Practice CIFAR-10
CIFAR-10 데이터셋 불러오기
import os
import random
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Conv2D, MaxPooling2D, Dropout, Flatten
from tensorflow.keras.datasets import cifar10
from tensorflow.keras.utils import to_categorical
from matplotlib.ticker import (MultipleLocator, FormatStrFormatter)
from dataclasses import dataclass
SEED_VALUE = 42
# 임의의 랜덤 시드 사용(42)
random.seed(SEED_VALUE)
np.random.seed(SEED_VALUE)
tf.random.set_seed(SEED_VALUE)
# load_data() 함수를 사용하여 데이터셋을 로드하고 샘플 수와 데이터의 형태를 확인
(X_train, y_train), (X_test, y_test) = cifar10.load_data()
print(X_train.shape)
print(X_test.shape)
Downloading data from https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
170498071/170498071 [==============================] - 13s 0us/step
(50000, 32, 32, 3)
(10000, 32, 32, 3)
데이터셋의 샘플 이미지 표시
plt.figure(figsize=(7, 7))
num_rows = 10
num_cols = 10
for i in range(num_rows*num_cols):
ax = plt.subplot(num_rows, num_cols, i + 1)
plt.imshow(X_train[i,:,:])
plt.axis("off")
데이터셋 전처리
# 이미지를 [0, 1] 범위로 정규화
X_train = X_train.astype("float32") / 255
X_test = X_test.astype("float32") / 255
# 레이블을 정수에서 범주형 데이터로 변경
print('training sample: ', y_train[0])
# 레이블을 원-핫 인코딩으로 변환
y_train = to_categorical(y_train)
y_test = to_categorical(y_test)
print('categorical one-hot encoded labels: ', y_train[0])
Original (integer) label for the first training sample: [6]
After conversion to categorical one-hot encoded labels: [0. 0. 0. 0. 0. 0. 1. 0. 0. 0.]
데이터 세트 및 학습 구성 매개 변수
@dataclass(frozen=True)
class DatasetConfig:
NUM_CLASSES: int = 10
IMG_HEIGHT: int = 32
IMG_WIDTH: int = 32
NUM_CHANNELS: int = 3
@dataclass(frozen=True)
class TrainingConfig:
EPOCHS: int = 31
BATCH_SIZE: int = 256
LEARNING_RATE: float = 0.001
Keras의 CNN 모델 구현
# Keras에서 사전 정의된 계층을 사용하여 네트워크 모델을 구축/정의
def cnn_model(input_shape=(32, 32, 3)):
model = Sequential()
#------------------------------------
# Conv Block 1: 32 Filters, MaxPool.
#------------------------------------
model.add(Conv2D(filters=32, kernel_size=3, padding='same', activation='relu', input_shape=input_shape))
model.add(Conv2D(filters=32, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
#------------------------------------
# Conv Block 2: 64 Filters, MaxPool.
#------------------------------------
model.add(Conv2D(filters=64, kernel_size=3, padding='same', activation='relu'))
model.add(Conv2D(filters=64, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
#------------------------------------
# Conv Block 3: 64 Filters, MaxPool.
#------------------------------------
model.add(Conv2D(filters=64, kernel_size=3, padding='same', activation='relu'))
model.add(Conv2D(filters=64, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
#------------------------------------
# Flatten the convolutional features.
#------------------------------------
model.add(Flatten())
model.add(Dense(512, activation='relu'))
model.add(Dense(10, activation='softmax'))
return model
# 모델 생성
model = cnn_model()
model.summary()
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 32, 32, 32) 896
conv2d_1 (Conv2D) (None, 32, 32, 32) 9248
max_pooling2d (MaxPooling2 (None, 16, 16, 32) 0
D)
conv2d_2 (Conv2D) (None, 16, 16, 64) 18496
conv2d_3 (Conv2D) (None, 16, 16, 64) 36928
max_pooling2d_1 (MaxPoolin (None, 8, 8, 64) 0
g2D)
conv2d_4 (Conv2D) (None, 8, 8, 64) 36928
conv2d_5 (Conv2D) (None, 8, 8, 64) 36928
max_pooling2d_2 (MaxPoolin (None, 4, 4, 64) 0
g2D)
flatten (Flatten) (None, 1024) 0
dense (Dense) (None, 512) 524800
dense_1 (Dense) (None, 10) 5130
=================================================================
Total params: 669354 (2.55 MB)
Trainable params: 669354 (2.55 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
# 모델 컴파일
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'],
)
# 모델 학습
history = model.fit(X_train,
y_train,
batch_size=TrainingConfig.BATCH_SIZE,
epochs=TrainingConfig.EPOCHS,
verbose=1,
validation_split=.3,
)
Epoch 1/31
137/137 [==============================] - 10s 33ms/step - loss: 2.1757 - accuracy: 0.2045 - val_loss: 1.8987 - val_accuracy: 0.3151
Epoch 2/31
137/137 [==============================] - 3s 22ms/step - loss: 1.7648 - accuracy: 0.3590 - val_loss: 1.5809 - val_accuracy: 0.4207
Epoch 3/31
137/137 [==============================] - 3s 24ms/step - loss: 1.5269 - accuracy: 0.4414 - val_loss: 1.4883 - val_accuracy: 0.4513
Epoch 4/31
137/137 [==============================] - 3s 20ms/step - loss: 1.3787 - accuracy: 0.4998 - val_loss: 1.2668 - val_accuracy: 0.5461
Epoch 5/31
137/137 [==============================] - 3s 22ms/step - loss: 1.2518 - accuracy: 0.5522 - val_loss: 1.4190 - val_accuracy: 0.5347
Epoch 6/31
137/137 [==============================] - 3s 20ms/step - loss: 1.1300 - accuracy: 0.5980 - val_loss: 1.2235 - val_accuracy: 0.5575
Epoch 7/31
137/137 [==============================] - 3s 23ms/step - loss: 1.0162 - accuracy: 0.6351 - val_loss: 1.0210 - val_accuracy: 0.6428
Epoch 8/31
137/137 [==============================] - 3s 23ms/step - loss: 0.9169 - accuracy: 0.6729 - val_loss: 1.0173 - val_accuracy: 0.6393
Epoch 9/31
137/137 [==============================] - 3s 20ms/step - loss: 0.8276 - accuracy: 0.7073 - val_loss: 0.9755 - val_accuracy: 0.6617
Epoch 10/31
137/137 [==============================] - 3s 20ms/step - loss: 0.7387 - accuracy: 0.7380 - val_loss: 1.0000 - val_accuracy: 0.6583
Epoch 11/31
137/137 [==============================] - 3s 23ms/step - loss: 0.6525 - accuracy: 0.7714 - val_loss: 0.8477 - val_accuracy: 0.7123
Epoch 12/31
137/137 [==============================] - 3s 21ms/step - loss: 0.5736 - accuracy: 0.7949 - val_loss: 0.8547 - val_accuracy: 0.7181
Epoch 13/31
137/137 [==============================] - 3s 20ms/step - loss: 0.4931 - accuracy: 0.8284 - val_loss: 0.8719 - val_accuracy: 0.7163
Epoch 14/31
137/137 [==============================] - 3s 20ms/step - loss: 0.4135 - accuracy: 0.8527 - val_loss: 1.0511 - val_accuracy: 0.6809
Epoch 15/31
137/137 [==============================] - 3s 20ms/step - loss: 0.3460 - accuracy: 0.8793 - val_loss: 0.9708 - val_accuracy: 0.7279
Epoch 16/31
137/137 [==============================] - 3s 22ms/step - loss: 0.2758 - accuracy: 0.9017 - val_loss: 1.0801 - val_accuracy: 0.7189
Epoch 17/31
137/137 [==============================] - 3s 21ms/step - loss: 0.2205 - accuracy: 0.9243 - val_loss: 1.2581 - val_accuracy: 0.7067
Epoch 18/31
137/137 [==============================] - 3s 23ms/step - loss: 0.1828 - accuracy: 0.9383 - val_loss: 1.2623 - val_accuracy: 0.7232
Epoch 19/31
137/137 [==============================] - 3s 20ms/step - loss: 0.1508 - accuracy: 0.9497 - val_loss: 1.3480 - val_accuracy: 0.7129
Epoch 20/31
137/137 [==============================] - 3s 22ms/step - loss: 0.1192 - accuracy: 0.9602 - val_loss: 1.5366 - val_accuracy: 0.7155
Epoch 21/31
137/137 [==============================] - 3s 21ms/step - loss: 0.1080 - accuracy: 0.9648 - val_loss: 1.8424 - val_accuracy: 0.6703
Epoch 22/31
137/137 [==============================] - 3s 23ms/step - loss: 0.1006 - accuracy: 0.9684 - val_loss: 1.6001 - val_accuracy: 0.7140
Epoch 23/31
137/137 [==============================] - 3s 21ms/step - loss: 0.0938 - accuracy: 0.9692 - val_loss: 1.7790 - val_accuracy: 0.7189
Epoch 24/31
137/137 [==============================] - 3s 23ms/step - loss: 0.0870 - accuracy: 0.9727 - val_loss: 1.7076 - val_accuracy: 0.7322
Epoch 25/31
137/137 [==============================] - 3s 25ms/step - loss: 0.0740 - accuracy: 0.9761 - val_loss: 1.8602 - val_accuracy: 0.7165
Epoch 26/31
137/137 [==============================] - 3s 25ms/step - loss: 0.0711 - accuracy: 0.9773 - val_loss: 1.8184 - val_accuracy: 0.7195
Epoch 27/31
137/137 [==============================] - 3s 23ms/step - loss: 0.0630 - accuracy: 0.9802 - val_loss: 2.0117 - val_accuracy: 0.7183
Epoch 28/31
137/137 [==============================] - 3s 24ms/step - loss: 0.0676 - accuracy: 0.9795 - val_loss: 2.1262 - val_accuracy: 0.7027
Epoch 29/31
137/137 [==============================] - 3s 21ms/step - loss: 0.0605 - accuracy: 0.9802 - val_loss: 2.1155 - val_accuracy: 0.7218
Epoch 30/31
137/137 [==============================] - 3s 23ms/step - loss: 0.0598 - accuracy: 0.9811 - val_loss: 2.1629 - val_accuracy: 0.7207
Epoch 31/31
137/137 [==============================] - 3s 23ms/step - loss: 0.0604 - accuracy: 0.9815 - val_loss: 2.1921 - val_accuracy: 0.7115
훈련 결과 플로팅
def plot_results(metrics, title=None, ylabel=None, ylim=None, metric_name=None, color=None):
fig, ax = plt.subplots(figsize=(15, 4))
if not (isinstance(metric_name, list) or isinstance(metric_name, tuple)):
metrics = [metrics,]
metric_name = [metric_name,]
for idx, metric in enumerate(metrics):
ax.plot(metric, color=color[idx])
plt.xlabel("Epoch")
plt.ylabel(ylabel)
plt.title(title)
plt.xlim([0, TrainingConfig.EPOCHS-1])
plt.ylim(ylim)
# Tailor x-axis tick marks
ax.xaxis.set_major_locator(MultipleLocator(5))
ax.xaxis.set_major_formatter(FormatStrFormatter('%d'))
ax.xaxis.set_minor_locator(MultipleLocator(1))
plt.grid(True)
plt.legend(metric_name)
plt.show()
plt.close()
# Retrieve training results.
train_loss = history.history["loss"]
train_acc = history.history["accuracy"]
valid_loss = history.history["val_loss"]
valid_acc = history.history["val_accuracy"]
plot_results([ train_loss, valid_loss ],
ylabel="Loss",
ylim = [0.0, 5.0],
metric_name=["Training Loss", "Validation Loss"],
color=["g", "b"]);
plot_results([ train_acc, valid_acc ],
ylabel="Accuracy",
ylim = [0.0, 1.0],
metric_name=["Training Accuracy", "Validation Accuracy"],
color=["g", "b"])
모델에 드롭아웃 추가
def cnn_model_dropout(input_shape=(32, 32, 3)):
model = Sequential()
#------------------------------------
# Conv Block 1: 32 Filters, MaxPool.
#------------------------------------
model.add(Conv2D(filters=32, kernel_size=3, padding='same', activation='relu', input_shape=input_shape))
model.add(Conv2D(filters=32, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
#------------------------------------
# Conv Block 2: 64 Filters, MaxPool.
#------------------------------------
model.add(Conv2D(filters=64, kernel_size=3, padding='same', activation='relu'))
model.add(Conv2D(filters=64, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
#------------------------------------
# Conv Block 3: 64 Filters, MaxPool.
#------------------------------------
model.add(Conv2D(filters=64, kernel_size=3, padding='same', activation='relu'))
model.add(Conv2D(filters=64, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
#------------------------------------
# Flatten the convolutional features.
#------------------------------------
model.add(Flatten())
model.add(Dense(512, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(10, activation='softmax'))
return model
# 모델 생성
model_dropout = cnn_model_dropout()
model_dropout.summary()
Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_6 (Conv2D) (None, 32, 32, 32) 896
conv2d_7 (Conv2D) (None, 32, 32, 32) 9248
max_pooling2d_3 (MaxPoolin (None, 16, 16, 32) 0
g2D)
dropout (Dropout) (None, 16, 16, 32) 0
conv2d_8 (Conv2D) (None, 16, 16, 64) 18496
conv2d_9 (Conv2D) (None, 16, 16, 64) 36928
max_pooling2d_4 (MaxPoolin (None, 8, 8, 64) 0
g2D)
dropout_1 (Dropout) (None, 8, 8, 64) 0
conv2d_10 (Conv2D) (None, 8, 8, 64) 36928
conv2d_11 (Conv2D) (None, 8, 8, 64) 36928
max_pooling2d_5 (MaxPoolin (None, 4, 4, 64) 0
g2D)
dropout_2 (Dropout) (None, 4, 4, 64) 0
flatten_1 (Flatten) (None, 1024) 0
dense_2 (Dense) (None, 512) 524800
dropout_3 (Dropout) (None, 512) 0
dense_3 (Dense) (None, 10) 5130
=================================================================
Total params: 669354 (2.55 MB)
Trainable params: 669354 (2.55 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
model_dropout.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'],
)
history = model_dropout.fit(X_train,
y_train,
batch_size=TrainingConfig.BATCH_SIZE,
epochs=TrainingConfig.EPOCHS,
verbose=1,
validation_split=.3,
)
Epoch 1/31
137/137 [==============================] - 8s 34ms/step - loss: 2.1466 - accuracy: 0.2049 - val_loss: 2.0399 - val_accuracy: 0.2543
Epoch 2/31
137/137 [==============================] - 5s 36ms/step - loss: 1.8586 - accuracy: 0.3310 - val_loss: 1.6229 - val_accuracy: 0.4019
Epoch 3/31
137/137 [==============================] - 4s 27ms/step - loss: 1.6434 - accuracy: 0.4079 - val_loss: 1.4500 - val_accuracy: 0.4736
Epoch 4/31
137/137 [==============================] - 4s 28ms/step - loss: 1.4957 - accuracy: 0.4597 - val_loss: 1.3105 - val_accuracy: 0.5251
Epoch 5/31
137/137 [==============================] - 4s 26ms/step - loss: 1.3728 - accuracy: 0.5065 - val_loss: 1.3360 - val_accuracy: 0.5172
Epoch 6/31
137/137 [==============================] - 4s 30ms/step - loss: 1.2613 - accuracy: 0.5488 - val_loss: 1.3085 - val_accuracy: 0.5423
Epoch 7/31
137/137 [==============================] - 4s 26ms/step - loss: 1.1760 - accuracy: 0.5827 - val_loss: 1.0483 - val_accuracy: 0.6257
Epoch 8/31
137/137 [==============================] - 4s 29ms/step - loss: 1.1000 - accuracy: 0.6099 - val_loss: 0.9638 - val_accuracy: 0.6559
Epoch 9/31
137/137 [==============================] - 4s 28ms/step - loss: 1.0316 - accuracy: 0.6346 - val_loss: 0.9656 - val_accuracy: 0.6584
Epoch 10/31
137/137 [==============================] - 4s 29ms/step - loss: 0.9723 - accuracy: 0.6551 - val_loss: 0.9065 - val_accuracy: 0.6806
Epoch 11/31
137/137 [==============================] - 4s 29ms/step - loss: 0.9263 - accuracy: 0.6726 - val_loss: 0.8328 - val_accuracy: 0.7091
Epoch 12/31
137/137 [==============================] - 4s 29ms/step - loss: 0.8802 - accuracy: 0.6909 - val_loss: 0.8497 - val_accuracy: 0.7026
Epoch 13/31
137/137 [==============================] - 4s 28ms/step - loss: 0.8335 - accuracy: 0.7038 - val_loss: 0.8664 - val_accuracy: 0.6953
Epoch 14/31
137/137 [==============================] - 4s 26ms/step - loss: 0.8004 - accuracy: 0.7174 - val_loss: 0.8244 - val_accuracy: 0.7139
Epoch 15/31
137/137 [==============================] - 4s 29ms/step - loss: 0.7768 - accuracy: 0.7269 - val_loss: 0.7989 - val_accuracy: 0.7231
Epoch 16/31
137/137 [==============================] - 4s 29ms/step - loss: 0.7370 - accuracy: 0.7401 - val_loss: 0.7107 - val_accuracy: 0.7546
Epoch 17/31
137/137 [==============================] - 4s 26ms/step - loss: 0.7163 - accuracy: 0.7487 - val_loss: 0.7445 - val_accuracy: 0.7460
Epoch 18/31
137/137 [==============================] - 4s 29ms/step - loss: 0.6881 - accuracy: 0.7579 - val_loss: 0.6884 - val_accuracy: 0.7600
Epoch 19/31
137/137 [==============================] - 4s 27ms/step - loss: 0.6670 - accuracy: 0.7633 - val_loss: 0.6727 - val_accuracy: 0.7661
Epoch 20/31
137/137 [==============================] - 4s 26ms/step - loss: 0.6448 - accuracy: 0.7722 - val_loss: 0.7363 - val_accuracy: 0.7433
Epoch 21/31
137/137 [==============================] - 4s 28ms/step - loss: 0.6266 - accuracy: 0.7790 - val_loss: 0.6804 - val_accuracy: 0.7686
Epoch 22/31
137/137 [==============================] - 4s 29ms/step - loss: 0.6086 - accuracy: 0.7865 - val_loss: 0.6885 - val_accuracy: 0.7609
Epoch 23/31
137/137 [==============================] - 4s 26ms/step - loss: 0.5923 - accuracy: 0.7921 - val_loss: 0.7677 - val_accuracy: 0.7401
Epoch 24/31
137/137 [==============================] - 4s 28ms/step - loss: 0.5711 - accuracy: 0.7998 - val_loss: 0.7026 - val_accuracy: 0.7614
Epoch 25/31
137/137 [==============================] - 4s 28ms/step - loss: 0.5580 - accuracy: 0.8035 - val_loss: 0.6414 - val_accuracy: 0.7824
Epoch 26/31
137/137 [==============================] - 4s 29ms/step - loss: 0.5387 - accuracy: 0.8097 - val_loss: 0.7137 - val_accuracy: 0.7622
Epoch 27/31
137/137 [==============================] - 4s 26ms/step - loss: 0.5267 - accuracy: 0.8132 - val_loss: 0.6783 - val_accuracy: 0.7748
Epoch 28/31
137/137 [==============================] - 4s 29ms/step - loss: 0.5143 - accuracy: 0.8167 - val_loss: 0.6377 - val_accuracy: 0.7879
Epoch 29/31
137/137 [==============================] - 4s 29ms/step - loss: 0.5033 - accuracy: 0.8238 - val_loss: 0.6172 - val_accuracy: 0.7927
Epoch 30/31
137/137 [==============================] - 4s 28ms/step - loss: 0.4858 - accuracy: 0.8275 - val_loss: 0.6394 - val_accuracy: 0.7884
Epoch 31/31
137/137 [==============================] - 4s 27ms/step - loss: 0.4810 - accuracy: 0.8299 - val_loss: 0.6545 - val_accuracy: 0.7846
드롭아웃이 적용된 훈련 결과 플로팅
# Retrieve training results.
train_loss = history.history["loss"]
train_acc = history.history["accuracy"]
valid_loss = history.history["val_loss"]
valid_acc = history.history["val_accuracy"]
plot_results([ train_loss, valid_loss ],
ylabel="Loss",
ylim = [0.0, 5.0],
metric_name=["Training Loss", "Validation Loss"],
color=["g", "b"]);
plot_results([ train_acc, valid_acc ],
ylabel="Accuracy",
ylim = [0.0, 1.0],
metric_name=["Training Accuracy", "Validation Accuracy"],
color=["g", "b"])
# 모델 저장
model_dropout.save('model_dropout')
from tensorflow.keras import models
reloaded_model_dropout = models.load_model('model_dropout')
# 테스트 데이터 세트에서 모델 평가
test_loss, test_acc = reloaded_model_dropout.evaluate(X_test, y_test)
print(f"Test accuracy: {test_acc*100:.3f}")
313/313 [==============================] - 2s 5ms/step - loss: 0.6712 - accuracy: 0.7822
Test accuracy: 78.220
샘플 테스트 이미지 예측
def evaluate_model(dataset, model):
class_names = ['airplane',
'automobile',
'bird',
'cat',
'deer',
'dog',
'frog',
'horse',
'ship',
'truck' ]
num_rows = 3
num_cols = 6
# Retrieve a number of _images/2024/05 from the dataset.
data_batch = dataset[0:num_rows*num_cols]
# Get predictions from model.
predictions = model.predict(data_batch)
plt.figure(figsize=(20, 8))
num_matches = 0
for idx in range(num_rows*num_cols):
ax = plt.subplot(num_rows, num_cols, idx + 1)
plt.axis("off")
plt.imshow(data_batch[idx])
pred_idx = tf.argmax(predictions[idx]).numpy()
truth_idx = np.nonzero(y_test[idx])
title = str(class_names[truth_idx[0][0]]) + " : " + str(class_names[pred_idx])
title_obj = plt.title(title, fontdict={'fontsize':13})
if pred_idx == truth_idx:
num_matches += 1
plt.setp(title_obj, color='g')
else:
plt.setp(title_obj, color='r')
acc = num_matches/(idx+1)
print("Prediction accuracy: ", int(100*acc)/100)
return
evaluate_model(X_test, reloaded_model_dropout)
1/1 [==============================] - 0s 289ms/step
Prediction accuracy: 0.83
References
https://learnopencv.com/implementing-cnn-tensorflow-keras/
댓글남기기