def output_types(self):
return {
"real_scores": NeuralType(('B', 'T'), VoidType()),
"fake_scores": NeuralType(('B', 'T'), VoidType()),
"real_feature_maps": NeuralType(("B", "C", "H", "W"), VoidType()),
"fake_feature_maps": NeuralType(("B", "C", "H", "W"), VoidType()),
}
def output_types(self):
return {
"real_scores": [NeuralType(('B', 'T'), VoidType())],
"fake_scores": [NeuralType(('B', 'T'), VoidType())],
"real_feature_maps": [[NeuralType(("B", "C", "T"), VoidType())]],
"fake_feature_maps": [[NeuralType(("B", "C", "T"), VoidType())]],
}
def output_types(self):
if self.mode == OperationMode.training or self.mode == OperationMode.validation:
return {
"pred_normal_dist": NeuralType(('B', 'flowgroup', 'T'), NormalDistributionSamplesType()),
"log_s_list": [NeuralType(('B', 'flowgroup', 'T'), VoidType())], # TODO: Figure out a good typing
"log_det_W_list": [NeuralType(elements_type=VoidType())], # TODO: Figure out a good typing
"audio_pred": NeuralType(('B', 'T'), AudioSignal()),
}
else:
return {
"audio": NeuralType(('B', 'T'), AudioSignal()),
}
def input_types(self):
return {
"z":
NeuralType(('B', 'flowgroup', 'T'),
NormalDistributionSamplesType()),
"log_s_list":
[NeuralType(('B', 'flowgroup', 'T'),
VoidType())], # TODO: Figure out a good typing
"log_det_W_list": [NeuralType(elements_type=VoidType())
], # TODO: Figure out a good typing
"sigma":
NeuralType(optional=True),
}
def output_types(self):
if self.mode == OperationMode.training or self.mode == OperationMode.validation:
output_dict = {
"pred_normal_dist": NeuralType(('B', 'flowgroup', 'T'), NormalDistributionSamplesType()),
"log_s_list": NeuralType(('B', 'flowgroup', 'T'), VoidType()), # TODO: Figure out a good typing
"log_det_W_list": NeuralType(elements_type=VoidType()), # TODO: Figure out a good typing
}
if self.mode == OperationMode.validation:
output_dict["audio_pred"] = NeuralType(('B', 'T'), AudioSignal())
output_dict["spec"] = NeuralType(('B', 'T', 'D'), MelSpectrogramType())
output_dict["spec_len"] = NeuralType(('B'), LengthsType())
return output_dict
return {
"audio_pred": NeuralType(('B', 'T'), AudioSignal()),
}
def input_types(self):
return {
"z": NeuralType(('B', 'flowgroup', 'T'), NormalDistributionSamplesType()),
"logdet": NeuralType(elements_type=VoidType()),
"gt_audio": NeuralType(('B', 'T'), AudioSignal()),
"predicted_audio": NeuralType(('B', 'T'), AudioSignal()),
"sigma": NeuralType(optional=True),
}
def __init__(self,
axes: Optional[Tuple] = None,
elements_type: ElementType = VoidType(),
optional=False):
if not isinstance(elements_type, ElementType):
raise ValueError(
"elements_type of NeuralType must be an instance of a class derived from ElementType. "
"Did you pass a class instead?")
self.elements_type = elements_type
if axes is not None:
NeuralType.__check_sanity(axes)
axes_list = []
for axis in axes:
if isinstance(axis, str):
axes_list.append(AxisType(AxisKind.from_str(axis), None))
elif isinstance(axis, AxisType):
axes_list.append(axis)
else:
raise ValueError(
"axis type must be either str or AxisType instance")
self.axes = tuple(axes_list)
else:
self.axes = None
self.optional = optional
def output_types(self):
return {
"decision": NeuralType(('B', 'T'), VoidType()),
"feature_maps": [NeuralType(("B", "C", "T"), VoidType())],
}
def output_types(self):
return {
"decision": NeuralType((('B', 'S', 'T')), VoidType()),
}
def input_types(self):
return {
"disc_real_outputs": NeuralType(('B', 'T'), VoidType()),
"disc_generated_outputs": NeuralType(('B', 'T'), VoidType()),
}
def input_types(self):
return {
"fmap_r": NeuralType(elements_type=VoidType()),
"fmap_g": NeuralType(elements_type=VoidType()),
}
def input_types(self):
return {
"disc_outputs": NeuralType(('B', 'T'), VoidType()),
}
