Keras迁移学习完整范例

Keras迁移学习完整范例:

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

# 基于预训练模型Xception的特征提取层，创建新的模型
# The default input image size for the Xception model is 299x299
base_model = tf.keras.applications.xception.Xception(
                include_top=False,
                weights="imagenet",
                input_shape=(299,299,3)
)
# 冻结该模型
base_model.trainable = False

# 载入训练和测试数据集
# 准备数据
TRAIN_DATASET_PATH = r"D:\nn_tf\cats_vs_dogs\train" #训练数据集路径
TEST_DATASET_PATH  = r"D:\nn_tf\cats_vs_dogs\test"  #测试数据集路径
batch_size = 32
image_size = (299,299)
# 训练数据集
train_dataset = keras.preprocessing.image_dataset_from_directory(
    TRAIN_DATASET_PATH,
    validation_split=0.2,
    image_size=image_size,
    seed=1337,
    subset='training',
    batch_size=batch_size
)
# 验证数据集
val_dataset = keras.preprocessing.image_dataset_from_directory(
    TRAIN_DATASET_PATH,
    validation_split=0.2,
    subset="validation",
    seed=1337,
    image_size=image_size,
    batch_size=batch_size,
)
# 测试数据集
test_dataset = keras.preprocessing.image_dataset_from_directory(
    TEST_DATASET_PATH,
    image_size=image_size,
    shuffle=False,
    batch_size=batch_size
)

# 查看数据形状
for image, label in train_dataset.take(1):
    print(image.shape)


# 设计数据增强算法
data_augmentation = tf.keras.Sequential([
  layers.RandomFlip("horizontal_and_vertical"),
  layers.RandomRotation(0.2),
])

# 将数据增强作用到训练数据集上
train_dataset = train_dataset.map(lambda x, y: (data_augmentation(x, training=True), y))
train_dataset = train_dataset.prefetch(buffer_size=tf.data.AUTOTUNE)

# 查看数据增强后的数据形状
for image, label in train_dataset.take(1):
    print(image.shape)


# 添加Xception model的预处理层
xception_preprocess = keras.applications.xception.preprocess_input
# 创建带有预处理的新模型
inputs = keras.Input(shape=(299,299,3), name="image_input")
x = xception_preprocess(inputs)
x = base_model(x, training=False)
x = keras.layers.GlobalAveragePooling2D()(x) #将特征值展平成向量
outputs = keras.layers.Dense(1, name="precisions")(x)
model_with_prepocess = keras.Model(inputs, outputs, name="model_with_prepocess")

# 编译模型
model_with_prepocess.compile(optimizer=keras.optimizers.Adam(),
              loss=keras.losses.BinaryCrossentropy(from_logits=True),
              metrics=[keras.metrics.BinaryAccuracy()])
# 训练模型
model_with_prepocess.fit(train_dataset, epochs=10, validation_data=val_dataset)

# 在测试数据集上测试模型
loss, acc = model_with_prepocess.evaluate(test_dataset)
print(f"Loss is {loss}; Accuracy is {acc} in Test Dataset")