tf-livepatch-lr 0.0.5

Simple project which allows you to change the learning rate schedule without stopping the training process (tensorflow only)

Tensorflow learning rate patcher

Description

This is a simple project which implements one feature: it allows you to change the learning rate schedule without stopping the training process by editing the specified file. The file is just a json containing the output of the tensorflow.keras.optimizers.schedules.deserialize method, for example:

{
  "class_name": "PolynomialDecay",
  "config":{
    "cycle": false,
    "decay_steps": 1000,
    "end_learning_rate": 0.01,
    "initial_learning_rate": 0.1,
    "power": 0.5
  }
}

Custom learning rate schedules are also supported. In order to use them pass custom_objects to LiveLrSchedule constructor as you'd do for tensorflow.keras.optimizers.schedules.deserialize.