UtilsRL.monitor

Monitor is designed to like a manager to monitor the progress of the experiment. For basic usage, it can be used as tqdm.tqdm, which wraps an iterable object and prints a progress bar on the screen. We enhanced Monitor with some more features on top of this, namely

• Context Variable Registration. Sometimes we hope to save products (like model checkpoint) or temporary states (like replay buffer) for reproducibility and staged training, and then Monitor may provide some help. You can register a context variable with register_context(), then the context variables will be saved at an interval of save_every. You can also resume training with such method in the following snippet, where we defined a class Trainer, load actor and local_var from path /path/to/ck_dir and resume training from the initial-th iteration.

class Trainer():
    def __init__(self):
        self.actor = ...

    def train(self):
        local_var = ...

        # load previous saved checkpoints specified by `load_path`
        self.actor, local_var = \
            monitor.register_context(["self.actor", "local_var"], load_path="/path/to/ck_dir").values()
        # use `initial` to designate the start point
        for i_epoch in monitor.listen(range(1000), initial=100):
            # continue training

• Callback Function Registration. You can register callback functions which will be triggered at certain state of the training with register_callback(). For example, a function which will email us when training is completed.

monitor = Monitor(desc="test_monitor")
monitor.register_callback(
    name= "email me at the end of training",
    on = "exit",
    callback = Monitor.email,
    ...
)

UtilsRL.monitor.Monitor