構造

実装

def InceptionV3_block1():
    def f(x):
        b1 = Convolution2D(64, (1,1), strides=2, padding='same')(x)
        b1 = BatchNormalization()(b1)
        b1 = Activation('relu')(b1)

        b2 = Convolution2D(48, (1,1))(x)
        b2 = BatchNormalization()(b2)
        b2 = Activation('relu')(b2)
        b2 = Convolution2D(96, (3,3), strides=2, padding='same')(b2)
        b2 = BatchNormalization()(b2)
        b2 = Activation('relu')(b2)


        b3 = AveragePooling2D(pool_size=(3, 3), strides=2,  padding='same')(x)
        b3 = Convolution2D(64, (3,3), padding='same')(b3)
        b3 = BatchNormalization()(b3)
        b3 = Activation('relu')(b3)

        b4 = Convolution2D(64, (1,1))(x)
        b4 = BatchNormalization()(b4)
        b4 = Activation('relu')(b4)
        b4 = Convolution2D(96, (3,3), padding='same')(b4)
        b4 = BatchNormalization()(b4)
        b4 = Activation('relu')(b4)
        b4 = Convolution2D(32, (3,3),strides=2, padding='same')(b4)
        b4 = BatchNormalization()(b4)
        b4 = Activation('relu')(b4)

        output = concatenate([b1, b2, b3, b4], axis=-1)
        return output
    return f

def InceptionV3_block2():
    def f(x):
        b1 = Convolution2D(64, (1,1), strides=2,  padding='same')(x)
        b1 = BatchNormalization()(b1)
        b1 = Activation('relu')(b1)

        b2 = Convolution2D(48, (1,1))(x)
        b2 = BatchNormalization()(b2)
        b2 = Activation('relu')(b2)
        b2 = Convolution2D(96, (3,3), strides=2,  padding='same')(b2)
        b2 = BatchNormalization()(b2)
        b2 = Activation('relu')(b2)


        b3 = AveragePooling2D(pool_size=(3, 3), strides=2,  padding='same')(x)
        b3 = Convolution2D(64, (3,3), padding='same')(b3)
        b3 = BatchNormalization()(b3)
        b3 = Activation('relu')(b3)

        b4 = Convolution2D(64, (1,1))(x)
        b4 = BatchNormalization()(b4)
        b4 = Activation('relu')(b4)
        b4 = Convolution2D(96, (3,3), padding='same')(b4)
        b4 = BatchNormalization()(b4)
        b4 = Activation('relu')(b4)
        b4 = Convolution2D(64, (3,3),strides=2, padding='same')(b4)
        b4 = BatchNormalization()(b4)
        b4 = Activation('relu')(b4)

        output = concatenate([b1, b2, b3, b4], axis=-1)
        return output
    return f

def InceptionV3():
    inputs = Input(shape=(32, 32, 3))
    x = Convolution2D(32, (1,1), strides=2)(inputs)
    x = BatchNormalization()(x)
    x = Activation('relu')(x)
    x = MaxPooling2D((3, 3), strides=(2,2), padding='same')(x)
  
    x = InceptionV3_block1()(x)
    x = InceptionV3_block2()(x)
  
    x = GlobalAveragePooling2D()(x)
    x = Dense(10, kernel_initializer='he_normal', activation='softmax')(x)
  
    model = Model(inputs=inputs, outputs=x)
    return model

model = InceptionV3()
adam = Adam()
model.compile(optimizer=adam, loss='categorical_crossentropy', metrics=['accuracy'])
model.summary()