def __init__(
self,
model_params: ContinuousActionModelParameters,
preprocess_handler: PreprocessHandler,
state_normalization: Dict[int, NormalizationParameters],
action_normalization: Dict[int, NormalizationParameters],
use_gpu: bool,
use_all_avail_gpus: bool,
):
logger.info("Running Parametric DQN workflow with params:")
logger.info(model_params)
model_params = model_params
trainer = ParametricDQNTrainer(
model_params,
state_normalization,
action_normalization,
use_gpu=use_gpu,
use_all_avail_gpus=use_all_avail_gpus,
)
trainer = update_model_for_warm_start(trainer)
assert (type(trainer) == ParametricDQNTrainer
), "Warm started wrong model type: " + str(type(trainer))
evaluator = Evaluator(
None,
model_params.rl.gamma,
trainer,
metrics_to_score=trainer.metrics_to_score,
)
super(ParametricDqnWorkflow,
self).__init__(preprocess_handler, trainer, evaluator,
model_params.training.minibatch_size)
def __init__(
self,
model_params: DiscreteActionModelParameters,
state_normalization: Dict[int, NormalizationParameters],
use_gpu: bool,
use_all_avail_gpus: bool,
):
logger.info("Running DQN workflow with params:")
logger.info(model_params)
model_params = model_params
trainer = create_dqn_trainer_from_params(
model_params,
state_normalization,
use_gpu=use_gpu,
use_all_avail_gpus=use_all_avail_gpus,
)
trainer = update_model_for_warm_start(trainer)
assert type(trainer) == DQNTrainer, "Warm started wrong model type: " + str(
type(trainer)
)
evaluator = Evaluator(
model_params.actions,
model_params.rl.gamma,
trainer,
metrics_to_score=trainer.metrics_to_score,
)
super().__init__(
DiscreteDqnBatchPreprocessor(Preprocessor(state_normalization, use_gpu)),
trainer,
evaluator,
model_params.training.minibatch_size,
)
def train(self, train_dataset: Dataset, eval_dataset: Optional[Dataset],
num_epochs: int) -> RLTrainingOutput:
"""
Train the model
Returns partially filled RLTrainningOutput. The field that should not be filled
are:
- output_path
- warmstart_output_path
- vis_metrics
- validation_output
"""
logger.info("Creating reporter")
reporter = DiscreteDQNReporter(
self.trainer_param.actions,
target_action_distribution=self.target_action_distribution,
)
logger.info("Adding reporter to trainer")
self.trainer.add_observer(reporter)
training_page_handler = TrainingPageHandler(self.trainer)
training_page_handler.add_observer(reporter)
evaluator = Evaluator(
self.action_names,
self.rl_parameters.gamma,
self.trainer,
metrics_to_score=self.metrics_to_score,
)
logger.info("Adding reporter to evaluator")
evaluator.add_observer(reporter)
evaluation_page_handler = EvaluationPageHandler(
self.trainer, evaluator, reporter)
batch_preprocessor = self.build_batch_preprocessor()
train_and_evaluate_generic(
train_dataset,
eval_dataset,
self.trainer,
num_epochs,
self.use_gpu,
batch_preprocessor,
training_page_handler,
evaluation_page_handler,
reader_options=self.reader_options,
)
return RLTrainingOutput(
training_report=reporter.generate_training_report())
def __init__(
self,
model_params: ContinuousActionModelParameters,
preprocess_handler: PreprocessHandler,
state_normalization: Dict[int, NormalizationParameters],
action_normalization: Dict[int, NormalizationParameters],
use_gpu: bool,
use_all_avail_gpus: bool,
):
logger.info("Running continuous workflow with params:")
logger.info(model_params)
model_params = model_params
min_action_range_tensor_serving, max_action_range_tensor_serving = construct_action_scale_tensor(
action_normalization, model_params.action_rescale_map)
trainer = DDPGTrainer(
model_params,
state_normalization,
action_normalization,
min_action_range_tensor_serving,
max_action_range_tensor_serving,
use_gpu=use_gpu,
use_all_avail_gpus=use_all_avail_gpus,
)
trainer = update_model_for_warm_start(trainer)
assert type(
trainer) == DDPGTrainer, "Warm started wrong model type: " + str(
type(trainer))
evaluator = Evaluator(
None,
model_params.rl.gamma,
trainer,
metrics_to_score=trainer.metrics_to_score,
)
super(ContinuousWorkflow, self).__init__(
preprocess_handler,
trainer,
evaluator,
model_params.shared_training.minibatch_size,
)
def train_network(params):
writer = None
if params["model_output_path"] is not None:
writer = SummaryWriter(log_dir=params["model_output_path"])
logger.info("Running DQN workflow with params:")
logger.info(params)
# Set minibatch size based on # of devices being used to train
params["training"]["minibatch_size"] *= minibatch_size_multiplier(
params["use_gpu"], params["use_all_avail_gpus"])
action_names = np.array(params["actions"])
rl_parameters = RLParameters(**params["rl"])
training_parameters = TrainingParameters(**params["training"])
rainbow_parameters = RainbowDQNParameters(**params["rainbow"])
trainer_params = DiscreteActionModelParameters(
actions=params["actions"],
rl=rl_parameters,
training=training_parameters,
rainbow=rainbow_parameters,
)
dataset = JSONDataset(params["training_data_path"],
batch_size=training_parameters.minibatch_size)
eval_dataset = JSONDataset(params["eval_data_path"], batch_size=16)
state_normalization = read_norm_file(params["state_norm_data_path"])
num_batches = int(len(dataset) / training_parameters.minibatch_size)
logger.info("Read in batch data set {} of size {} examples. Data split "
"into {} batches of size {}.".format(
params["training_data_path"],
len(dataset),
num_batches,
training_parameters.minibatch_size,
))
trainer = DQNTrainer(
trainer_params,
state_normalization,
use_gpu=params["use_gpu"],
use_all_avail_gpus=params["use_all_avail_gpus"],
)
trainer = update_model_for_warm_start(trainer)
preprocessor = Preprocessor(state_normalization, False)
evaluator = Evaluator(
trainer_params.actions,
trainer_params.rl.gamma,
trainer,
metrics_to_score=trainer.metrics_to_score,
)
start_time = time.time()
for epoch in range(int(params["epochs"])):
dataset.reset_iterator()
for batch_idx in range(num_batches):
report_training_status(batch_idx, num_batches, epoch,
int(params["epochs"]))
batch = dataset.read_batch(batch_idx)
tdp = preprocess_batch_for_training(preprocessor, batch,
action_names)
tdp.set_type(trainer.dtype)
trainer.train(tdp)
eval_dataset.reset_iterator()
accumulated_edp = None
while True:
batch = eval_dataset.read_batch(batch_idx)
if batch is None:
break
tdp = preprocess_batch_for_training(preprocessor, batch,
action_names)
edp = EvaluationDataPage.create_from_tdp(tdp, trainer)
if accumulated_edp is None:
accumulated_edp = edp
else:
accumulated_edp = accumulated_edp.append(edp)
accumulated_edp = accumulated_edp.compute_values(trainer.gamma)
cpe_start_time = time.time()
details = evaluator.evaluate_post_training(accumulated_edp)
details.log()
logger.info("CPE evaluation took {} seconds.".format(time.time() -
cpe_start_time))
through_put = (len(dataset) * int(params["epochs"])) / (time.time() -
start_time)
logger.info("Training finished. Processed ~{} examples / s.".format(
round(through_put)))
if writer is not None:
writer.close()
return export_trainer_and_predictor(trainer, params["model_output_path"])
def train_network(params):
logger.info("Running Parametric DQN workflow with params:")
logger.info(params)
# Set minibatch size based on # of devices being used to train
params["training"]["minibatch_size"] *= minibatch_size_multiplier(
params["use_gpu"], params["use_all_avail_gpus"])
rl_parameters = RLParameters(**params["rl"])
training_parameters = TrainingParameters(**params["training"])
rainbow_parameters = RainbowDQNParameters(**params["rainbow"])
if params["in_training_cpe"] is not None:
in_training_cpe_parameters = InTrainingCPEParameters(
**params["in_training_cpe"])
else:
in_training_cpe_parameters = None
trainer_params = ContinuousActionModelParameters(
rl=rl_parameters,
training=training_parameters,
rainbow=rainbow_parameters,
in_training_cpe=in_training_cpe_parameters,
)
dataset = JSONDataset(params["training_data_path"],
batch_size=training_parameters.minibatch_size)
eval_dataset = JSONDataset(params["eval_data_path"],
batch_size=training_parameters.minibatch_size)
state_normalization = read_norm_file(params["state_norm_data_path"])
action_normalization = read_norm_file(params["action_norm_data_path"])
num_batches = int(len(dataset) / training_parameters.minibatch_size)
logger.info("Read in batch data set {} of size {} examples. Data split "
"into {} batches of size {}.".format(
params["training_data_path"],
len(dataset),
num_batches,
training_parameters.minibatch_size,
))
trainer = ParametricDQNTrainer(
trainer_params,
state_normalization,
action_normalization,
use_gpu=params["use_gpu"],
use_all_avail_gpus=params["use_all_avail_gpus"],
)
trainer = update_model_for_warm_start(trainer)
state_preprocessor = Preprocessor(state_normalization, False)
action_preprocessor = Preprocessor(action_normalization, False)
if trainer_params.in_training_cpe is not None:
evaluator = Evaluator(
None,
trainer_params.rl.gamma,
trainer,
trainer_params.in_training_cpe.mdp_sampled_rate,
metrics_to_score=trainer.metrics_to_score,
)
else:
evaluator = Evaluator(
None,
trainer_params.rl.gamma,
trainer,
float(DEFAULT_NUM_SAMPLES_FOR_CPE) / len(dataset),
metrics_to_score=trainer.metrics_to_score,
)
start_time = time.time()
for epoch in range(params["epochs"]):
dataset.reset_iterator()
for batch_idx in range(num_batches):
report_training_status(batch_idx, num_batches, epoch,
params["epochs"])
batch = dataset.read_batch(batch_idx)
tdp = preprocess_batch_for_training(
state_preprocessor,
batch,
action_preprocessor=action_preprocessor)
tdp.set_type(trainer.dtype)
trainer.train(tdp)
eval_dataset.reset_iterator()
accumulated_edp = None
for batch_idx in range(num_batches):
batch = eval_dataset.read_batch(batch_idx)
tdp = preprocess_batch_for_training(
state_preprocessor,
batch,
action_preprocessor=action_preprocessor)
edp = EvaluationDataPage.create_from_tdp(tdp, trainer)
if accumulated_edp is None:
accumulated_edp = edp
else:
accumulated_edp = accumulated_edp.append(edp)
accumulated_edp = accumulated_edp.compute_values(trainer.gamma)
cpe_start_time = time.time()
details = evaluator.evaluate_post_training(accumulated_edp)
details.log()
logger.info("CPE evaluation took {} seconds.".format(time.time() -
cpe_start_time))
through_put = (len(dataset) * params["epochs"]) / (time.time() -
start_time)
logger.info("Training finished. Processed ~{} examples / s.".format(
round(through_put)))
return export_trainer_and_predictor(trainer, params["model_output_path"])
def train_network(params):
writer = None
if params["model_output_path"] is not None:
writer = SummaryWriter(log_dir=params["model_output_path"])
logger.info("Running DQN workflow with params:")
logger.info(params)
# Set minibatch size based on # of devices being used to train
params["training"]["minibatch_size"] *= minibatch_size_multiplier(
params["use_gpu"], params["use_all_avail_gpus"]
)
action_names = np.array(params["actions"])
rl_parameters = RLParameters(**params["rl"])
training_parameters = TrainingParameters(**params["training"])
rainbow_parameters = RainbowDQNParameters(**params["rainbow"])
trainer_params = DiscreteActionModelParameters(
actions=params["actions"],
rl=rl_parameters,
training=training_parameters,
rainbow=rainbow_parameters,
)
dataset = JSONDataset(
params["training_data_path"], batch_size=training_parameters.minibatch_size
)
eval_dataset = JSONDataset(params["eval_data_path"], batch_size=16)
state_normalization = read_norm_file(params["state_norm_data_path"])
num_batches = int(len(dataset) / training_parameters.minibatch_size)
logger.info(
"Read in batch data set {} of size {} examples. Data split "
"into {} batches of size {}.".format(
params["training_data_path"],
len(dataset),
num_batches,
training_parameters.minibatch_size,
)
)
trainer = DQNTrainer(
trainer_params,
state_normalization,
use_gpu=params["use_gpu"],
use_all_avail_gpus=params["use_all_avail_gpus"],
)
trainer = update_model_for_warm_start(trainer)
preprocessor = Preprocessor(state_normalization, False)
evaluator = Evaluator(
trainer_params.actions,
trainer_params.rl.gamma,
trainer,
metrics_to_score=trainer.metrics_to_score,
)
start_time = time.time()
for epoch in range(int(params["epochs"])):
dataset.reset_iterator()
batch_idx = -1
while True:
batch_idx += 1
report_training_status(batch_idx, num_batches, epoch, int(params["epochs"]))
batch = dataset.read_batch()
if batch is None:
break
tdp = preprocess_batch_for_training(preprocessor, batch, action_names)
tdp.set_type(trainer.dtype)
trainer.train(tdp)
eval_dataset.reset_iterator()
accumulated_edp = None
while True:
batch = eval_dataset.read_batch()
if batch is None:
break
tdp = preprocess_batch_for_training(preprocessor, batch, action_names)
tdp.set_type(trainer.dtype)
edp = EvaluationDataPage.create_from_tdp(tdp, trainer)
if accumulated_edp is None:
accumulated_edp = edp
else:
accumulated_edp = accumulated_edp.append(edp)
accumulated_edp = accumulated_edp.compute_values(trainer.gamma)
cpe_start_time = time.time()
details = evaluator.evaluate_post_training(accumulated_edp)
details.log()
logger.info(
"CPE evaluation took {} seconds.".format(time.time() - cpe_start_time)
)
through_put = (len(dataset) * int(params["epochs"])) / (time.time() - start_time)
logger.info(
"Training finished. Processed ~{} examples / s.".format(round(through_put))
)
if writer is not None:
writer.close()
return export_trainer_and_predictor(trainer, params["model_output_path"])
def __init__(
self,
model_params: ContinuousActionModelParameters,
preprocess_handler: PreprocessHandler,
state_normalization: Dict[int, NormalizationParameters],
action_normalization: Dict[int, NormalizationParameters],
use_gpu: bool,
use_all_avail_gpus: bool,
):
logger.info("Running continuous workflow with params:")
logger.info(model_params)
model_params = model_params
min_action_range_tensor_serving, max_action_range_tensor_serving = construct_action_scale_tensor(
action_normalization, model_params.action_rescale_map
)
state_dim = get_num_output_features(state_normalization)
action_dim = get_num_output_features(action_normalization)
# Build Actor Network
actor_network = ActorNetModel(
layers=(
[state_dim] + model_params.actor_training.layers[1:-1] + [action_dim]
),
activations=model_params.actor_training.activations,
fl_init=model_params.shared_training.final_layer_init,
state_dim=state_dim,
action_dim=action_dim,
)
# Build Critic Network
critic_network = CriticNetModel(
# Ensure dims match input state and scalar output
layers=[state_dim] + model_params.critic_training.layers[1:-1] + [1],
activations=model_params.critic_training.activations,
fl_init=model_params.shared_training.final_layer_init,
state_dim=state_dim,
action_dim=action_dim,
)
trainer = DDPGTrainer(
actor_network,
critic_network,
model_params,
state_normalization,
action_normalization,
min_action_range_tensor_serving,
max_action_range_tensor_serving,
use_gpu=use_gpu,
use_all_avail_gpus=use_all_avail_gpus,
)
trainer = update_model_for_warm_start(trainer)
assert type(trainer) == DDPGTrainer, "Warm started wrong model type: " + str(
type(trainer)
)
evaluator = Evaluator(
None,
model_params.rl.gamma,
trainer,
metrics_to_score=trainer.metrics_to_score,
)
super().__init__(
preprocess_handler,
trainer,
evaluator,
model_params.shared_training.minibatch_size,
)