def _run_env_eval(self, step, do_sampler_step=True, calculate_holdout=True):
timeit.start('eval')
### calculate holdout costs
self._policy.eval_holdout()
timeit.stop('eval')
def _run_env_eval(self,
step,
do_sampler_step=True,
calculate_holdout=True):
timeit.start('eval')
### add to eval buffer
if self._sampler_eval and do_sampler_step:
self._sampler_eval.reset()
eval_step = step
num_dones = 0
while num_dones < self._rollouts_per_eval:
_, _, _, _, done, _ = \
self._sampler_eval.step(eval_step, explore=False)
eval_step += 1
num_dones += int(done)
self._sampler.reset()
### calculate holdout costs
if self._replay_pool_eval.can_sample(
batch_size=self._batch_size) and calculate_holdout:
indices, weights, steps, observations, goals, actions, rewards, dones, _ = \
self._replay_pool_eval.sample(self._batch_size)
self._policy.eval_holdout(step,
steps=steps,
observations=observations,
goals=goals,
actions=actions,
rewards=rewards,
dones=dones)
timeit.stop('eval')
def _run_env_step(self, step):
timeit.start('sample')
self._sampler.step(
step,
take_random_actions=(step < self._onpolicy_after_n_steps),
explore=True)
timeit.stop('sample')
return step
def _run_init(self):
self._restore()
# note this is the right step, but the trajectories might not all have been saved
save_itr = self._fm.get_train_itr()
start_step = save_itr * self._save_every_n_steps
self._sampler.reset()
timeit.reset()
timeit.start('total')
return start_step, save_itr
def _run_init_inference(self):
inference_itr = self._fm.get_inference_itr()
self._restore_rollouts('train')
self._restore_rollouts('eval')
save_itr = inference_itr
start_step = save_itr * self._save_every_n_steps
timeit.reset()
timeit.start('total')
return start_step, save_itr
def _run_init_train(self):
train_itr = self._fm.get_train_itr()
if train_itr > 0:
logger.info('Restore train iteration {0}'.format(train_itr - 1))
self._policy.restore(self._fm.train_policy_fname(train_itr - 1), train=True)
save_itr = train_itr
start_step = save_itr * self._save_every_n_steps
timeit.reset()
timeit.start('total')
return start_step, save_itr
def _run_init_train(self):
train_itr = self._fm.get_train_itr()
if train_itr > 0:
self._policy.restore(self._fm.train_policy_fname(train_itr - 1),
train=True)
save_itr = train_itr
start_step = save_itr * self._save_every_n_steps
timeit.reset()
timeit.start('total')
return start_step, save_itr
def train(self):
self._start_train_batch()
logger.info('Training model')
alg_args = self._params['alg']
total_steps = int(alg_args['total_steps'])
save_every_n_steps = int(alg_args['save_every_n_steps'])
update_target_after_n_steps = int(
alg_args['update_target_after_n_steps'])
update_target_every_n_steps = int(
alg_args['update_target_every_n_steps'])
log_every_n_steps = int(alg_args['log_every_n_steps'])
timeit.reset()
timeit.start('total')
save_itr = 0
for step in range(total_steps):
timeit.start('sample')
# steps, observations, actions, rewards, dones, _ = self._replay_pool.sample(batch_size)
steps, observations, actions, rewards, dones, _ = self._batch_queue.get(
)
timeit.stop('sample')
timeit.start('train')
self._model.train_step(step,
steps=steps,
observations=observations,
actions=actions,
rewards=rewards,
dones=dones,
use_target=True)
timeit.stop('train')
### update target network
if step > update_target_after_n_steps and step % update_target_every_n_steps == 0:
self._model.update_target()
### log
if step > 0 and step % log_every_n_steps == 0:
logger.record_tabular('Step', step)
self._model.log()
logger.dump_tabular(print_func=logger.info)
timeit.stop('total')
for line in str(timeit).split('\n'):
logger.debug(line)
timeit.reset()
timeit.start('total')
### save model
if step > 0 and step % save_every_n_steps == 0:
logger.info('Saving files for itr {0}'.format(save_itr))
self._save_train_policy(save_itr)
save_itr += 1
### always save the end
self._save_train_policy(save_itr)
self._stop_train_batch()
def _run_log(self, step):
logger.record_tabular('Step', step)
self._env.log()
self._replay_pool.log()
if self._env_eval:
self._env_eval.log(prefix='Eval')
if self._replay_pool_eval:
self._replay_pool_eval.log(prefix='Eval')
self._policy.log()
logger.dump_tabular(print_func=logger.info)
timeit.stop('total')
for line in str(timeit).split('\n'):
logger.debug(line)
timeit.reset()
timeit.start('total')
def do_train_step():
timeit.start('batch')
indices, weights, steps, observations, goals, actions, rewards, dones, _ = \
self._replay_pool.sample(self._batch_size)
timeit.stop('batch')
timeit.start('train')
rew_errors = self._policy.train_step(step,
steps=steps,
observations=observations,
goals=goals,
actions=actions,
rewards=rewards,
dones=dones,
weights=weights)
self._replay_pool.update_priorities(indices, rew_errors)
timeit.stop('train')
def train(self):
### restore where we left off
save_itr = self._restore()
target_updated = False
eval_rollouts = []
self._sampler.reset()
if self._eval_sampler is not None:
self._eval_sampler.reset()
timeit.reset()
timeit.start('total')
for step in range(0, self._total_steps, self._sampler.n_envs):
### sample and add to buffer
if step > self._sample_after_n_steps:
timeit.start('sample')
self._sampler.step(
step,
take_random_actions=(step <= self._onpolicy_after_n_steps),
explore=True)
timeit.stop('sample')
### sample and DON'T add to buffer (for validation)
if self._eval_sampler is not None and step > 0 and step % self._eval_every_n_steps == 0:
timeit.start('eval')
for _ in range(self._rollouts_per_eval):
eval_rollouts_step = []
eval_step = step
while len(eval_rollouts_step) == 0:
self._eval_sampler.step(eval_step, explore=False)
eval_rollouts_step = self._eval_sampler.get_recent_paths(
)
eval_step += 1
eval_rollouts += eval_rollouts_step
timeit.stop('eval')
if step >= self._learn_after_n_steps:
### training step
if self._train_every_n_steps >= 1:
if step % int(self._train_every_n_steps) == 0:
timeit.start('batch')
steps, observations, goals, actions, rewards, dones, _ = \
self._sampler.sample(self._batch_size)
timeit.stop('batch')
timeit.start('train')
self._policy.train_step(step,
steps=steps,
observations=observations,
goals=goals,
actions=actions,
rewards=rewards,
dones=dones,
use_target=target_updated)
timeit.stop('train')
else:
for _ in range(int(1. / self._train_every_n_steps)):
timeit.start('batch')
steps, observations, goals, actions, rewards, dones, _ = \
self._sampler.sample(self._batch_size)
timeit.stop('batch')
timeit.start('train')
self._policy.train_step(step,
steps=steps,
observations=observations,
goals=goals,
actions=actions,
rewards=rewards,
dones=dones,
use_target=target_updated)
timeit.stop('train')
### update target network
if step > self._update_target_after_n_steps and step % self._update_target_every_n_steps == 0:
self._policy.update_target()
target_updated = True
### log
if step % self._log_every_n_steps == 0:
logger.record_tabular('Step', step)
self._sampler.log()
self._eval_sampler.log(prefix='Eval')
self._policy.log()
logger.dump_tabular(print_func=logger.info)
timeit.stop('total')
for line in str(timeit).split('\n'):
logger.debug(line)
timeit.reset()
timeit.start('total')
### save model
if step > 0 and step % self._save_every_n_steps == 0:
logger.info('Saving files for itr {0}'.format(save_itr))
self._save(save_itr, self._sampler.get_recent_paths(),
eval_rollouts)
save_itr += 1
eval_rollouts = []
self._save(save_itr, self._sampler.get_recent_paths(), eval_rollouts)
def inference(self):
### restore where we left off
self._restore_inference()
inference_itr = self._get_inference_itr()
inference_step = self._get_inference_step()
train_itr = self._get_train_itr()
self._run_rsync()
train_rollouts = []
eval_rollouts = []
self._inference_reset_sampler()
timeit.reset()
timeit.start('total')
while True:
train_step = self._get_train_step()
if inference_step > self._total_steps:
break
### sample and add to buffer
if inference_step > self._sample_after_n_steps:
timeit.start('sample')
inference_step = self._inference_step(inference_step)
timeit.stop('sample')
else:
inference_step += self._sampler.n_envs
### sample and DON'T add to buffer (for validation)
if self._eval_sampler is not None and inference_step > 0 and inference_step % self._eval_every_n_steps == 0:
timeit.start('eval')
eval_rollouts_step = []
eval_step = inference_step
while len(eval_rollouts_step) == 0:
self._eval_sampler.step(eval_step, explore=False)
eval_rollouts_step = self._eval_sampler.get_recent_paths()
eval_step += 1
eval_rollouts += eval_rollouts_step
timeit.stop('eval')
### log
if inference_step % self._log_every_n_steps == 0:
logger.info('train itr {0:04d} inference itr {1:04d}'.format(
train_itr, inference_itr))
logger.record_tabular('Train step', train_step)
logger.record_tabular('Inference step', inference_step)
self._sampler.log()
if self._eval_sampler:
self._eval_sampler.log(prefix='Eval')
logger.dump_tabular(print_func=logger.info)
timeit.stop('total')
for line in str(timeit).split('\n'):
logger.debug(line)
timeit.reset()
timeit.start('total')
### save rollouts / load model
train_rollouts += self._sampler.get_recent_paths()
if inference_step > 0 and inference_step % self._inference_save_every_n_steps == 0:
self._inference_reset_sampler()
### save rollouts
logger.debug('Saving files for itr {0}'.format(inference_itr))
self._save_inference(inference_itr, train_rollouts,
eval_rollouts)
inference_itr += 1
train_rollouts = []
eval_rollouts = []
### load model
with self._rsync_lock: # to ensure the ckpt has been fully transferred over
new_train_itr = self._get_train_itr()
if train_itr < new_train_itr:
logger.debug(
'Loading policy for itr {0}'.format(new_train_itr -
1))
try:
self._policy.restore(
self._inference_policy_file_name(
new_train_itr - 1),
train=False)
train_itr = new_train_itr
except:
logger.debug(
'Failed to load model for itr {0}'.format(
new_train_itr - 1))
self._policy.restore(
self._inference_policy_file_name(train_itr -
1),
train=False)
logger.debug('As backup, restored itr {0}'.format(
train_itr - 1))
self._save_inference(inference_itr, self._sampler.get_recent_paths(),
eval_rollouts)
def train(self):
### restore where we left off
init_inference_step = len(self._sampler) # don't count offpolicy
self._restore_train()
train_itr = self._get_train_itr()
train_step = self._get_train_step()
inference_itr = self._get_inference_itr()
target_updated = False
timeit.reset()
timeit.start('total')
while True:
inference_step = len(self._sampler) - init_inference_step
if inference_step > self._total_steps or train_step > self._train_total_steps:
break
if inference_step >= self._learn_after_n_steps:
### training step
train_step += 1
timeit.start('batch')
steps, observations, goals, actions, rewards, dones, _ = \
self._sampler.sample(self._batch_size)
timeit.stop('batch')
timeit.start('train')
self._policy.train_step(train_step,
steps=steps,
observations=observations,
goals=goals,
actions=actions,
rewards=rewards,
dones=dones,
use_target=target_updated)
timeit.stop('train')
### update target network
if train_step > self._update_target_after_n_steps and train_step % self._update_target_every_n_steps == 0:
self._policy.update_target()
target_updated = True
### log
if train_step % self._log_every_n_steps == 0:
logger.info(
'train itr {0:04d} inference itr {1:04d}'.format(
train_itr, inference_itr))
logger.record_tabular('Train step', train_step)
logger.record_tabular('Inference step', inference_step)
self._policy.log()
logger.dump_tabular(print_func=logger.info)
timeit.stop('total')
for line in str(timeit).split('\n'):
logger.debug(line)
timeit.reset()
timeit.start('total')
else:
time.sleep(1)
### save model
if train_step > 0 and train_step % self._train_save_every_n_steps == 0:
logger.debug('Saving files for itr {0}'.format(train_itr))
self._save_train(train_itr)
train_itr += 1
### reset model
if train_step > 0 and self._train_reset_every_n_steps is not None and \
train_step % self._train_reset_every_n_steps == 0:
logger.debug('Resetting model')
self._policy.reset_weights()
### load data
inference_itr = self._train_load_data(inference_itr)
def inference(self):
### restore where we left off
self._restore_inference()
inference_itr = self._get_inference_itr()
inference_step = self._get_inference_step()
train_itr = self._get_train_itr()
self._run_rsync()
assert (self._eval_sampler is None) # TODO: temporary
train_rollouts = []
eval_rollouts = []
self._reset_sampler()
timeit.reset()
timeit.start('total')
while True:
train_step = self._get_train_step()
if inference_step > self._total_steps:
break
### sample and add to buffer
if inference_step > self._sample_after_n_steps:
timeit.start('sample')
try:
self._sampler.step(
inference_step,
take_random_actions=(
inference_step <= self._learn_after_n_steps
or inference_step <= self._onpolicy_after_n_steps),
explore=True)
inference_step += self._sampler.n_envs
except Exception as e:
logger.warn('Sampler exception {0}'.format(str(e)))
trashed_steps = self._sampler.trash_current_rollouts()
inference_step -= trashed_steps
logger.warn('Trashed {0} steps'.format(trashed_steps))
while not self._env.ros_is_good(
print=False): # TODO hard coded
time.sleep(0.25)
self._reset_sampler()
logger.warn('Continuing...')
timeit.stop('sample')
else:
inference_step += self._sampler.n_envs
### sample and DON'T add to buffer (for validation)
if self._eval_sampler is not None and inference_step > 0 and inference_step % self._eval_every_n_steps == 0:
timeit.start('eval')
eval_rollouts_step = []
eval_step = inference_step
while len(eval_rollouts_step) == 0:
self._eval_sampler.step(eval_step, explore=False)
eval_rollouts_step = self._eval_sampler.get_recent_paths()
eval_step += 1
eval_rollouts += eval_rollouts_step
timeit.stop('eval')
### log
if inference_step % self._log_every_n_steps == 0:
logger.info('train itr {0:04d} inference itr {1:04d}'.format(
train_itr, inference_itr))
logger.record_tabular('Train step', train_step)
logger.record_tabular('Inference step', inference_step)
self._sampler.log()
if self._eval_sampler:
self._eval_sampler.log(prefix='Eval')
logger.dump_tabular(print_func=logger.info)
timeit.stop('total')
for line in str(timeit).split('\n'):
logger.debug(line)
timeit.reset()
timeit.start('total')
### save rollouts / load model
train_rollouts += self._sampler.get_recent_paths()
if inference_step > 0 and inference_step % self._inference_save_every_n_steps == 0 and \
len(train_rollouts) > 0:
response = input('Keep rollouts?')
if response != 'y':
train_rollouts = []
continue
### reset to stop rollout
self._sampler.reset()
### save rollouts
logger.debug('Saving files for itr {0}'.format(inference_itr))
self._save_inference(inference_itr, train_rollouts,
eval_rollouts)
inference_itr += 1
train_rollouts = []
eval_rollouts = []
### load model
with self._rsync_lock: # to ensure the ckpt has been fully transferred over
new_train_itr = self._get_train_itr()
if train_itr < new_train_itr:
logger.debug(
'Loading policy for itr {0}'.format(new_train_itr -
1))
try:
self._policy.restore(
self._inference_policy_file_name(
new_train_itr - 1),
train=False)
train_itr = new_train_itr
except:
logger.debug(
'Failed to load model for itr {0}'.format(
new_train_itr - 1))
self._policy.restore(
self._inference_policy_file_name(train_itr -
1),
train=False)
logger.debug('As backup, restored itr {0}'.format(
train_itr - 1))
self._save_inference(inference_itr, self._sampler.get_recent_paths(),
eval_rollouts)
def do_train_step():
timeit.start('train')
self._policy.train_step(step)
timeit.stop('train')
def train(self):
### restore where we left off
self._restore_train()
train_itr = self._get_train_itr()
train_step = self._get_train_step()
inference_itr = self._get_inference_itr()
init_inference_step = len(self._sampler)
target_updated = False
timeit.reset()
timeit.start('total')
while True:
inference_step = len(self._sampler) - init_inference_step
if inference_step > self._total_steps:
break
if inference_step >= self._learn_after_n_steps:
### update preprocess
if train_step % self._update_preprocess_every_n_steps == 0:
self._policy.update_preprocess(self._sampler.statistics)
### training step
train_step += 1
timeit.start('batch')
batch = self._sampler.sample(self._batch_size)
timeit.stop('batch')
timeit.start('train')
self._policy.train_step(train_step,
*batch,
use_target=target_updated)
timeit.stop('train')
### update target network
if train_step > self._update_target_after_n_steps and train_step % self._update_target_every_n_steps == 0:
self._policy.update_target()
target_updated = True
### log
if train_step % self._log_every_n_steps == 0:
logger.info(
'train itr {0:04d} inference itr {1:04d}'.format(
train_itr, inference_itr))
logger.record_tabular('Train step', train_step)
logger.record_tabular('Inference step', inference_step)
self._policy.log()
logger.dump_tabular(print_func=logger.info)
timeit.stop('total')
for line in str(timeit).split('\n'):
logger.debug(line)
timeit.reset()
timeit.start('total')
else:
time.sleep(1)
### save model
if train_step > 0 and train_step % self._train_save_every_n_steps == 0:
logger.debug('Saving files for itr {0}'.format(train_itr))
self._save_train(train_itr)
train_itr += 1
### reset model
if train_step > 0 and self._train_reset_every_n_steps is not None and \
train_step % self._train_reset_every_n_steps == 0:
logger.debug('Resetting model')
self._policy.reset_weights()
### load data
new_inference_itr = self._get_inference_itr()
if inference_itr < new_inference_itr:
for i in range(inference_itr, new_inference_itr):
try:
logger.debug('Loading files for itr {0}'.format(i))
self._sampler.add_rollouts(
[self._train_rollouts_file_name(i)])
inference_itr = i + 1
except:
logger.debug(
'Failed to load files for itr {0}'.format(i))
def _run_save(self, save_itr):
timeit.start('save')
logger.info('Saving files for itr {0}'.format(save_itr))
self._save(save_itr, self._replay_pool.get_recent_rollouts(),
self._replay_pool_eval.get_recent_rollouts())
timeit.stop('save')
