본문 바로가기
DEEP LEARNING/Tensorflow Training

cnn 실습_나만의 이미지 데이터를 만들고 학습하기(4) - additional models

Mj.Hong 2020. 3. 26. 16:18

MLP

cnn이전에 먼저 multi layer percetron으로도 학습 가능

## model function
def create_model():
    model = keras.Sequential()
    model.add(keras.layers.Flatten(input_shape=(128,128,3)))
    model.add(keras.layers.Dense(512, activation='relu'))
    model.add(keras.layers.Dense(256, activation='relu'))
    model.add(keras.layers.Dense(128, activation='relu'))
    model.add(keras.layers.Dense(5, activation='softmax'))
    return model

GoogLeNet

## GoogLeNet
from tensorflow.keras.layers import Input, Conv2D, Dense, MaxPool2D, 
GlobalAveragePooling2D, Concatenate, Dropout

IMG_SIZE = 224

def create_model():
    ## Stem
    inputs = Input(shape=(IMG_SIZE, IMG_SIZE, 3))
    net = Conv2D(64, 7, 2, 'SAME', activation='relu')(inputs)
    net = MaxPool2D(3, 2, 'SAME')(net)
    net = Conv2D(64, 1, 1, 'SAME', activation='relu')(net)
    net = Conv2D(192, 3, 1, 'SAME', activation='relu')(net)
    net = MaxPool2D(3, 2, 'SAME')(net)
    ## inception 3a
    b1 = Conv2D(64, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(96, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(128, 3, 1, 'SAME', activation='relu')(b2)
    b3 = Conv2D(16, 1, 1, 'SAME', activation='relu')(net)
    b3 = Conv2D(32, 5, 1, 'SAME', activation='relu')(b3)
    b4 = MaxPool2D(3, 1, 'SAME')(net)
    b4 = Conv2D(32, 1, 1, 'SAME', activation='relu')(b4)
    net = Concatenate()([b1, b2, b3, b4])
    ## inception 3b
    b1 = Conv2D(128, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(128, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(192, 3, 1, 'SAME', activation='relu')(b2)
    b3 = Conv2D(32, 1, 1, 'SAME', activation='relu')(net)
    b3 = Conv2D(96, 5, 1, 'SAME', activation='relu')(b3)
    b4 = MaxPool2D(3, 1, 'SAME')(net)
    b4 = Conv2D(64, 1, 1, 'SAME', activation='relu')(b4)
    net = Concatenate()([b1, b2, b3, b4])
    ## 
    net = MaxPool2D(3, 2, 'SAME')(net)
    ## inception 4a
    b1 = Conv2D(192, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(96, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(208, 3, 1, 'SAME', activation='relu')(b2)
    b3 = Conv2D(16, 1, 1, 'SAME', activation='relu')(net)
    b3 = Conv2D(48, 5, 1, 'SAME', activation='relu')(b3)
    b4 = MaxPool2D(3, 1, 'SAME')(net)
    b4 = Conv2D(64, 1, 1, 'SAME', activation='relu')(b4)
    net = Concatenate()([b1, b2, b3, b4])
    ## inception 4b
    b1 = Conv2D(160, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(112, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(224, 3, 1, 'SAME', activation='relu')(b2)
    b3 = Conv2D(24, 1, 1, 'SAME', activation='relu')(net)
    b3 = Conv2D(64, 5, 1, 'SAME', activation='relu')(b3)
    b4 = MaxPool2D(3, 1, 'SAME')(net)
    b4 = Conv2D(64, 1, 1, 'SAME', activation='relu')(b4)
    net = Concatenate()([b1, b2, b3, b4])
    ## inception 4c
    b1 = Conv2D(128, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(128, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(256, 3, 1, 'SAME', activation='relu')(b2)
    b3 = Conv2D(24, 1, 1, 'SAME', activation='relu')(net)
    b3 = Conv2D(64, 5, 1, 'SAME', activation='relu')(b3)
    b4 = MaxPool2D(3, 1, 'SAME')(net)
    b4 = Conv2D(64, 1, 1, 'SAME', activation='relu')(b4)
    net = Concatenate()([b1, b2, b3, b4])
    ## inception 4d
    b1 = Conv2D(112, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(144, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(288, 3, 1, 'SAME', activation='relu')(b2)
    b3 = Conv2D(32, 1, 1, 'SAME', activation='relu')(net)
    b3 = Conv2D(64, 5, 1, 'SAME', activation='relu')(b3)
    b4 = MaxPool2D(3, 1, 'SAME')(net)
    b4 = Conv2D(64, 1, 1, 'SAME', activation='relu')(b4)
    net = Concatenate()([b1, b2, b3, b4])
    ## inception 4e
    b1 = Conv2D(256, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(160, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(320, 3, 1, 'SAME', activation='relu')(b2)
    b3 = Conv2D(32, 1, 1, 'SAME', activation='relu')(net)
    b3 = Conv2D(128, 5, 1, 'SAME', activation='relu')(b3)
    b4 = MaxPool2D(3, 1, 'SAME')(net)
    b4 = Conv2D(128, 1, 1, 'SAME', activation='relu')(b4)
    net = Concatenate()([b1, b2, b3, b4])
    ##
    net = MaxPool2D(3, 2, 'SAME')(net)
    ## inception 5a
    b1 = Conv2D(256, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(160, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(320, 3, 1, 'SAME', activation='relu')(b2)
    b3 = Conv2D(32, 1, 1, 'SAME', activation='relu')(net)
    b3 = Conv2D(128, 5, 1, 'SAME', activation='relu')(b3)
    b4 = MaxPool2D(3, 1, 'SAME')(net)
    b4 = Conv2D(128, 1, 1, 'SAME', activation='relu')(b4)
    net = Concatenate()([b1, b2, b3, b4])
    ## inception 5b
    b1 = Conv2D(384, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(192, 1, 1, 'SAME', activation='relu')(net)
    b2 = Conv2D(384, 3, 1, 'SAME', activation='relu')(b2)
    b3 = Conv2D(48, 1, 1, 'SAME', activation='relu')(net)
    b3 = Conv2D(128, 5, 1, 'SAME', activation='relu')(b3)
    b4 = MaxPool2D(3, 1, 'SAME')(net)
    b4 = Conv2D(128, 1, 1, 'SAME', activation='relu')(b4)
    net = Concatenate()([b1, b2, b3, b4])
    ## global average pooling
    net = GlobalAveragePooling2D()(net)
    net = Dropout(0.4)(net)
    net = Dense(5, activation='softmax')(net)
    
    return keras.Model(inputs=inputs, outputs=net)

MobileMet

## MobileNet
from tensorflow.keras.layers import Conv2D, Dense, ReLU, Softmax, BatchNormalization, 
DepthwiseConv2D, GlobalAveragePooling2D

IMG_SIZE = 224

def create_model():
    model = keras.Sequential()
    model.add(Conv2D(filters=32, kernel_size=3, strides=2, padding='SAME', 
                     use_bias=False, input_shape=(IMG_SIZE, IMG_SIZE, 3)))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=1, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=64, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=2, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=128, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=1, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=128, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=2, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=256, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=1, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=256, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=2, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=512, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=1, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=512, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=1, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=512, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=1, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=512, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=1, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=512, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=1, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=512, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=2, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=1024, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(DepthwiseConv2D(kernel_size=3, strides=1, padding='SAME', use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Conv2D(filters=1024, kernel_size=1, use_bias=False))
    model.add(BatchNormalization())
    model.add(ReLU())
    
    model.add(GlobalAveragePooling2D())
    model.add(Dense(units=128))
    model.add(BatchNormalization())
    model.add(ReLU())
    model.add(Dense(units=5))
    model.add(BatchNormalization())
    model.add(Softmax())
    return model

 

Module: tf.keras.applications

저작자표시 비영리 변경금지

'DEEP LEARNING > Tensorflow Training' 카테고리의 다른 글

cnn 실습_나만의 이미지 데이터를 만들고 학습하기(3) - functional API & save_model  (0) 2020.03.26
cnn 실습_나만의 이미지 데이터를 만들고 학습하기(2) - sequential API  (0) 2020.03.21
cnn 실습_나만의 이미지 데이터를 만들고 학습하기(1) - image crawling  (0) 2020.03.20
Linear Regression & Logistic Regression  (0) 2020.03.01
Make A Dataset  (0) 2020.02.26

'DEEP LEARNING/Tensorflow Training' Related Articles

티스토리툴바