def __init__(
self,
input_features_def,
combiner_def,
output_features_def,
random_seed=None,
):
# Deep copy to prevent TensorFlow from hijacking the dicts within the config and
# transforming them into _DictWrapper classes, which are not JSON serializable.
self._input_features_df = copy.deepcopy(input_features_def)
self._combiner_def = copy.deepcopy(combiner_def)
self._output_features_df = copy.deepcopy(output_features_def)
self._random_seed = random_seed
if random_seed is not None:
torch.random.manual_seed(random_seed)
super().__init__()
# ================ Inputs ================
self.input_features = torch.nn.ModuleDict()
self.input_features.update(build_inputs(input_features_def))
# ================ Combiner ================
logger.debug(f"Combiner {combiner_def[TYPE]}")
combiner_class = get_combiner_class(combiner_def[TYPE])
config, kwargs = load_config_with_kwargs(
combiner_class.get_schema_cls(),
combiner_def,
)
self.combiner = combiner_class(input_features=self.input_features, config=config, **kwargs)
# ================ Outputs ================
self.output_features = torch.nn.ModuleDict()
def __init__(
self,
input_features_def,
combiner_def,
output_features_def,
**kwargs
):
super().__init__()
# ================ Inputs ================
self.input_features = build_inputs(
input_features_def
)
# ================ Combiner ================
logger.debug('Combiner {}'.format(combiner_def[TYPE]))
combiner_class = get_combiner_class(combiner_def[TYPE])
self.combiner = combiner_class(
input_features=self.input_features,
**combiner_def,
**kwargs
)
# ================ Outputs ================
self.output_features = build_outputs(
output_features_def,
self.combiner
)
# ================ Combined loss metric ================
self.eval_loss_metric = tf.keras.metrics.Mean()
# After constructing all layers, clear the cache to free up memory
clear_data_cache()
def __init__(
self,
input_features_def,
combiner_def,
output_features_def,
random_seed=None,
):
self._input_features_def = copy.deepcopy(input_features_def)
self._combiner_def = copy.deepcopy(combiner_def)
self._output_features_def = copy.deepcopy(output_features_def)
self._random_seed = random_seed
if random_seed is not None:
torch.random.manual_seed(random_seed)
super().__init__()
# ================ Inputs ================
self.input_features = LudwigFeatureDict()
try:
self.input_features.update(build_inputs(self._input_features_def))
except KeyError as e:
raise KeyError(
f"An input feature has a name that conflicts with a class attribute of torch's ModuleDict: {e}"
)
# ================ Combiner ================
logger.debug(f"Combiner {combiner_def[TYPE]}")
combiner_class = get_combiner_class(combiner_def[TYPE])
config, kwargs = load_config_with_kwargs(
combiner_class.get_schema_cls(),
combiner_def,
)
self.combiner = combiner_class(input_features=self.input_features, config=config, **kwargs)
# ================ Outputs ================
self.output_features = LudwigFeatureDict()
self.output_features.update(build_outputs(self._output_features_def, self.combiner))
# ================ Combined loss metric ================
self.eval_loss_metric = torchmetrics.MeanMetric()
self.eval_additional_losses_metrics = torchmetrics.MeanMetric()
# After constructing all layers, clear the cache to free up memory
clear_data_cache()
def __init__(
self,
input_features_def,
combiner_def,
output_features_def,
random_seed=None,
):
# Deep copy to prevent TensorFlow from hijacking the dicts within the config and
# transforming them into _DictWrapper classes, which are not JSON serializable.
self._input_features_df = copy.deepcopy(input_features_def)
self._combiner_def = copy.deepcopy(combiner_def)
self._output_features_df = copy.deepcopy(output_features_def)
self._random_seed = random_seed
if random_seed is not None:
tf.random.set_seed(random_seed)
super().__init__()
# ================ Inputs ================
self.input_features = build_inputs(
input_features_def
)
# ================ Combiner ================
logger.debug('Combiner {}'.format(combiner_def[TYPE]))
combiner_class = get_combiner_class(combiner_def[TYPE])
self.combiner = combiner_class(
input_features=self.input_features,
**combiner_def,
)
# ================ Outputs ================
self.output_features = build_outputs(
output_features_def,
self.combiner
)
# ================ Combined loss metric ================
self.eval_loss_metric = tf.keras.metrics.Mean()
# After constructing all layers, clear the cache to free up memory
clear_data_cache()