Callback functions for Trainers ?

[AxKo]

Hi Everybody,

I recently discovered skrl and I like the coding style very much.
I wonder if it is possible to use callbacks with SequentialTrainer (I didn't find any documentation) ?
Specifically I like to write my own callback for progress monitoring (I don't like Tensorboard so much).

Many thanks

[Toni-SM]

Hi @AxKo

Currently there are no callbacks in the library implementation.
However, a possible solution for a custom progress monitoring is to overwrite the Agent.write_tracking_data (see implementation) method to handle the self.tracking_data content according to the desired implementation.

from skrl.agents.torch.ppo import PPO

class DerivedPPO(PPO):
    def write_tracking_data(self, timestep: int, timesteps: int) -> None:
        # process data for custom progress monitoring
        #===================================
        for k, v in self.tracking_data.items():
            pass
        #===================================
            
        # reset data containers for next iteration
        self._track_rewards.clear()
        self._track_timesteps.clear()
        self.tracking_data.clear()

# ...

agent = DerivedPPO(...)