CNN

합성곱

import numpy as np
from PIL import Image

img = Image.open('balloon.webp').resize((256, 256)).convert('L')
img

img = np.array(img, dtype='float32')[np.newaxis, ..., np.newaxis]
img.shape

(1, 256, 256, 1)

from tensorflow.keras.layers import Conv2D
conv = Conv2D(1, (3, 3))
conv.build(input_shape=(None, 256, 256, 1))

# 블러
# filter = [
#     [1/8, 1/8, 1/8], 
#     [1/8,   0, 1/8], 
#     [1/8, 1/8, 1/8]]
# 경계선
# filter = [
#     [-1/8, -1/8, -1/8], 
#     [-1/8,    1, -1/8], 
#     [-1/8, -1/8, -1/8]]
# 수평선
# filter = [
#     [-1/4, -2/4, -1/4], 
#     [   0,    0,    0], 
#     [ 1/4,  2/4,  1/4]]
# 수직선
filter = [
    [1/4, 0, -1/4], 
    [2/4, 0, -2/4], 
    [1/4, 0, -1/4]]

filter = np.array(filter, dtype='float32').reshape((3, 3, 1, 1))
conv.weights[0].assign(filter);

feature = conv(img)

Image.fromarray(feature[0, ..., 0].numpy().astype('uint8'))

MNIST 데이터셋

import tensorflow as tf
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()

x_train의 형태는 (N, 28, 28)
컬러 이미지의 경우 삼원색의 이미지가 겹쳐져서 (N, 28, 28, 3) 형태
CNN의 경우 입력에서 마지막 차원이 색상의 수를 나타냄
흑백 이미지이므로 (N, 28, 28, 1) 형태로 바꿔줌

import numpy as np
x_train = np.expand_dims(x_train, -1)
x_test = np.expand_dims(x_test, -1)

CNN 모델 정의

from tensorflow.keras.layers import *
model = tf.keras.Sequential([
    Rescaling(1/255),
    Conv2D(32, kernel_size=(3, 3), activation="relu"),
    MaxPooling2D(pool_size=(2, 2)),
    Conv2D(64, kernel_size=(3, 3), activation="relu"),
    MaxPooling2D(pool_size=(2, 2)),
    Flatten(),
    Dropout(0.5),
    Dense(10, activation="softmax"),
])

훈련

설정

model.compile(loss='sparse_categorical_crossentropy', 
              optimizer='adam', metrics=['accuracy'])

훈련

model.fit(x_train, y_train, epochs=10)

테스트

model.evaluate(x_test, y_test)

313/313 [==============================] - 1s 3ms/step - loss: 0.0238 - accuracy: 0.9919

[0.023816542699933052, 0.9919000267982483]

CIFAR10

10 종류의 32x32 크기 컬러 이미지 데이터

훈련용 5만개, 테스트용 1만개

컬러 이미지이므로 데이터 로딩 후 차원을 추가할 필요 없음

(x_train, y_train), (x_test, y_test) = \
    tf.keras.datasets.cifar10.load_data()

Downloading data from https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
170498071/170498071 [==============================] - 24s 0us/step