迁移学习 colab 完整示例:fruits-360 数据集

这里当前目录下已经有fruits-360这个数据集. 关于调用数据集的方法可以查看我另一篇文章.

准备

import tensorflow as tf
import tensorflow.keras as keras
from tensorflow.keras.preprocessing.image import load_img, img_to_array, array_to_img, ImageDataGenerator

创建 Generator

创建 ImageDataGenerator. 由于这个数据集足够大, 所以不需要进行 image augmentation.

train_datagen = ImageDataGenerator(rescale=1./255)
test_datagen = ImageDataGenerator(rescale=1./255)
train_generator = train_datagen.flow_from_directory(
        "fruits-360/Training",
        target_size=(100, 100),
        batch_size=32,
        class_mode='categorical')

validation_generator = test_datagen.flow_from_directory(
    "fruits-360/Test",
    target_size=(100, 100),
    batch_size=32,
    class_mode='categorical')

运行后看到如下输出表示创建成功.

Found 67692 images belonging to 131 classes.
Found 22688 images belonging to 131 classes.

模型

这里使用的是 Xception 模型.

from tensorflow.keras.applications.xception import preprocess_input
from tensorflow.keras.applications.xception import decode_predictions
from tensorflow.keras.applications.xception import Xception

tf.keras.backend.clear_session()
base_model = tf.keras.applications.Xception(
    weights='imagenet',  # Load weights pre-trained on ImageNet.
    input_shape=(100, 100, 3),
    include_top=False)  # Do not include the ImageNet classifier at the top.
input_layer = tf.keras.Input(shape=(100, 100, 3))
base_model.trainable = False
# x = data_augmentation(input_layer)
x = base_model(input_layer, training = False)
x = tf.keras.layers.GlobalAveragePooling2D()(x)
x = tf.keras.layers.Dense(64, activation = 'relu')(x)
x = tf.keras.layers.Dropout(0.2)(x)  # Regularize with dropout
output_layer = tf.keras.layers.Dense(131, activation = 'softmax')(x)
model = tf.keras.Model(input_layer, output_layer)
model.summary()

base_model.trainable = False将会冻结 Xception 模型的权重, 在训练中不会被更新. 即使用已经训练好的权重.

x = base_model(input_layer, training = False) 中training=False可以确保 base模型处于 Inference phase, 而不是Training phase.

opt = tf.keras.optimizers.Adam(learning_rate=0.01)
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics = ['accuracy'])

model.fit(train_generator, epochs=5, steps_per_epoch = 67692//32, validation_data=validation_generator)