In some cases, it'd be beneficial to have hooks in the trainer loop. PFRL already provides step_hooks and global_step_hooks in train_agent and train_agent_async respectively, but they are not for learners. In particular, the latter global_step_hook works in actors and there's no hook mechanism in learners.
This PR provides optimize_hook in agents under pfrl.agents package. Each agent takes optimize_hooks, which is a list of callable objects, as a parameter of their constructors. In the learner loop, the hooks are invoked just after optimizer.step() method is called.
In some cases, it'd be beneficial to have hooks in the trainer loop. PFRL already provides
step_hooksandglobal_step_hooksintrain_agentandtrain_agent_asyncrespectively, but they are not for learners. In particular, the latterglobal_step_hookworks in actors and there's no hook mechanism in learners.This PR provides
optimize_hookin agents underpfrl.agentspackage. Each agent takesoptimize_hooks, which is a list of callable objects, as a parameter of their constructors. In the learner loop, the hooks are invoked just afteroptimizer.step()method is called.