05. Transfer Learning - Fine Tuning
(참고) udemy - TF Developer in 2022
Contents
- Fine Tuning 한 장 요약
- Model 0
- Data Augmentation을 모델에 넣기
- Model 1
- Model 2
- Model checkpoint
- Model 3 & 4
(1) Fine Tuning 한 장 요약
다양한 모델 생성
- Model 0: a transfer learning model using the Keras Functional API
- Model 1: a feature extraction transfer learning model on 1% of the data with data augmentation
- Model 2: a feature extraction transfer learning model on 10% of the data with data augmentation
- Model 3: a fine-tuned transfer learning model on 10% of the data
- Model 4: a fine-tuned transfer learning model on 100% of the data
(2) Model 0
Model 0: a transfer learning model using the Keras Functional API
- Data 불러오기
IMG_SIZE = (224, 224) # define image size
train_data_10_percent = tf.keras.preprocessing.image_dataset_from_directory(directory=train_dir,
image_size=IMG_SIZE,
label_mode="categorical",
batch_size=32)
- Base Model 불러오기
- trainable=False : FE 부분 전부 freeze할 것
- include_top=False : custom output layer 할 것
base_model = tf.keras.applications.EfficientNetB0(include_top=False)
base_model.trainable = False
- layer 쌓기
inputs = tf.keras.layers.Input(shape=(224, 224, 3), name="input_layer")
#----------------------------------------------------------------------#
x = base_model(inputs)
x = tf.keras.layers.GlobalAveragePooling2D(name="global_average_pooling_layer")(x)
#----------------------------------------------------------------------#
outputs = tf.keras.layers.Dense(10, activation="softmax", name="output_layer")(x)
- 모델 생성 & 학습
tf.keras.Model를 사용하여, input & output을 지정해줄 수 있다.
model_0 = tf.keras.Model(inputs, outputs)
model_0.compile(loss='categorical_crossentropy',
optimizer=tf.keras.optimizers.Adam(),
metrics=["accuracy"])
history_10_percent = model_0.fit(train_data_10_percent,
epochs=5,
steps_per_epoch=len(train_data_10_percent),
validation_data=test_data_10_percent,
validation_steps=int(0.25 * len(test_data_10_percent)),
callbacks=[create_tensorboard_callback("transfer_learning",
"10_percent_feature_extract")])
- 모델 summary 확인하기
model_0.summary()
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_layer (InputLayer) [(None, 224, 224, 3)] 0
_________________________________________________________________
efficientnetb0 (Functional) (None, None, None, 1280) 4049571
_________________________________________________________________
global_average_pooling_layer (None, 1280) 0
_________________________________________________________________
output_layer (Dense) (None, 10) 12810
=================================================================
Total params: 4,062,381
Trainable params: 12,810
Non-trainable params: 4,049,571
_________________________________________________________________
참고 : Global Average Pooling 이후 shape의 변화 :
-
input : (1, 4, 4, 3)
-
output : (1, 3)
아래의 둘은 동일하다!
tf.reduce_mean(input_tensor, axis=[1, 2])
tf.keras.layers.GlobalAveragePooling2D()(input_tensor)
(3) Data Augmentation을 모델에 넣기
( ImageDataGenerator 클래스를 사용하지 않고 )
tf.keras.layers.experimental.preprocessing 모듈을 사용해서!
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras.layers.experimental import preprocessing
data_augmentation = tf.keras.Sequential([
preprocessing.RandomFlip("horizontal"),
preprocessing.RandomRotation(0.2),
preprocessing.RandomZoom(0.2),
preprocessing.RandomHeight(0.2),
preprocessing.RandomWidth(0.2),
], name ="data_augmentation")
(4) Model 1
Model 1: a feature extraction transfer learning model on 1% of the data with data augmentation
- Data 불러오기
IMG_SIZE = (224, 224)
train_data_1_percent = tf.keras.preprocessing.image_dataset_from_directory(train_dir_1_percent,
label_mode="categorical",
batch_size=32,
image_size=IMG_SIZE)
- Base Model 불러오기
base_model = tf.keras.applications.EfficientNetB0(include_top=False)
base_model.trainable = False
- layer 쌓기
input_shape = (224, 224, 3)
inputs = layers.Input(shape=input_shape, name="input_layer")
#----------------------------------------------------------------------#
x = data_augmentation(inputs)
x = base_model(x, training=False)
x = layers.GlobalAveragePooling2D(name="global_average_pooling_layer")(x)
#----------------------------------------------------------------------#
outputs = layers.Dense(10, activation="softmax", name="output_layer")(x)
- 모델 생성 & 학습
model_1 = keras.Model(inputs, outputs)
model_1.compile(loss="categorical_crossentropy",
optimizer=tf.keras.optimizers.Adam(),
metrics=["accuracy"])
history_1_percent = model_1.fit(train_data_1_percent,
epochs=5,
steps_per_epoch=len(train_data_1_percent),
validation_data=test_data,
validation_steps=int(0.25* len(test_data)),
callbacks=[create_tensorboard_callback("transfer_learning", "1_percent_data_aug")])
(4) Model 2
Model 2: a feature extraction transfer learning model on 10% of the data with data augmentation
- 데이터 불러오기
IMG_SIZE = (224, 224)
train_data_10_percent = tf.keras.preprocessing.image_dataset_from_directory(train_dir_10_percent,
label_mode="categorical",
image_size=IMG_SIZE)
나머지는 위의 (3) Model 1 과 동일
(5) Model Checkpoint
checkpoint_path = "ten_percent_model_checkpoints_weights/checkpoint.ckpt"
checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_path,
save_weights_only=True,
save_best_only=False,
save_freq="epoch", # save every epoch
verbose=1)
이렇게 생성한 checkpoint를 callback안에 넣어준다.
(5) Model 3 & 4
base_model.trainable = True
for layer in base_model.layers[:-10]:
layer.trainable = False
이 외에는 전부 동일!