Skip to content

SOTA Open Source TTS

optimize dpo behavior

Lengyue0e0332396fbd
diff --git a/fish_speech/models/text2semantic/lit_module.py b/fish_speech/models/text2semantic/lit_module.py
index ddffc75..02a92fb 100644
--- a/fish_speech/models/text2semantic/lit_module.py
+++ b/fish_speech/models/text2semantic/lit_module.py
@@ -197,9 +197,7 @@ class TextToSemantic(L.LightningModule):
chosen_rewards = self.dpo_beta * positive_codebook_logps.detach()
rejected_rewards = self.dpo_beta * negative_codebook_logps.detach()
- reward_accuracy = (
- (positive_codebook_logps > negative_codebook_logps).float().mean()
- )
+ reward_accuracy = (chosen_rewards > rejected_rewards).float().mean()
chosen_rewards, rejected_rewards = (
chosen_rewards.mean(),
rejected_rewards.mean(),