def accumulate(self, f0_target, f0_pred, seq_len, is_voiced):
sequence_mask = utils.sequence_mask(seq_len,
max_len=f0_target.shape[1],
dtype=f0_target.dtype)
mask = sequence_mask * is_voiced.type(f0_target.dtype)
# Accumulate the squared difference.
square_diff = (f0_target - f0_pred)**2
self.sum += torch.sum(square_diff * mask)
self.count += torch.sum(mask).item()
def accumulate(self, tensor, seq_len=None):
r"""tensor much have shape [batch_size, seq_len, feat_dim]."""
if seq_len is None:
self.sum += torch.sum(tensor)
self.count += tensor.numel()
else:
sequence_mask = utils.sequence_mask(seq_len,
max_len=tensor.shape[1],
dtype=tensor.dtype)
self.sum += torch.sum(tensor * sequence_mask)
self.count += torch.sum(sequence_mask).item()
def accumulate(self, f0_target, f0_pred, is_voiced, seq_len=None):
StatefulMetric.accumulate(self)
mask = is_voiced.type(f0_target.dtype)
if seq_len is not None:
sequence_mask = utils.sequence_mask(seq_len,
max_len=f0_target.shape[1],
dtype=f0_target.dtype)
mask *= sequence_mask
# Accumulate the squared difference.
square_diff = (f0_target - f0_pred)**2
self.sum += torch.sum(square_diff * mask)
self.count += torch.sum(mask).item()
def wrapped_loss(predictions, targets, seq_len=None):
feature_loss = loss_fn(predictions, targets)
if seq_len is None:
max_num_frames = feature_loss.shape[1]
feature_loss = torch.sum(feature_loss, dim=1) / max_num_frames
else:
mask = utils.sequence_mask(seq_len,
max_len=feature_loss.shape[1],
dtype=feature_loss.dtype)
num_valid_frames = torch.sum(mask, dim=1)
feature_loss = torch.sum(feature_loss * mask,
dim=1) / num_valid_frames
# Average across all batch items and all feature dimensions.
feature_loss = torch.mean(feature_loss)
return feature_loss