def log_diagnostics(self, itr, sampler_itr, throttle_time, prefix='Diagnostics/'):
with logger.tabular_prefix(prefix):
logger.record_tabular('CumCompletedTrajs', self._cum_completed_trajs)
logger.record_tabular('NewCompletedTrajs', self._new_completed_trajs)
logger.record_tabular('StepsInTrajWindow',
sum(info["Length"] for info in self._traj_infos))
super().log_diagnostics(itr, sampler_itr, throttle_time, prefix=prefix)
self._new_completed_trajs = 0
def log_diagnostics(self, itr, sampler_itr, throttle_time, prefix='Diagnostics/'):
if not self._traj_infos:
logger.log("WARNING: had no complete trajectories in eval.")
steps_in_eval = sum([info["Length"] for info in self._traj_infos])
with logger.tabular_prefix(prefix):
logger.record_tabular('StepsInEval', steps_in_eval)
logger.record_tabular('TrajsInEval', len(self._traj_infos))
logger.record_tabular('CumEvalTime', self.ctrl.eval_time.value)
super().log_diagnostics(itr, sampler_itr, throttle_time, prefix=prefix)
self._traj_infos = list() # Clear after each eval.
def log_diagnostics(self,
itr,
traj_infos=None,
eval_time=0,
prefix='Diagnostics/'):
"""
Write diagnostics (including stored ones) to csv via the logger.
"""
if itr > 0:
self.pbar.stop()
self.save_itr_snapshot(itr)
new_time = time.time()
self._cum_time = new_time - self._start_time
train_time_elapsed = new_time - self._last_time - eval_time
new_updates = self.algo.update_counter - self._last_update_counter
new_samples = (self.sampler.batch_size * self.world_size *
self.log_interval_itrs)
updates_per_second = (float('nan') if itr == 0 else new_updates /
train_time_elapsed)
samples_per_second = (float('nan') if itr == 0 else new_samples /
train_time_elapsed)
replay_ratio = (new_updates * self.algo.batch_size * self.world_size /
new_samples)
cum_replay_ratio = (self.algo.batch_size * self.algo.update_counter /
((itr + 1) * self.sampler.batch_size)
) # world_size cancels.
cum_steps = (itr + 1) * self.sampler.batch_size * self.world_size
with logger.tabular_prefix(prefix):
if self._eval:
logger.record_tabular(
'CumTrainTime', self._cum_time -
self._cum_eval_time) # Already added new eval_time.
logger.record_tabular('Iteration', itr)
logger.record_tabular('CumTime (s)', self._cum_time)
logger.record_tabular('CumSteps', cum_steps)
logger.record_tabular('CumCompletedTrajs',
self._cum_completed_trajs)
logger.record_tabular('CumUpdates', self.algo.update_counter)
logger.record_tabular('StepsPerSecond', samples_per_second)
logger.record_tabular('UpdatesPerSecond', updates_per_second)
logger.record_tabular('ReplayRatio', replay_ratio)
logger.record_tabular('CumReplayRatio', cum_replay_ratio)
self._log_infos(traj_infos)
logger.dump_tabular(with_prefix=False)
self._last_time = new_time
self._last_update_counter = self.algo.update_counter
if itr < self.n_itr - 1:
logger.log(f"Optimizing over {self.log_interval_itrs} iterations.")
self.pbar = ProgBarCounter(self.log_interval_itrs)
def log_diagnostics(self,
itr,
eval_traj_infos,
eval_time,
prefix='Diagnostics/'):
if not eval_traj_infos:
logger.log("WARNING: had no complete trajectories in eval.")
steps_in_eval = sum([info["Length"] for info in eval_traj_infos])
with logger.tabular_prefix(prefix):
logger.record_tabular('StepsInEval', steps_in_eval)
logger.record_tabular('TrajsInEval', len(eval_traj_infos))
self._cum_eval_time += eval_time
logger.record_tabular('CumEvalTime', self._cum_eval_time)
super().log_diagnostics(itr, eval_traj_infos, eval_time, prefix=prefix)
def log_diagnostics(self, itr, sampler_itr, throttle_time, prefix='Diagnostics/'):
self.pbar.stop()
self.save_itr_snapshot(itr, sampler_itr)
new_time = time.time()
time_elapsed = new_time - self._last_time
new_updates = self.algo.update_counter - self._last_update_counter
new_samples = self.sampler.batch_size * (sampler_itr - self._last_sampler_itr)
updates_per_second = (float('nan') if itr == 0 else
new_updates / time_elapsed)
samples_per_second = (float('nan') if itr == 0 else
new_samples / time_elapsed)
if self._eval:
new_eval_time = self.ctrl.eval_time.value
eval_time_elapsed = new_eval_time - self._last_eval_time
non_eval_time_elapsed = time_elapsed - eval_time_elapsed
non_eval_samples_per_second = (float('nan') if itr == 0 else
new_samples / non_eval_time_elapsed)
self._last_eval_time = new_eval_time
cum_steps = sampler_itr * self.sampler.batch_size # No * world_size.
replay_ratio = (new_updates * self.algo.batch_size * self.world_size /
max(1, new_samples))
cum_replay_ratio = (self.algo.update_counter * self.algo.batch_size *
self.world_size / max(1, cum_steps))
with logger.tabular_prefix(prefix):
logger.record_tabular('Iteration', itr)
logger.record_tabular('SamplerIteration', sampler_itr)
logger.record_tabular('CumTime (s)', new_time - self._start_time)
logger.record_tabular('CumSteps', cum_steps)
logger.record_tabular('CumUpdates', self.algo.update_counter)
logger.record_tabular('ReplayRatio', replay_ratio)
logger.record_tabular('CumReplayRatio', cum_replay_ratio)
logger.record_tabular('StepsPerSecond', samples_per_second)
if self._eval:
logger.record_tabular('NonEvalSamplesPerSecond', non_eval_samples_per_second)
logger.record_tabular('UpdatesPerSecond', updates_per_second)
logger.record_tabular('OptThrottle', (time_elapsed - throttle_time) /
time_elapsed)
self._log_infos()
self._last_time = new_time
self._last_itr = itr
self._last_sampler_itr = sampler_itr
self._last_update_counter = self.algo.update_counter
logger.dump_tabular(with_prefix=False)
logger.log(f"Optimizing over {self.log_interval_itrs} sampler "
"iterations.")
self.pbar = ProgBarCounter(self.log_interval_itrs)