From acb9a0fec8d8a4dabeedcbb8e08c26f66d7083f0 Mon Sep 17 00:00:00 2001
From: haoneng.lhn <haoneng.lhn@alibaba-inc.com>
Date: 星期五, 08 十二月 2023 16:19:00 +0800
Subject: [PATCH] fix loss normalization for ddp training

---
 funasr/models/e2e_asr_contextual_paraformer.py |    3 +++
 1 files changed, 3 insertions(+), 0 deletions(-)

diff --git a/funasr/models/e2e_asr_contextual_paraformer.py b/funasr/models/e2e_asr_contextual_paraformer.py
index a2f7078..b474dbc 100644
--- a/funasr/models/e2e_asr_contextual_paraformer.py
+++ b/funasr/models/e2e_asr_contextual_paraformer.py
@@ -125,6 +125,7 @@
         if self.crit_attn_weight > 0:
             self.attn_loss = torch.nn.L1Loss()
         self.crit_attn_smooth = crit_attn_smooth
+        self.length_normalized_loss = length_normalized_loss
 
     def forward(
             self,
@@ -231,6 +232,8 @@
 
         stats["loss"] = torch.clone(loss.detach())
         # force_gatherable: to-device and to-tensor if scalar for DataParallel
+        if self.length_normalized_loss:
+            batch_size = int((text_lengths + self.predictor_bias).sum())
         loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
         return loss, stats, weight
     

--
Gitblit v1.9.1