"""
# Parameters
tf.app.flags.DEFINE_string("test_file", './data/test.txt', "the path of test data")
tf.app.flags.DEFINE_integer("batch_size", 128, "batch_size(default:128)")
tf.app.flags.DEFINE_integer("num_classes", 5, "num_classes(default:2)")
FLAGS = tf.app.flags.FLAGS
num_validation = 10000
train_layers = ["Conv2d_1c_1x1"]
# Load data on the cpu
print("Loading data...")
with tf.device('/cpu:0'):
test_iterator = ImageDataGenerator(txt_file=FLAGS.test_file,
mode='inference',
batch_size=FLAGS.batch_size,
num_classes=FLAGS.num_classes,
shuffle=True,
img_out_size=inception.inception_v3.default_image_size
)
test_next_batch = test_iterator.iterator.get_next()
# Initialize model
inceptionv3 = InceptionV3(num_classes=FLAGS.num_classes, train_layers=train_layers)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(var_list=tf.global_variables())
# Parameters
tf.app.flags.DEFINE_string("test_file", './data/test.txt',
"the path of test data")
tf.app.flags.DEFINE_integer("batch_size", 128, "batch_size(default:128)")
tf.app.flags.DEFINE_integer("num_classes", 5, "num_classes(default:2)")
FLAGS = tf.app.flags.FLAGS
num_validation = 10000
train_layers = ["logits"]
# Load data on the cpu
print("Loading data...")
with tf.device('/cpu:0'):
test_iterator = ImageDataGenerator(
txt_file=FLAGS.test_file,
mode='inference',
batch_size=FLAGS.batch_size,
num_classes=FLAGS.num_classes,
shuffle=True,
img_out_size=resnet_v1.resnet_v1_152.default_image_size)
test_next_batch = test_iterator.iterator.get_next()
# Initialize model
resnetv1_152 = ResNetv1_152(num_classes=FLAGS.num_classes,
train_layers=train_layers)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(var_list=tf.global_variables())
model_file = tf.train.latest_checkpoint(
tf.app.flags.DEFINE_integer("evaluate_every", 200, "Evaluate model on dev set after this many steps (default: 100)")
tf.app.flags.DEFINE_integer("checkpoint_every", 400, "Save model after this many steps (default: 100)")
tf.app.flags.DEFINE_integer("num_checkpoints", 3, "num_checkpoints(default:3)")
FLAGS = tf.app.flags.FLAGS
train_layers = ['fc6', 'fc7', 'fc8']
"""
Main Part of the finetuning Script.
"""
# Load data on the cpu
print("Loading data...")
with tf.device('/cpu:0'):
train_iterator = ImageDataGenerator(txt_file=FLAGS.train_file,
mode='training',
batch_size=FLAGS.batch_size,
num_classes=FLAGS.num_classes,
shuffle=True)
val_iterator = ImageDataGenerator(txt_file=FLAGS.val_file,
mode='inference',
batch_size=FLAGS.batch_size,
num_classes=FLAGS.num_classes,
shuffle=False)
train_next_batch = train_iterator.iterator.get_next()
val_next_batch = val_iterator.iterator.get_next()
# Initialize model
vgg16 = VGG16(keep_prob=FLAGS.keep_prob,
num_classes=FLAGS.num_classes,
tf.app.flags.DEFINE_integer("num_checkpoints", 3, "num_checkpoints(default:3)")
FLAGS = tf.app.flags.FLAGS
num_validation = 10000
train_layers = ["logits"]
"""
Main Part of the finetuning Script.
"""
# Load data on the cpu
print("Loading data...")
with tf.device('/cpu:0'):
train_iterator = ImageDataGenerator(txt_file=FLAGS.train_file,
mode='training',
batch_size=FLAGS.batch_size,
num_classes=FLAGS.num_classes,
shuffle=True,
img_out_size=resnet_v1.resnet_v1_50.default_image_size
)
val_iterator = ImageDataGenerator(txt_file=FLAGS.val_file,
mode='inference',
batch_size=FLAGS.batch_size,
num_classes=FLAGS.num_classes,
shuffle=False,
img_out_size=resnet_v1.resnet_v1_50.default_image_size
)
train_next_batch = train_iterator.iterator.get_next()
val_next_batch = val_iterator.iterator.get_next()
def train(train_data_dir, validation_data_dir, model_path):
base_model = Xception(input_shape=(img_width, img_height, 3),
include_top=False)
x = base_model.output
# 필터를 기준으로 나누어 평균값만 뽑아냄
x = GlobalAveragePooling2D()(x)
# 완전연결계층을 만들어줌
# sigmoid - True(0.5~1), False(0~0.5), 선형을 비선형으로 변경, output layer에서 사용
# Dense() - hidden layer의 Node 수 정의
# multi label일 경우 sigmoid, 아닐 경우 softmax
predictions = layers.Dense(nb_classes, activation='sigmoid')(x)
# 모델 만들기(input, output)
model = Model(base_model.input, predictions)
print(model.summary())
for layer in model.layers:
layer.trainable = True # layer 학습 가능하게 설정
layer.kernel_regularizer = l2(0.05) # 가중치 규제(L2 정규화)
# binary_crossentropy - 0 or 1로 판단
# crossentropy - 하나만 1, 나머지 0으로 판단
model.compile(optimizer='nadam',
loss='binary_crossentropy',
metrics=['accuracy'])
# 학습 데이터 가공
# multi label일 경우 multi_categorical, 아닐 경우 categorical
train_datagen = ImageDataGenerator(rescale=1. / 255,
width_shift_range=0.2,
height_shift_range=0.2,
horizontal_flip=True,
zoom_range=[0.8, 4.5],
fill_mode='nearest')
train_generator = train_datagen.flow_from_directory(
train_data_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
class_mode='multi_categorical')
# ImageDataGenerator 이미지 확인
# import matplotlib.pyplot as plt
# for item in train_generator:
# X, y = item
# print(X.shape)
# plt.figure()
# plt.imshow(X[0])
# plt.show()
print(train_generator.class_indices)
# 검증 데이터 가공
validation_datagen = ImageDataGenerator(rescale=1. / 255)
validation_generator = validation_datagen.flow_from_directory(
validation_data_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
class_mode='multi_categorical')
# 가중치 저장 경로
final_acc_weights_path = os.path.join(
os.path.abspath(model_path), 'acc_weights_190920(b8_e200_r0.05).h5')
final_loss_weights_path = os.path.join(
os.path.abspath(model_path), 'loss_weights_190920(b8_e200_r0.05).h5')
# 콜백 설정
callbacks_list = [
ModelCheckpoint(final_acc_weights_path,
monitor='val_acc',
verbose=1,
save_best_only=True),
ModelCheckpoint(final_loss_weights_path,
monitor='val_loss',
verbose=1,
save_best_only=True),
TensorBoard(log_dir='./graph',
histogram_freq=0,
write_graph=True,
write_images=True)
]
# ImageDataGenerator 사용 시 fit_generator()로 학습
hist = model.fit_generator(train_generator,
epochs=nb_epoch,
validation_data=validation_generator,
callbacks=callbacks_list,
steps_per_epoch=1000)
# 모델 저장
model_json = model.to_json()
with open(os.path.join(os.path.abspath(model_path), 'model.json'),
'w') as json_file:
json_file.write(model_json)
# 결과 그래프 출력
fig, loss_ax = plt.subplots()
acc_ax = loss_ax.twinx()
loss_ax.plot(hist.history['loss'], 'y', label='Train - Loss')
loss_ax.plot(hist.history['val_loss'], 'r', label='Val - Loss')
acc_ax.plot(hist.history['acc'], 'b', label='Train - Accuracy')
acc_ax.plot(hist.history['val_acc'], 'g', label='Val - Accuracy')
loss_ax.set_xlabel('Epoch')
loss_ax.set_ylabel('Loss')
acc_ax.set_ylabel('Accuracy')
loss_ax.legend(loc='upper left')
acc_ax.legend(loc='lower left')
plt.show()
