def normalize_configuration(cfg: Namespace, train_mode: bool) -> None:
"""Given a configuration namespace, normalize the values it contains.
Arguments:
cfg: The namespace object returned by `Configuration.make_namespace`
train_mode: Boolean flag controlling normalization of parameters only
used during training.
"""
if train_mode:
_normalize_train_cfg(cfg)
if cfg.tf_manager is None:
cfg.tf_manager = get_default_tf_manager()
cfg.evaluation = [(e[0], e[0], e[1]) if len(e) == 2 else e
for e in cfg.evaluation]
if cfg.evaluation:
cfg.main_metric = "{}/{}".format(cfg.evaluation[-1][0],
cfg.evaluation[-1][-1].name)
else:
cfg.main_metric = "{}/{}".format(cfg.runners[-1].decoder_data_id,
cfg.runners[-1].loss_names[0])
if not cfg.tf_manager.minimize_metric:
raise ValueError("minimize_metric must be set to True in "
"TensorFlowManager when using loss as "
"the main metric")
def normalize_configuration(cfg: Namespace, train_mode: bool) -> None:
"""Given a configuration namespace, normalize the values it contains.
Arguments:
cfg: The namespace object returned by `Configuration.make_namespace`
train_mode: Boolean flag controlling normalization of parameters only
used during training.
"""
if train_mode:
_normalize_train_cfg(cfg)
if cfg.tf_manager is None:
cfg.tf_manager = get_default_tf_manager()
if (cfg.batch_size is None) == (cfg.batching_scheme is None):
raise ValueError("You must specify either batch_size or "
"batching_scheme (not both).")
if cfg.batch_size is not None:
assert cfg.batching_scheme is None
cfg.batching_scheme = BatchingScheme(batch_size=cfg.batch_size)
else:
assert cfg.batching_scheme is not None
cfg.batch_size = cfg.batching_scheme.batch_size
if cfg.runners_batch_size is None:
cfg.runners_batch_size = cfg.batching_scheme.batch_size
cfg.runners_batching_scheme = BatchingScheme(
batch_size=cfg.runners_batch_size,
token_level_batching=cfg.batching_scheme.token_level_batching,
use_leftover_buckets=True)
cfg.evaluation = [(e[0], e[0], e[1]) if len(e) == 2 else e
for e in cfg.evaluation]
if cfg.evaluation:
cfg.main_metric = "{}/{}".format(cfg.evaluation[-1][0],
cfg.evaluation[-1][-1].name)
else:
cfg.main_metric = "{}/{}".format(cfg.runners[-1].decoder_data_id,
cfg.runners[-1].loss_names[0])
if not cfg.tf_manager.minimize_metric:
raise ValueError("minimize_metric must be set to True in "
"TensorFlowManager when using loss as "
"the main metric")
def build_model(self) -> None:
if self._model_built:
raise RuntimeError("build_model() called twice")
random.seed(self.config.args.random_seed)
np.random.seed(self.config.args.random_seed)
with self.graph.as_default():
tf.set_random_seed(self.config.args.random_seed)
# Enable the created model parts to find this experiment.
type(self)._current_experiment = self # type: ignore
self.config.build_model(warn_unused=self.train_mode)
type(self)._current_experiment = None
self._model = self.config.model
self._model_built = True
if self.model.runners_batch_size is None:
self.model.runners_batch_size = self.model.batch_size
if self.model.tf_manager is None:
self.model.tf_manager = get_default_tf_manager()
if self.train_mode:
check_dataset_and_coders(self.model.train_dataset,
self.model.runners)
if isinstance(self.model.val_dataset, Dataset):
check_dataset_and_coders(self.model.val_dataset,
self.model.runners)
else:
for val_dataset in self.model.val_dataset:
check_dataset_and_coders(val_dataset,
self.model.runners)
if self.train_mode and self.model.visualize_embeddings:
visualize_embeddings(self.model.visualize_embeddings,
self.model.output)
self._check_unused_initializers()