dpo改实验性勾选而非必须。勾选后batch size自动减半。

dpo改实验性勾选而非必须。勾选后batch size自动减半。

Author: RVC-Boss

GPT_SoVITS/AR/models/t2s_lightning_module.py

@@ -11,7 +11,6 @@
 from AR.modules.lr_schedulers import WarmupCosineLRSchedule
 from AR.modules.optim import ScaledAdam
 
-
 class Text2SemanticLightningModule(LightningModule):
     def __init__(self, config, output_dir, is_train=True):
         super().__init__()
@@ -35,7 +34,8 @@ def __init__(self, config, output_dir, is_train=True):
     def training_step(self, batch: Dict, batch_idx: int):
         opt = self.optimizers()
         scheduler = self.lr_schedulers()
-        loss, acc = self.model.forward(
+        forward=self.model.forward if self.config["train"].get("if_dpo",False)==True else self.model.forward_old
+        loss, acc = forward(
             batch["phoneme_ids"],
             batch["phoneme_ids_len"],
             batch["semantic_ids"],