Reverted to l1 loss for durations.

geneing · Jul 6, 2020 · 07dd1ac · 07dd1ac
1 parent 6b275f8
commit 07dd1ac
Show file tree

Hide file tree

Showing 2 changed files with 6 additions and 3 deletions.
diff --git a/hparams.py b/hparams.py
@@ -120,7 +120,8 @@
 # Training
 
 forward_schedule = [(1e-4, 10_000,  32),    # progressive training schedule
-                    (1e-4, 300_000,  32)]   # (lr, step, batch_size)
+                    (1e-4, 300_000,  32),   # (lr, step, batch_size)
+                    (2e-5, 600_000,  32)]   # (lr, step, batch_size)
 
 forward_max_mel_len = 1250              # if you have a couple of extremely long spectrograms you might want to use this
 forward_clip_grad_norm = 1.0            # clips the gradient norm to prevent explosion - set to None if not needed

diff --git a/trainer/forward_trainer.py b/trainer/forward_trainer.py
@@ -1,4 +1,6 @@
 import time
+import torch.nn.functional as F
+
 from typing import Tuple
 
 import torch
@@ -65,7 +67,7 @@ def train_session(self, model: ForwardTacotron,
                 m1_loss = self.l1_loss(m1_hat, m, lens)
                 m2_loss = self.l1_loss(m2_hat, m, lens)
 
-                dur_loss = self.log_l1_loss(dur_hat, dur)
+                dur_loss = F.l1_loss(dur_hat, dur)
 
                 loss = m1_loss + m2_loss + dur_loss
                 optimizer.zero_grad()
@@ -117,7 +119,7 @@ def evaluate(self, model: ForwardTacotron, val_set: Dataset) -> Tuple[float, flo
                 m1_hat, m2_hat, dur_hat = model(x, m, dur)
                 m1_loss = self.l1_loss(m1_hat, m, lens)
                 m2_loss = self.l1_loss(m2_hat, m, lens)
-                dur_loss = self.log_l1_loss(dur_hat, dur)
+                dur_loss = F.l1_loss(dur_hat, dur)
                 m_val_loss += m1_loss.item() + m2_loss.item()
                 dur_val_loss += dur_loss.item()
         return m_val_loss / len(val_set), dur_val_loss / len(val_set)