|
""" |
|
Taken from ESPNet |
|
Adapted by Flux |
|
""" |
|
|
|
import torch |
|
|
|
from Utility.utils import make_non_pad_mask |
|
|
|
|
|
class StochasticToucanTTSLoss(torch.nn.Module): |
|
|
|
def __init__(self): |
|
super().__init__() |
|
self.l1_criterion = torch.nn.L1Loss(reduction="none") |
|
|
|
def forward(self, predicted_features, gold_features, features_lengths): |
|
""" |
|
Args: |
|
predicted_features (Tensor): Batch of outputs (B, Lmax, odim). |
|
gold_features (Tensor): Batch of target features (B, Lmax, odim). |
|
features_lengths (LongTensor): Batch of the lengths of each target (B,). |
|
|
|
Returns: |
|
Tensor: L1 loss value. |
|
""" |
|
|
|
|
|
l1_loss = self.l1_criterion(predicted_features, gold_features) |
|
|
|
|
|
out_masks = make_non_pad_mask(features_lengths).unsqueeze(-1).to(gold_features.device) |
|
out_masks = torch.nn.functional.pad(out_masks.transpose(1, 2), [0, gold_features.size(1) - out_masks.size(1), 0, 0, 0, 0], value=False).transpose(1, 2) |
|
out_weights = out_masks.float() / out_masks.sum(dim=1, keepdim=True).float() |
|
out_weights /= gold_features.size(0) * gold_features.size(2) |
|
|
|
|
|
l1_loss = l1_loss.mul(out_weights).masked_select(out_masks).sum() |
|
|
|
return l1_loss |
|
|