def _train(self, env, policy, pool): """Perform RL training. Args: env (`rllab.Env`): Environment used for training policy (`Policy`): Policy used for training pool (`PoolBase`): Sample pool to add samples to """ self._init_training() self.sampler.initialize(env, policy, pool) evaluation_env = deep_clone(env) if self._eval_n_episodes else None with tf_utils.get_default_session().as_default(): gt.rename_root('RLAlgorithm') gt.reset() gt.set_def_unique(False) for epoch in gt.timed_for(range(self._n_epochs + 1), save_itrs=True): logger.push_prefix('Epoch #%d | ' % epoch) for t in range(self._epoch_length): self.sampler.sample() if not self.sampler.batch_ready(): continue gt.stamp('sample') for i in range(self._n_train_repeat): self._do_training(iteration=t + epoch * self._epoch_length, batch=self.sampler.random_batch()) gt.stamp('train') self._evaluate(policy, evaluation_env) gt.stamp('eval') params = self.get_snapshot(epoch) logger.save_itr_params(epoch, params) time_itrs = gt.get_times().stamps.itrs time_eval = time_itrs['eval'][-1] time_total = gt.get_times().total time_train = time_itrs.get('train', [0])[-1] time_sample = time_itrs.get('sample', [0])[-1] logger.record_tabular('time-train', time_train) logger.record_tabular('time-eval', time_eval) logger.record_tabular('time-sample', time_sample) logger.record_tabular('time-total', time_total) logger.record_tabular('epoch', epoch) self.sampler.log_diagnostics() logger.dump_tabular(with_prefix=False) logger.pop_prefix() self.sampler.terminate()
def _train(self, env, policy, pool): """Perform RL training. Args: env (`rllab.Env`): Environment used for training policy (`Policy`): Policy used for training pool (`PoolBase`): Sample pool to add samples to """ self._init_training() self.sampler.initialize(env, policy, pool) evaluation_env = deep_clone(env) if self._eval_n_episodes else None with tf_utils.get_default_session().as_default(): gt.rename_root('RLAlgorithm') gt.reset() gt.set_def_unique(False) for epoch in gt.timed_for( range(self._n_epochs + 1), save_itrs=True): logger.push_prefix('Epoch #%d | ' % epoch) for t in range(self._epoch_length): self.sampler.sample() if not self.sampler.batch_ready(): continue gt.stamp('sample') for i in range(self._n_train_repeat): self._do_training( iteration=t + epoch * self._epoch_length, batch=self.sampler.random_batch()) gt.stamp('train') self._evaluate(policy, evaluation_env) gt.stamp('eval') params = self.get_snapshot(epoch) logger.save_itr_params(epoch, params) time_itrs = gt.get_times().stamps.itrs time_eval = time_itrs['eval'][-1] time_total = gt.get_times().total time_train = time_itrs.get('train', [0])[-1] time_sample = time_itrs.get('sample', [0])[-1] logger.record_tabular('time-train', time_train) logger.record_tabular('time-eval', time_eval) logger.record_tabular('time-sample', time_sample) logger.record_tabular('time-total', time_total) logger.record_tabular('epoch', epoch) self.sampler.log_diagnostics() logger.dump_tabular(with_prefix=False) logger.pop_prefix()
def _init_training(self, env, policy, pool): """Method to be called at the start of training. :param env: Environment instance. :param policy: Policy instance. :return: None """ self.env = env if self._eval_n_episodes > 0: self._eval_env = deep_clone(env) self.policy = policy self.pool = pool
def _init_training(self, env, policy, pool): """Method to be called at the start of training. :param env: Environment instance. :param policy: Policy instance. :return: None """ self.env = env if self._eval_n_episodes > 0: self._eval_env = deep_clone(env) self.policy = policy self.pool = pool