def _test_parametric_dqn_workflow(self,
use_gpu=False,
use_all_avail_gpus=False):
"""Run Parametric DQN workflow to ensure no crashes, algorithm correctness
not tested here."""
with tempfile.TemporaryDirectory() as tmpdirname:
lockfile = os.path.join(tmpdirname, "multiprocess_lock")
Path(lockfile).touch()
params = {
"training_data_path":
os.path.join(
curr_dir,
"test_data/parametric_action/cartpole_training.json.bz2"),
"eval_data_path":
os.path.join(
curr_dir,
"test_data/parametric_action/cartpole_eval.json.bz2"),
"state_norm_data_path":
os.path.join(
curr_dir,
"test_data/parametric_action/state_features_norm.json"),
"action_norm_data_path":
os.path.join(curr_dir,
"test_data/parametric_action/action_norm.json"),
"model_output_path":
tmpdirname,
"use_gpu":
use_gpu,
"use_all_avail_gpus":
use_all_avail_gpus,
"init_method":
"file://" + lockfile,
"num_nodes":
1,
"node_index":
0,
"epochs":
1,
"rl": {},
"rainbow": {},
"training": {
"minibatch_size": 128
},
}
parametric_dqn_workflow.main(params)
predictor_files = glob.glob(tmpdirname + "/predictor_*.c2")
assert len(
predictor_files) == 1, "Somehow created two predictor files!"
predictor = ParametricDQNPredictor.load(predictor_files[0],
"minidb")
test_float_state_features = [{
"0": 1.0,
"1": 1.0,
"2": 1.0,
"3": 1.0
}]
test_action_features = [{"4": 0.0, "5": 1.0}]
q_values = predictor.predict(test_float_state_features,
test_action_features)
assert len(q_values[0].keys()) == 1
def predictor(self) -> ParametricDQNPredictor:
"""Builds a ParametricDQNPredictor."""
return ParametricDQNPredictor.export(
self,
self.state_normalization_parameters,
self.action_normalization_parameters,
self.use_gpu,
)
def critic_prediction(self, float_state_features, actions):
""" Critic Prediction - Returns values for each state/action pair. Accepts
:param float_state_features states as list of feature -> float value dict
:param actions actions as list of feature -> value dict
"""
return ParametricDQNPredictor.predict(self, float_state_features,
actions)
def predictor(self) -> ParametricDQNPredictor:
"""Builds a ParametricDQNPredictor."""
return ParametricDQNPredictor.export(
self,
self.state_normalization_parameters,
self.action_normalization_parameters,
self._additional_feature_types.int_features,
self.use_gpu,
)
def predictor(self) -> ParametricDQNPredictor:
"""Builds a ParametricDQNPredictor."""
return ParametricDQNPredictor.export(
self,
self.state_normalization_parameters,
self.action_normalization_parameters,
self._additional_feature_types.int_features,
self.use_gpu,
)
def critic_prediction(self, float_state_features, int_state_features, actions):
""" Critic Prediction - Returns values for each state/action pair. Accepts
int_features as 3rd optional parameter.
:param float_state_features states as list of feature -> float value dict
:param int_state_features states as list of feature -> int value dict
:param actions actions as list of feature -> value dict
"""
return ParametricDQNPredictor.predict(
self, float_state_features, int_state_features, actions
)
def export_critic(
cls,
trainer,
state_normalization_parameters,
action_normalization_parameters,
model_on_gpu=False,
):
return ParametricDQNPredictor.export(
trainer,
state_normalization_parameters,
action_normalization_parameters,
model_on_gpu=model_on_gpu,
normalize_actions=False,
)
def export_critic(
cls,
trainer,
state_normalization_parameters,
action_normalization_parameters,
int_features=False,
model_on_gpu=False,
):
return ParametricDQNPredictor.export(
trainer,
state_normalization_parameters,
action_normalization_parameters,
int_features=int_features,
model_on_gpu=model_on_gpu,
normalize_actions=False,
)
