def get_actor_predictor(self, trainer, environment):
state_preprocessor = Preprocessor(environment.normalization, False)
postprocessor = Postprocessor(
environment.normalization_continuous_action, False)
actor_with_preprocessor = ActorWithPreprocessor(
trainer.actor_network.cpu_model().eval(), state_preprocessor,
postprocessor)
serving_module = ActorPredictorWrapper(actor_with_preprocessor)
predictor = ActorTorchPredictor(
serving_module,
sort_features_by_normalization(
environment.normalization_continuous_action)[0],
)
return predictor
def test_prepare_normalization_and_normalize(self):
feature_value_map = read_data()
normalization_parameters = {}
for name, values in feature_value_map.items():
normalization_parameters[name] = normalization.identify_parameter(
name,
values,
10,
feature_type=self._feature_type_override(name))
for k, v in normalization_parameters.items():
if id_to_type(k) == CONTINUOUS:
self.assertEqual(v.feature_type, CONTINUOUS)
self.assertIs(v.boxcox_lambda, None)
self.assertIs(v.boxcox_shift, None)
elif id_to_type(k) == BOXCOX:
self.assertEqual(v.feature_type, BOXCOX)
self.assertIsNot(v.boxcox_lambda, None)
self.assertIsNot(v.boxcox_shift, None)
else:
assert v.feature_type == id_to_type(k)
preprocessor = Preprocessor(normalization_parameters, False)
sorted_features, _ = sort_features_by_normalization(
normalization_parameters)
input_matrix = torch.zeros([10000, len(sorted_features)])
for i, feature in enumerate(sorted_features):
input_matrix[:, i] = torch.from_numpy(feature_value_map[feature])
normalized_feature_matrix = preprocessor(
input_matrix, (input_matrix != MISSING_VALUE))
normalized_features = {}
on_column = 0
for feature in sorted_features:
norm = normalization_parameters[feature]
if norm.feature_type == ENUM:
column_size = len(norm.possible_values)
else:
column_size = 1
normalized_features[
feature] = normalized_feature_matrix[:,
on_column:(on_column +
column_size)]
on_column += column_size
self.assertTrue(
all([
np.isfinite(parameter.stddev) and np.isfinite(parameter.mean)
for parameter in normalization_parameters.values()
]))
for k, v in six.iteritems(normalized_features):
v = v.numpy()
self.assertTrue(np.all(np.isfinite(v)))
feature_type = normalization_parameters[k].feature_type
if feature_type == identify_types.PROBABILITY:
sigmoidv = special.expit(v)
self.assertTrue(
np.all(
np.logical_and(np.greater(sigmoidv, 0),
np.less(sigmoidv, 1))))
elif feature_type == identify_types.ENUM:
possible_values = normalization_parameters[k].possible_values
self.assertEqual(v.shape[0], len(feature_value_map[k]))
self.assertEqual(v.shape[1], len(possible_values))
possible_value_map = {}
for i, possible_value in enumerate(possible_values):
possible_value_map[possible_value] = i
for i, row in enumerate(v):
original_feature = feature_value_map[k][i]
if abs(original_feature - MISSING_VALUE) < 0.01:
self.assertEqual(0.0, np.sum(row))
else:
self.assertEqual(
possible_value_map[original_feature],
np.where(row == 1)[0][0],
)
elif feature_type == identify_types.QUANTILE:
for i, feature in enumerate(v[0]):
original_feature = feature_value_map[k][i]
expected = NumpyFeatureProcessor.value_to_quantile(
original_feature,
normalization_parameters[k].quantiles)
self.assertAlmostEqual(feature, expected, 2)
elif feature_type == identify_types.BINARY:
pass
elif (feature_type == identify_types.CONTINUOUS
or feature_type == identify_types.BOXCOX):
one_stddev = np.isclose(np.std(v, ddof=1), 1, atol=0.01)
zero_stddev = np.isclose(np.std(v, ddof=1), 0, atol=0.01)
zero_mean = np.isclose(np.mean(v), 0, atol=0.01)
self.assertTrue(
np.all(zero_mean),
"mean of feature {} is {}, not 0".format(k, np.mean(v)),
)
self.assertTrue(np.all(np.logical_or(one_stddev, zero_stddev)))
elif feature_type == identify_types.CONTINUOUS_ACTION:
less_than_max = v < 1
more_than_min = v > -1
self.assertTrue(
np.all(less_than_max),
"values are not less than 1: {}".format(
v[less_than_max == False]),
)
self.assertTrue(
np.all(more_than_min),
"values are not more than -1: {}".format(
v[more_than_min == False]),
)
else:
raise NotImplementedError()
def single_process_main(gpu_index, *args):
params = args[0]
# 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 = from_json(params["rl"], RLParameters)
training_parameters = from_json(params["training"], TrainingParameters)
rainbow_parameters = from_json(params["rainbow"], RainbowDQNParameters)
model_params = ContinuousActionModelParameters(
rl=rl_parameters,
training=training_parameters,
rainbow=rainbow_parameters)
state_normalization = BaseWorkflow.read_norm_file(
params["state_norm_data_path"])
action_normalization = BaseWorkflow.read_norm_file(
params["action_norm_data_path"])
writer = SummaryWriter(log_dir=params["model_output_path"])
logger.info("TensorBoard logging location is: {}".format(writer.log_dir))
if params["use_all_avail_gpus"]:
BaseWorkflow.init_multiprocessing(
int(params["num_processes_per_node"]),
int(params["num_nodes"]),
int(params["node_index"]),
gpu_index,
params["init_method"],
)
workflow = ParametricDqnWorkflow(
model_params,
state_normalization,
action_normalization,
params["use_gpu"],
params["use_all_avail_gpus"],
)
state_sorted_features, _ = sort_features_by_normalization(
state_normalization)
action_sorted_features, _ = sort_features_by_normalization(
action_normalization)
preprocess_handler = ParametricDqnPreprocessHandler(
StringKeySparseToDenseProcessor(state_sorted_features),
StringKeySparseToDenseProcessor(action_sorted_features),
)
train_dataset = JSONDatasetReader(
params["training_data_path"],
batch_size=training_parameters.minibatch_size,
preprocess_handler=preprocess_handler,
)
eval_dataset = JSONDatasetReader(params["eval_data_path"],
batch_size=16,
preprocess_handler=preprocess_handler)
with summary_writer_context(writer):
workflow.train_network(train_dataset, eval_dataset,
int(params["epochs"]))
if int(params["node_index"]) == 0 and gpu_index == 0:
workflow.save_models(params["model_output_path"])
def preprocess_samples(
self,
samples: Samples,
minibatch_size: int,
use_gpu: bool = False,
one_hot_action: bool = True,
normalize_actions: bool = True,
) -> List[TrainingDataPage]:
logger.info("Shuffling...")
samples = shuffle_samples(samples)
logger.info("Sparse2Dense...")
sorted_state_features, _ = sort_features_by_normalization(self.normalization)
sorted_action_features, _ = sort_features_by_normalization(
self.normalization_action
)
state_sparse_to_dense_processor = PythonSparseToDenseProcessor(
sorted_state_features
)
action_sparse_to_dense_processor = PythonSparseToDenseProcessor(
sorted_action_features
)
state_matrix, state_matrix_presence = state_sparse_to_dense_processor(
samples.states
)
next_state_matrix, next_state_matrix_presence = state_sparse_to_dense_processor(
samples.next_states
)
action_matrix, action_matrix_presence = action_sparse_to_dense_processor(
samples.actions
)
(
next_action_matrix,
next_action_matrix_presence,
) = action_sparse_to_dense_processor(samples.next_actions)
action_probabilities = torch.tensor(
samples.action_probabilities, dtype=torch.float32
).reshape(-1, 1)
rewards = torch.tensor(samples.rewards, dtype=torch.float32).reshape(-1, 1)
max_action_size = 4
pnas_mask_list: List[List[int]] = []
pnas_flat: List[Dict[str, float]] = []
for pnas in samples.possible_next_actions:
pnas_mask_list.append([1] * len(pnas) + [0] * (max_action_size - len(pnas)))
pnas_flat.extend(pnas)
for _ in range(max_action_size - len(pnas)):
pnas_flat.append({}) # Filler
pnas_mask = torch.Tensor(pnas_mask_list)
(
possible_next_actions_matrix,
possible_next_actions_matrix_presence,
) = action_sparse_to_dense_processor(pnas_flat)
logger.info("Preprocessing...")
state_preprocessor = Preprocessor(self.normalization, False)
action_preprocessor = Preprocessor(self.normalization_action, False)
states_ndarray = state_preprocessor(state_matrix, state_matrix_presence)
if normalize_actions:
actions_ndarray = action_preprocessor(action_matrix, action_matrix_presence)
else:
actions_ndarray = action_matrix
next_states_ndarray = state_preprocessor(
next_state_matrix, next_state_matrix_presence
)
state_pnas_tile = next_states_ndarray.repeat(1, max_action_size).reshape(
-1, next_states_ndarray.shape[1]
)
if normalize_actions:
next_actions_ndarray = action_preprocessor(
next_action_matrix, next_action_matrix_presence
)
else:
next_actions_ndarray = next_action_matrix
if normalize_actions:
logged_possible_next_actions = action_preprocessor(
possible_next_actions_matrix, possible_next_actions_matrix_presence
)
else:
logged_possible_next_actions = possible_next_actions_matrix
assert state_pnas_tile.shape[0] == logged_possible_next_actions.shape[0], (
"Invalid shapes: "
+ str(state_pnas_tile.shape)
+ " != "
+ str(logged_possible_next_actions.shape)
)
logged_possible_next_state_actions = torch.cat(
(state_pnas_tile, logged_possible_next_actions), dim=1
)
logger.info("Reward Timeline to Torch...")
time_diffs = torch.ones([len(samples.states), 1])
tdps = []
pnas_start = 0
logger.info("Batching...")
for start in range(0, states_ndarray.shape[0], minibatch_size):
end = start + minibatch_size
if end > states_ndarray.shape[0]:
break
pnas_end = pnas_start + (minibatch_size * max_action_size)
tdp = TrainingDataPage(
states=states_ndarray[start:end],
actions=actions_ndarray[start:end],
propensities=action_probabilities[start:end],
rewards=rewards[start:end],
next_states=next_states_ndarray[start:end],
next_actions=next_actions_ndarray[start:end],
not_terminal=(pnas_mask[start:end, :].sum(dim=1, keepdim=True) > 0),
time_diffs=time_diffs[start:end],
possible_next_actions_mask=pnas_mask[start:end, :],
possible_next_actions_state_concat=logged_possible_next_state_actions[
pnas_start:pnas_end, :
],
)
pnas_start = pnas_end
tdp.set_type(torch.cuda.FloatTensor if use_gpu else torch.FloatTensor)
tdps.append(tdp)
return tdps
def preprocess_samples_discrete(
self,
samples: Samples,
minibatch_size: int,
one_hot_action: bool = True,
use_gpu: bool = False,
do_shuffle: bool = True,
) -> List[TrainingDataPage]:
if do_shuffle:
logger.info("Shuffling...")
samples = shuffle_samples(samples)
logger.info("Preprocessing...")
sorted_features, _ = sort_features_by_normalization(self.normalization)
sparse_to_dense_processor = PythonSparseToDenseProcessor(
sorted_features)
state_matrix, state_matrix_presence = sparse_to_dense_processor(
samples.states)
next_state_matrix, next_state_matrix_presence = sparse_to_dense_processor(
samples.next_states)
logger.info("Converting to Torch...")
actions_one_hot = torch.tensor((np.array(samples.actions).reshape(
-1, 1) == np.array(self.ACTIONS)).astype(np.int64))
actions = actions_one_hot.argmax(dim=1, keepdim=True)
rewards = torch.tensor(samples.rewards,
dtype=torch.float32).reshape(-1, 1)
action_probabilities = torch.tensor(samples.action_probabilities,
dtype=torch.float32).reshape(
-1, 1)
next_actions_one_hot = torch.tensor(
(np.array(samples.next_actions).reshape(-1, 1) == np.array(
self.ACTIONS)).astype(np.int64))
logger.info("Converting PA to Torch...")
possible_action_strings = np.array(
list(itertools.zip_longest(*samples.possible_actions,
fillvalue=""))).T
possible_actions_mask = torch.zeros(
[len(samples.actions), len(self.ACTIONS)])
for i, action in enumerate(self.ACTIONS):
possible_actions_mask[:, i] = torch.tensor(
np.max(possible_action_strings == action,
axis=1).astype(np.int64))
logger.info("Converting PNA to Torch...")
possible_next_action_strings = np.array(
list(
itertools.zip_longest(*samples.possible_next_actions,
fillvalue=""))).T
possible_next_actions_mask = torch.zeros(
[len(samples.next_actions),
len(self.ACTIONS)])
for i, action in enumerate(self.ACTIONS):
possible_next_actions_mask[:, i] = torch.tensor(
np.max(possible_next_action_strings == action,
axis=1).astype(np.int64))
terminals = torch.tensor(samples.terminals,
dtype=torch.int32).reshape(-1, 1)
not_terminal = 1 - terminals
logger.info("Converting RT to Torch...")
time_diffs = torch.ones([len(samples.states), 1])
logger.info("Preprocessing...")
preprocessor = Preprocessor(self.normalization, False)
states_ndarray = preprocessor(state_matrix, state_matrix_presence)
next_states_ndarray = preprocessor(next_state_matrix,
next_state_matrix_presence)
logger.info("Batching...")
tdps = []
for start in range(0, states_ndarray.shape[0], minibatch_size):
end = start + minibatch_size
if end > states_ndarray.shape[0]:
break
tdp = TrainingDataPage(
states=states_ndarray[start:end],
actions=actions_one_hot[start:end]
if one_hot_action else actions[start:end],
propensities=action_probabilities[start:end],
rewards=rewards[start:end],
next_states=next_states_ndarray[start:end],
# pyre-fixme[16]: `int` has no attribute `__getitem__`.
not_terminal=not_terminal[start:end],
next_actions=next_actions_one_hot[start:end],
possible_actions_mask=possible_actions_mask[start:end],
possible_next_actions_mask=possible_next_actions_mask[
start:end],
time_diffs=time_diffs[start:end],
)
tdp.set_type(
torch.cuda.FloatTensor if use_gpu else torch.FloatTensor)
tdps.append(tdp)
return tdps
