def integrate_new_skill(self, new_skill_id, new_skill_subpath):
skill_integration_method = CategoricalMLPSkillIntegrator.Method.SUBPATH_SKILLS_AVG
## Hierarchized environment
hrl_env = HierarchizedEnv(
# base env that was wrapped in HierarchizedEnv (not fully unwrapped - may be normalized!)
env=self.env.env.env,
num_orig_skills=self._hrl_policy.num_skills
)
tf_hrl_env = TfEnv(hrl_env)
## Top policy
# 1) Get old policy from saved data
old_top_policy = self._hrl_policy.get_top_policy()
# 2) Get weights of old top policy
otp_weights = unflatten_tensors(
old_top_policy.get_param_values(),
old_top_policy.get_param_shapes()
)
# 3) Create weights for new top policy
skill_integrator = CategoricalMLPSkillIntegrator()
ntp_weight_values = skill_integrator.integrate_skill(
old_policy_weights=otp_weights,
method=skill_integration_method,
# Specific parameters for START_OBSS_SKILLS_AVG
subpath_start_obss=new_skill_subpath['start_observations'],
top_policy=old_top_policy,
# Specific parameters for SUBPATH_SKILLS_AVG, SUBPATH_SKILLS_SMOOTH_AVG and SUBPATH_FIRST_SKILL
subpath_actions=new_skill_subpath['actions']
)
# 4) Create new policy and randomly initialize its weights
new_top_policy = CategoricalMLPPolicy(
env_spec=tf_hrl_env.spec, # This env counts with new skill (action space = n + 1)
hidden_sizes=(32, 32), # As was in asa_test.py,
name='CategoricalMLPPolicyWithSkill{}'.format(new_skill_id)
)
ntp_init_op = tf.variables_initializer(new_top_policy.get_params())
ntp_init_op.run()
# 5) Fill new policy with adjusted weights
new_top_policy.set_param_values(
flattened_params=flatten_tensors(ntp_weight_values)
)
## Adjust HRL policy and training algorithms
self._hrl_policy.top_policy = new_top_policy
hrl_env.set_hrl_policy(self._hrl_policy)
self.env = tf_hrl_env
self.policy=self._hrl_policy.get_top_policy()
self._top_algo = self._top_algo_cls(
env=tf_hrl_env,
policy=self._hrl_policy.get_top_policy(),
baseline=self.baseline,
**self._top_algo_kwargs
)
self.sampler = self._top_algo.sampler
self.start_worker(self._tf_sess)
def run_task(*_):
# Configure TF session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config).as_default() as tf_session:
## Load data from itr_N.pkl
with open(snapshot_file, 'rb') as file:
saved_data = dill.load(file)
## Load data of new skill
global new_skill_subpath
if new_skill_policy_file:
with open(new_skill_policy_file, 'rb') as file:
new_skill_data = dill.load(file)
new_skill_policy = new_skill_data['policy']
new_skill_subpath = new_skill_data['subpath']
unique_end_obss = np.unique(new_skill_subpath['end_observations'], axis=0)
new_skill_stop_func = lambda path: (path['observations'][-1] == unique_end_obss).all(axis=1).any()
## Lower level environment & policies
# Base (original) environment.
base_env = saved_data['env'].env.env # <NormalizedEnv<MinibotEnv instance>>
# Skill policies, operating in base environment
skill_targets = [ # 13 basic room regions
( 6, 5), ( 6, 18), ( 6, 33), ( 6, 47), ( 6, 61),
(21, 5), (21, 18), (21, 33), (21, 47), (21, 61),
(37, 5), (37, 18), (37, 33),
]
trained_skill_policies = \
[GridworldTargetPolicy(env_spec=base_env.spec, target=t) for t in skill_targets] + \
[GridworldStepPolicy(env_spec=base_env.spec, direction=d, n=7) for d in range(4)] + \
[
new_skill_policy
# GridworldTargetPolicy(env_spec=base_env.spec, target=(43, 54)) # DEBUG use GridworldTargetPolicy as new skill
# GridworldRandomPolicy(env_spec=base_env.spec, n=25) # DEBUG use GridworldRandomPolicy as new skill
# GridworldStayPolicy(env_spec=base_env.spec, n=25) # DEBUG use GridworldStayPolicy as new skill
]
trained_skill_policies_stop_funcs = \
[pol.skill_stopping_func for pol in trained_skill_policies[:-1]] + \
[
new_skill_stop_func
# trained_skill_policies[-1].skill_stopping_func # DEBUG use Gridworld*Policy as new skill
]
skill_policy_prototype = saved_data['hrl_policy'].skill_policy_prototype
## Upper level environment & policies
# Hierarchized environment
hrl_env = HierarchizedEnv(
env=base_env,
num_orig_skills=len(trained_skill_policies)
)
tf_hrl_env = TfEnv(hrl_env)
## Top policy
# 1) Get old policy from saved data
old_top_policy = saved_data['policy']
# 2) Get weights of old top policy
otp_weights = unflatten_tensors(
old_top_policy.get_param_values(),
old_top_policy.get_param_shapes()
)
# 3) Create weights for new top policy
skill_integrator = CategoricalMLPSkillIntegrator()
ntp_weight_values = skill_integrator.integrate_skill(
old_policy_weights=otp_weights,
method=skill_integration_method,
# Specific parameters for START_OBSS_SKILLS_AVG
subpath_start_obss=new_skill_subpath['start_observations'],
top_policy=old_top_policy,
# Specific parameters for SUBPATH_SKILLS_AVG, SUBPATH_SKILLS_SMOOTH_AVG and SUBPATH_FIRST_SKILL
subpath_actions=new_skill_subpath['actions']
)
# 4) Create new policy and randomly initialize its weights
new_top_policy = CategoricalMLPPolicy(
env_spec=tf_hrl_env.spec, # This env counts with new skill (action space = n + 1)
hidden_sizes=(32, 32), # As was in asa_basic_run.py,
name="TopCategoricalMLPPolicy2"
)
ntp_init_op = tf.variables_initializer(new_top_policy.get_params())
ntp_init_op.run()
# 5) Fill new policy with adjusted weights
new_top_policy.set_param_values(
flattened_params=flatten_tensors(ntp_weight_values)
)
## Hierarchy of policies
hrl_policy = HierarchicalPolicy(
top_policy=new_top_policy,
skill_policy_prototype=skill_policy_prototype,
skill_policies=trained_skill_policies,
skill_stop_functions=trained_skill_policies_stop_funcs,
skill_max_timesteps=150
)
# Link hrl_policy and hrl_env, so that hrl_env can use skills
hrl_env.set_hrl_policy(hrl_policy)
## Other
# Baseline
baseline = saved_data['baseline'] # Take trained baseline
# Main ASA algorithm
asa_algo = AdaptiveSkillAcquisition(
env=tf_hrl_env,
hrl_policy=hrl_policy,
baseline=baseline,
top_algo_cls=TRPO,
low_algo_cls=TRPO,
# Top algo kwargs
batch_size=5000,
max_path_length=50,
n_itr=300,
start_itr=saved_data['itr'] + 1, # Continue from previous iteration number
discount=0.99,
force_batch_sampler=True,
low_algo_kwargs={
'batch_size': 20000,
'max_path_length': 800,
'n_itr': 300,
'discount': 0.99,
}
)
## Launch training
train_info = asa_algo.train(
sess=tf_session,
snapshot_mode='none'
)
## Save last iteration
out_file = os.path.join(train_info['snapshot_dir'], 'final.pkl')
empty_samples_data = {'paths': None}
with open(out_file, 'wb') as file:
out_data = asa_algo.get_itr_snapshot(
itr=asa_algo.n_itr - 1,
samples_data=empty_samples_data
)
dill.dump(out_data, file)