Merge pull request #1161 from alibaba-damo-academy/dev_lhn

hnluo

2023-12-08 202ab8a2c9e2af5c147faf080f96e97abbb7be42

Merge pull request #1161 from alibaba-damo-academy/dev_lhn

fix loss normalization for ddp training

 funasr/models/e2e_uni_asr.py

@@ -442,7 +442,7 @@
        stats["loss"] = torch.clone(loss.detach())
        # force_gatherable: to-device and to-tensor if scalar for DataParallel
        if self.length_normalized_loss:
            batch_size = (text_lengths + 1).sum().type_as(batch_size)
            batch_size = int((text_lengths + 1).sum())
        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
        return loss, stats, weight