keras tune 使用

安装

pip install keras-tuner [–upgrade]

import kerastuner as kt

from tensorflow import keras

写函数的时候使用hp来指代超参，

tuner = kt.RandomSearch(bulid_model,objective=‘val_loss’,max_trial=5)

上面使用随机搜索通过检测验证损失来尝试五个不同模型

  1. tuner.search(x_train, y_train, epochs=5, validation_data=(x_val, y_val))
  2. best_model = tuner.get_best_models()[0]

搜索模型超参，按照目标度量来进行排序

  1. tuner.search_space_summary()

查看参数搜索空间

  1. 例子
  2. def build_model(hp):
  3.     model = keras.Sequential()
  4.     model.add(layers.Flatten())
  5.     for i in range(hp.Int("num_layers", 2, 20)):
  6.         model.add(
  7.             layers.Dense(
  8.                 units=hp.Int("units_" + str(i), min_value=32, max_value=512, step=32),
  9.                 activation="relu",
  10.             )
  11.         )
  12.     model.add(layers.Dense(10, activation="softmax"))
  13.     model.compile(
  14.         optimizer=keras.optimizers.Adam(hp.Choice("learning_rate", [1e-2, 1e-3, 1e-4])),
  15.         loss="categorical_crossentropy",
  16.         metrics=["accuracy"],
  17.     )
  18.     return model
  1. 使用HyperModel类的子类来
  2. from keras_tuner import HyperModel
  5. class MyHyperModel(HyperModel):
  6.     def __init__(self, classes):
  7.         self.classes = classes
  9.     def build(self, hp):
  10.         model = keras.Sequential()
  11.         model.add(layers.Flatten())
  12.         model.add(
  13.             layers.Dense(
  14.                 units=hp.Int("units", min_value=32, max_value=512, step=32),
  15.                 activation="relu",
  16.             )
  17.         )
  18.         model.add(layers.Dense(self.classes, activation="softmax"))
  19.         model.compile(
  20.             optimizer=keras.optimizers.Adam(
  21.                 hp.Choice("learning_rate", values=[1e-2, 1e-3, 1e-4])
  22.             ),
  23.             loss="categorical_crossentropy",
  24.             metrics=["accuracy"],
  25.         )
  26.         return model
  29. hypermodel = MyHyperModel(classes=10)
  31. tuner = RandomSearch(
  32.     hypermodel,
  33.     objective="val_accuracy",
  34.     max_trials=3,
  35.     overwrite=True,
  36.     directory="my_dir",
  37.     project_name="helloworld",
  38. )
  40. tuner.search(x_train, y_train, epochs=2, validation_data=(x_val, y_val))

只要重写build方法就能实现很好的模型共享和复用

主要是hp.Int（name,min_value,max_value,step,default）

和hp.choice(name,value(这应该是个可以迭代的对象))

hp = HyperParameters()