Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
decoding
游雁407625a73478
diff --git a/funasr/bin/train_ds.py b/funasr/bin/train_ds.pyindex 67ed2ba93..41ecbe4bc 100644--- a/funasr/bin/train_ds.py+++ b/funasr/bin/train_ds.py@@ -124,6 +124,7 @@ def main(**kwargs):use_ddp=use_ddp,use_fsdp=use_fsdp,device=kwargs["device"],+ excludes=kwargs.get("excludes", None),output_dir=kwargs.get("output_dir", "./exp"),**kwargs.get("train_conf"),)diff --git a/funasr/train_utils/trainer_ds.py b/funasr/train_utils/trainer_ds.pyindex ec76531d6..c97372851 100644--- a/funasr/train_utils/trainer_ds.py+++ b/funasr/train_utils/trainer_ds.py@@ -147,6 +147,10 @@ class Trainer:self.use_deepspeed = use_deepspeedself.deepspeed_config = kwargs.get("deepspeed_config", "")+ self.excludes = kwargs.get("excludes", None)+ if self.excludes is not None:+ if isinstance(self.excludes, str):+ self.excludes = self.excludes.split(",")def save_checkpoint(self,@@ -440,6 +444,12 @@ class Trainer:src_state = checkpoint["state_dict"]dst_state = model.state_dict()for k in dst_state.keys():+ if excludes is not None:+ for k_ex in excludes:+ k_tmp = k.replace("module.", "")+ if k_tmp.startswith(k_ex):+ logging.info(f"key: {{k}} matching: {k_ex}, excluded")+ continueif not k.startswith("module.") and "module." + k in src_state.keys():k_ddp = "module." + kelif k.startswith("module.") and "module." + k not in src_state.keys():