from enum import Enum
from torch import nn
[docs]
class PredictionType(Enum):
"""Prediction types"""
DETERMINISTIC = 1
"""Deterministic prediction
Non-MDN single-stream models should use this type.
Pseudo code:
.. code-block::
# training
y = model(x)
# inference
y = model.inference(x)
"""
PROBABILISTIC = 2
"""Probabilistic prediction with mixture density networks
MDN-based models should use this type.
Pseudo code:
.. code-block::
# training
mdn_params = model(x)
# inference
mu, sigma = model.inference(x)
"""
MULTISTREAM_HYBRID = 3
"""Multi-stream preodictions where each prediction can be
detereministic or probabilistic
Multi-stream models should use this type.
Pseudo code:
.. code-block::
# training
feature_streams = model(x) # e.g. (mgc, lf0, vuv, bap) or (mel, lf0, vuv)
# inference
y = model.inference(x)
Note that concatenated features are assumed to be returned during inference.
"""
DIFFUSION = 4
"""Diffusion model's prediction
NOTE: may subject to change in the future
Pseudo code:
.. code-block::
# training
noise, x_recon = model(x)
# inference
y = model.inference(x)
"""
[docs]
class BaseModel(nn.Module):
"""Base class for all models
If you want to implement your custom model, you should inherit from this class.
You must need to implement the forward method. Other methods are optional.
"""
[docs]
def forward(self, x, lengths=None, y=None):
"""Forward pass
Args:
x (tensor): input features
lengths (tensor): lengths of the input features
y (tensor): optional target features
Returns:
tensor: output features
"""
pass
[docs]
def inference(self, x, lengths=None):
"""Inference method
If you want to implement custom inference method such as autoregressive sampling,
please override this method.
Defaults to call the forward method.
Args:
x (tensor): input features
lengths (tensor): lengths of the input features
Returns:
tensor: output features
"""
return self(x, lengths)
[docs]
def preprocess_target(self, y):
"""Preprocess target signals at training time
This is useful for shallow AR models in which a FIR filter
is used for the target signals. For other types of model, you don't need to
implement this method.
Defaults to do nothing.
Args:
y (tensor): target features
Returns:
tensor: preprocessed target features
"""
return y
[docs]
def prediction_type(self):
"""Prediction type.
If your model has a MDN layer, please return ``PredictionType.PROBABILISTIC``.
Returns:
PredictionType: Determisitic or probabilistic. Default is deterministic.
"""
return PredictionType.DETERMINISTIC
[docs]
def is_autoregressive(self):
"""Is autoregressive or not
If your custom model is an autoregressive model, please return ``True``. In that case,
you would need to implement autoregressive sampling in :py:meth:`inference`.
Returns:
bool: True if autoregressive. Default is False.
"""
return False
[docs]
def has_residual_lf0_prediction(self):
"""Whether the model has residual log-F0 prediction or not.
This should only be used for acoustic models.
Returns:
bool: True if the model has residual log-F0 prediction. Default is False.
"""
return False