Use the last attention value as a threshold to stop decoding. since stoptoken prediction is not precies enough to synthsis at the right time.

2026-09-09 11:16:00 +08:00 · 2019-01-16 12:32:40 +01:00
parent ed1f648b83
commit 8969d59902
1 changed files with 1 additions and 1 deletions
@@ -434,7 +434,7 @@ class Decoder(nn.Module):
                if t >= T_decoder:
                    break
            else:
-                if t > inputs.shape[1] / 4 and stop_token > 0.6:
+                if t > inputs.shape[1] / 4 and (stop_token > 0.6 or attention[:, -1].item() > 0.6):
                    break
                elif t > self.max_decoder_steps:
                    print("   | > Decoder stopped with 'max_decoder_steps")