32, 32, 3) - (size, h, w, c)
y_train.shape #(50000, 1)

first_img = x_train[0]
first_img.shape #(32, 32, 3)

plt.imshow(X=first_img)
plt.show()

y_train[0] #[6] - 개구리 

x_val.shape #(10000, 32, 32, 3)

 

2. image dataset 전처리
1) image pixel 실수형 변환 

x_train = x_train.astype(dtype='float32') #image vs filter 
x_val = x_val.astype(dtype='float32')

2) image 정규화 : 0~1

x_train = x_train / 255 
x_val = x_val / 255

3) label 전처리 : class(10진수) -> one hot encoding(2진수)

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

y_train[0] #[0., 0., 0., 0., 0., 0., 1., 0., 0., 0.] - 6

3. CNN model 생성 

model = Sequential()

4. CNN model layer 구축 

input_shape = (32, 32, 3) #입력 image 모양

Conv layer1 : 1층 [28, 28, 32] -> [13, 13, 32]

model.add(Conv2D(filters=32, kernel_size=(5,5),
                 input_shape = input_shape, activation='relu')) #합성곱+활성함수

filters : 특징맵 개수, kernel_size : Filter 크기  

model.add(MaxPool2D(pool_size=(3,3), strides=(2,2))) #폴링 : 픽셀축소

pool_size : 폴링 윈도, strides : 윈도 이동 크기 

model.add(Dropout(rate=0.3)) #30% 무작위 n/w 제거

Conv layer2 : 2층 [9, 9, 64] -> [4, 4, 64]

model.add(Conv2D(filters=64, kernel_size=(5,5),
                 input_shape = input_shape, activation='relu')) #합성곱+활성함수 
model.add(MaxPool2D(pool_size=(3,3), strides=(2,2))) #폴링 : 픽셀축소 
model.add(Dropout(rate=0.1)) #10% 무작위 n/w 제거

Conv layer3 : 3층 [2, 2, 128]

model.add(Conv2D(filters=128, kernel_size=(3,3),
                 input_shape = input_shape, activation='relu')) #합성곱+활성함수 
model.add(Dropout(rate=0.1)) #10% 무작위 n/w 제거

전결합층 : Flatten layer : 3d(h,w,c) -> 1d(h*w*c)

model.add(Flatten())

DNN layer1 : 4층(hidden layer)

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

DNN layer2 : 5층(output layer)

model.add(Dense(units=10, activation='softmax'))          
model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_8 (Conv2D)            (None, 28, 28, 32)        2432      
_________________________________________________________________
max_pooling2d_6 (MaxPooling2 (None, 13, 13, 32)        0         
_________________________________________________________________
conv2d_9 (Conv2D)            (None, 9, 9, 64)          51264     
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 4, 4, 64)          0         
_________________________________________________________________
conv2d_10 (Conv2D)           (None, 2, 2, 128)         73856     
_________________________________________________________________
flatten_1 (Flatten)          (None, 512)               0         
_________________________________________________________________
dense_5 (Dense)              (None, 64)                32832     
_________________________________________________________________
dense_6 (Dense)              (None, 10)                650       
=================================================================



5. model compile : 학습환경(다항분류기)

model.compile(optimizer='adam',
              loss='categorical_crossentropy', #y = one_hot encoding 
              metrics=['accuracy'])

 

               
6. model training : 학습

model_fit = model.fit(x=x_train, y=y_train, #훈련셋 
          epochs=10, #반복학습
          batch_size = 100, #1회공급 image 개수(1epoch=100*500) 
          verbose=1,
          validation_data=(x_val, y_val)) #검증셋

7. model 평가

print('='*30)
model.evaluate(x=x_val, y=y_val)

8. model history 
loss vs val_loss : 과적합 시작점 : 3.5

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

accuracy vs val_accuracy : 과적합 시작점 : 4

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

 

 

 

 

 

real image cnn basic

real image + CNN basic

import tensorflow as tf 
import numpy as np 
import matplotlib.pyplot as plt #image show 
from matplotlib.image import imread #image  read 

path = r'C:\ITWILL\5_Tensorflow\data\images'

1. image load 

img = imread(path + '/parrots.png')
type(img) #numpy.ndarray
plt.imshow(X=img)
plt.show()

RGB 픽셀값 

img.shape #(512, 768, 3) - (h, w, color)

2. image reshape

Img = img.reshape(1, 512, 768, 3) #(size, h, w, color)

3. Filter 정의 

Filter = tf.Variable(tf.random.normal([9,9,3,5])) #[h,w,color,f-map]

 

4. 합성곱층 : 이미지 특징 추출 

conv2d = tf.nn.conv2d(input=Img, filters=Filter, 
                      strides=[1,2,2,1], padding='SAME')

conv2d.shape # [1, 256, 384, 5]

5. 폴링층 : 이미지 특징 강조(픽셀 축소)

pool = tf.nn.max_pool(input=conv2d, ksize=[1,7,7,1], 
               strides=[1,4,4,1], padding='SAME')
pool.shape #[1, 64, 96, 5]

합성곱 연산 결과  시각화 

print(conv2d.shape) #(1, 256, 384, 5)
conv2d_img = np.swapaxes(conv2d, 0, 3) #(5, 28, 28, 1) - 축교환 

for i, img in enumerate(conv2d_img) : #index, image 넘김 
    plt.subplot(1, 5, i+1) #1행5열 격자 : 1 -> 5 이미지 배열  
    plt.imshow(img.reshape(256, 384), cmap='gray')  
plt.show()

폴링 연산 결과 시각화

print(pool.shape) #(1, 64, 96, 5)
pool_img = np.swapaxes(pool, 0, 3) #(5, 14, 14, 1) - 축교환 

for i, img in enumerate(pool_img) : #index, image 넘김 
    plt.subplot(1,5, i+1) #1행5열 격자 : 1 -> 5 이미지 배열
    plt.imshow(img.reshape(64, 96), cmap='gray') 
plt.show()

 

 

 

 

 

keras cnn tensorboard

* keras_cnn_cifar10 참고
- Keras CNN layers tensorboard 시각화 

from tensorflow.keras.datasets import cifar10 #color image dataset
from tensorflow.keras.utils import to_categorical #y변수 encoding
from tensorflow.keras import Sequential #keras model
from tensorflow.keras.layers import Conv2D, MaxPool2D #Conv layer
from tensorflow.keras.layers import Flatten, Dense, Dropout #DNN layer
import matplotlib.pyplot as plt #image show   
import tensorflow as tf #[추가]

[추가] tensorboard 초기화 

tf.keras.backend.clear_session()

1. image dataset load

(x_train, y_train), (x_val, y_val) = cifar10.load_data()

x_train.shape #(50000, 32, 32, 3) - (size, h, w, c)
y_train.shape #(50000, 1)

first_img = x_train[0]
first_img.shape #(32, 32, 3)

plt.imshow(X=first_img)
plt.show()

y_train[0] #[6] - 개구리 

x_val.shape #(10000, 32, 32, 3)

2. image dataset 전처리
1) image pixel 실수형 변환 

x_train = x_train.astype(dtype='float32') #image vs filter 
x_val = x_val.astype(dtype='float32')

2) image 정규화 : 0~1

x_train = x_train / 255 
x_val = x_val / 255

3) label 전처리 : class(10진수) -> one hot encoding(2진수)

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

y_train[0] #[0., 0., 0., 0., 0., 0., 1., 0., 0., 0.] - 6

3. CNN model 생성 

model = Sequential()

4. CNN model layer 구축 

input_shape = (32, 32, 3) # 입력 image 모양

Conv layer1 : 1층 [28, 28, 32] -> [13, 13, 32]

model.add(Conv2D(filters=32, kernel_size=(5,5),
                 input_shape = input_shape, activation='relu')) #합성곱+활성함수

filters : 특징맵 개수, kernel_size : Filter 크기  

model.add(MaxPool2D(pool_size=(3,3), strides=(2,2))) #폴링 : 픽셀축소

pool_size : 폴링 윈도, strides : 윈도 이동 크기 

model.add(Dropout(rate=0.3)) # 30% 무작위 n/w 제거

Conv layer2 : 2층 [9, 9, 64] -> [4, 4, 64]

model.add(Conv2D(filters=64, kernel_size=(5,5),
                 input_shape = input_shape, activation='relu')) #합성곱+활성함수 
model.add(MaxPool2D(pool_size=(3,3), strides=(2,2))) #폴링 : 픽셀축소 
model.add(Dropout(rate=0.1)) #10% 무작위 n/w 제거

Conv layer3 : 3층 [2, 2, 128]

model.add(Conv2D(filters=128, kernel_size=(3,3),
                 input_shape = input_shape, activation='relu')) #합성곱+활성함수 
model.add(Dropout(rate=0.1)) #10% 무작위 n/w 제거

전결합층 : Flatten layer : 3d(h,w,c) -> 1d(h*w*c)

model.add(Flatten())

DNN layer1 : 4층(hidden layer)

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

DNN layer2 : 5층(output layer)

model.add(Dense(units=10, activation='softmax'))          
model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_8 (Conv2D)            (None, 28, 28, 32)        2432      
_________________________________________________________________
max_pooling2d_6 (MaxPooling2 (None, 13, 13, 32)        0         
_________________________________________________________________
conv2d_9 (Conv2D)            (None, 9, 9, 64)          51264     
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 4, 4, 64)          0         
_________________________________________________________________
conv2d_10 (Conv2D)           (None, 2, 2, 128)         73856     
_________________________________________________________________
flatten_1 (Flatten)          (None, 512)               0         
_________________________________________________________________
dense_5 (Dense)              (None, 64)                32832     
_________________________________________________________________
dense_6 (Dense)              (None, 10)                650       
=================================================================



5. model compile : 학습환경(다항분류기)

model.compile(optimizer='adam',
              loss='categorical_crossentropy', #y = one_hot encoding 
              metrics=['accuracy'])

[추가] Tensorboad

from tensorflow.keras.callbacks import TensorBoard
from datetime import datetime #'20211215-134135'

log file 저장 위치 

logdir = 'c:/graph/' + datetime.now().strftime('%Y%m%d-%H%M%S')
callback = TensorBoard(log_dir = logdir)

6. model training : 학습

model_fit = model.fit(x=x_train, y=y_train, #훈련셋 
          epochs=10, #반복학습
          batch_size = 100, #1회공급 image 개수(1epoch=100*500) 
          verbose=1,
          validation_data=(x_val, y_val), #검증셋 
          callbacks = [callback]) #[추가] log file

7. model 평가

print('='*30)
model.evaluate(x=x_val, y=y_val)

8. model history 
loss vs val_loss : 과적합 시작점 : 3.5

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

accuracy vs val_accuracy : 과적합 시작점 : 4

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