Skip to content

다항 분류

Fashion MNIST

불러오기

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

from PIL import Image
Image.fromarray(x_train[0]).resize((200, 200))

전처리

x_train = x_train / 127.5 - 1
x_test = x_test / 127.5 - 1
  • 이미지 데이터는 0-255로 밝기를 표현
  • 신경망에서 사용하는 시그모이드 등의 함수는 지나치게 큰 값이 들어오면 경사가 0에 가까워져서 학습이 잘 되지 않음
  • $[-1, +1]$ 범위로 재조정

모형

from tensorflow.keras.layers import *
model = tf.keras.Sequential([
    Flatten(),
    Dense(10, activation='softmax')
])

소프트맥스 함수

tf.nn.softmax([-1. ,  0.5,  2. ])

<tf.Tensor: shape=(3,), dtype=float32, numpy=array([0.03911257, 0.17529039, 0.785597  ], dtype=float32)>

시그모이드 vs. 소프트맥스

시그모이드

x = 1.0
tf.nn.sigmoid(x)

<tf.Tensor: shape=(), dtype=float32, numpy=0.7310586>

소프트맥스

tf.nn.softmax([0, x]) 

<tf.Tensor: shape=(2,), dtype=float32, numpy=array([0.26894143, 0.7310586 ], dtype=float32)>

훈련

설정

model.compile(
    loss=tf.keras.losses.sparse_categorical_crossentropy,
    metrics=['accuracy'])

훈련

model.fit(x_train, y_train, epochs=100, validation_split=0.1,
          callbacks=[tf.keras.callbacks.EarlyStopping(patience=2)])

Epoch 1/100
1688/1688 [==============================] - 2s 946us/step - loss: 0.5530 - accuracy: 0.8067 - val_loss: 0.4958 - val_accuracy: 0.8222
Epoch 2/100
1688/1688 [==============================] - 1s 839us/step - loss: 0.4627 - accuracy: 0.8386 - val_loss: 0.4651 - val_accuracy: 0.8325
Epoch 3/100
1688/1688 [==============================] - 1s 788us/step - loss: 0.4448 - accuracy: 0.8443 - val_loss: 0.4562 - val_accuracy: 0.8380
Epoch 4/100
1688/1688 [==============================] - 1s 801us/step - loss: 0.4351 - accuracy: 0.8481 - val_loss: 0.4623 - val_accuracy: 0.8368
Epoch 5/100
1688/1688 [==============================] - 1s 710us/step - loss: 0.4277 - accuracy: 0.8501 - val_loss: 0.4601 - val_accuracy: 0.8345

<keras.callbacks.History at 0x232dd3ee0b0>

평가

정확도

model.evaluate(x_test, y_test)

313/313 [==============================] - 0s 781us/step - loss: 0.4857 - accuracy: 0.8243

[0.48574209213256836, 0.8242999911308289]

예측

prob = model.predict(x_test)

313/313 [==============================] - 0s 547us/step

가장 확률이 높은 카테고리

y_pred = prob.argmax(axis=1)

혼동행렬

from sklearn.metrics import confusion_matrix
confusion_matrix(y_test, y_pred)

array([[658,   4,  17,  67,   4,   0, 242,   0,   8,   0],
       [  1, 963,   4,  22,   3,   0,   6,   0,   1,   0],
       [  8,   6, 696,  18, 120,   0, 143,   0,   9,   0],
       [ 10,  22,   7, 860,  10,   1,  85,   0,   5,   0],
       [  0,   3, 108,  61, 687,   0, 133,   0,   8,   0],
       [  0,   0,   0,   1,   0, 943,   2,  35,   4,  15],
       [ 69,   2, 109,  52,  85,   0, 665,   0,  18,   0],
       [  0,   0,   0,   0,   0,  60,   0, 917,   0,  23],
       [  2,   1,   7,   8,   2,   6,  35,   7, 932,   0],
       [  0,   0,   0,   0,   0,  27,   0,  50,   1, 922]], dtype=int64)