def _setup(self, variant):
set_seed(variant['run_params']['seed'])
self._variant = variant
self._session = tf.keras.backend.get_session()
self.train_generator = None
self._built = False
def _setup(self, variant):
set_seed(variant['run_params']['seed'])
self._variant = variant
gpu_options = tf.GPUOptions(allow_growth=True)
session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
tf.keras.backend.set_session(session)
self._session = tf.keras.backend.get_session()
self.train_generator = None
self._built = False
def _setup(self, variant):
set_seed(variant['run_params']['seed'])
self._variant = variant
self.SAVE_PER_ALGO = {
'default': self.save_mbpo,
'MBPO': self.save_mbpo,
'CMBPO': self.save_cmbpo
}
self.RESTORE_PER_ALGO = {
'default': self.restore_mbpo,
'MBPO': self.restore_mbpo,
'CMBPO': self.restore_cmbpo,
}
gpu_options = tf.GPUOptions(allow_growth=True)
session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
tf.keras.backend.set_session(session)
self._session = tf.keras.backend.get_session()
self.train_generator = None
self._built = False
def main():
import sys
example_args = get_parser().parse_args(sys.argv[1:])
variant_spec = get_variant_spec(example_args)
command_line_args = example_args
print('vriant spec: {}'.format(variant_spec))
params = variant_spec.get('algorithm_params')
local_dir = os.path.join(params.get('log_dir'), params.get('domain'))
resources_per_trial = _normalize_trial_resources(
command_line_args.resources_per_trial, command_line_args.trial_cpus,
command_line_args.trial_gpus, command_line_args.trial_extra_cpus,
command_line_args.trial_extra_gpus)
experiment_id = params.get('exp_name')
#### add pool_load_max_size to experiment_id
if 'pool_load_max_size' in variant_spec['algorithm_params']['kwargs']:
max_size = variant_spec['algorithm_params']['kwargs'][
'pool_load_max_size']
experiment_id = '{}_{}e3'.format(experiment_id, int(max_size / 1000))
####
variant_spec = add_command_line_args_to_variant_spec(
variant_spec, command_line_args)
if command_line_args.video_save_frequency is not None:
assert 'algorithm_params' in variant_spec
variant_spec['algorithm_params']['kwargs']['video_save_frequency'] = (
command_line_args.video_save_frequency)
variant = variant_spec
# init
set_seed(variant['run_params']['seed'])
gpu_options = tf.GPUOptions(allow_growth=True)
session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
tf.keras.backend.set_session(session)
# build
variant = copy.deepcopy(variant)
tester.set_hyper_param(**variant)
tester.add_record_param(['run_params.seed', 'info'])
tester.configure(task_name='policy_learn',
private_config_path=os.path.join(get_package_path(),
'rla_config.yaml'),
run_file='main.py',
log_root=get_package_path())
tester.log_files_gen()
tester.print_args()
environment_params = variant['environment_params']
training_environment = (get_environment_from_params(
environment_params['training']))
evaluation_environment = (get_environment_from_params(
environment_params['evaluation'](variant)) if 'evaluation'
in environment_params else training_environment)
replay_pool = (get_replay_pool_from_variant(variant, training_environment))
sampler = get_sampler_from_variant(variant)
Qs = get_Q_function_from_variant(variant, training_environment)
policy = get_policy_from_variant(variant, training_environment, Qs)
initial_exploration_policy = (get_policy('UniformPolicy',
training_environment))
#### get termination function
domain = environment_params['training']['domain']
static_fns = mopo.static[domain.lower()]
####
print("[ DEBUG ] KWARGS: {}".format(variant['algorithm_params']['kwargs']))
algorithm = get_algorithm_from_variant(
variant=variant,
training_environment=training_environment,
evaluation_environment=evaluation_environment,
policy=policy,
initial_exploration_policy=initial_exploration_policy,
Qs=Qs,
pool=replay_pool,
static_fns=static_fns,
sampler=sampler,
session=session)
print('[ DEBUG ] finish construct model, start training')
# train
list(algorithm.train())
