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
Dev gzf (#1402)
zhifu gaob9cfd9953a88
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.pyindex 1abdcea57..a6be69115 100644--- a/funasr/auto/auto_model.py+++ b/funasr/auto/auto_model.py@@ -193,7 +193,7 @@ class AutoModel:path=init_param,ignore_init_mismatch=kwargs.get("ignore_init_mismatch", False),oss_bucket=kwargs.get("oss_bucket", None),- scope_map=kwargs.get("scope_map", "module.,None"),+ scope_map=kwargs.get("scope_map", []),excludes=kwargs.get("excludes", None),)else:diff --git a/funasr/bin/train.py b/funasr/bin/train.pyindex 6650f0a87..569757a6a 100644--- a/funasr/bin/train.py+++ b/funasr/bin/train.py@@ -105,7 +105,7 @@ def main(**kwargs):path=p,ignore_init_mismatch=kwargs.get("ignore_init_mismatch", True),oss_bucket=kwargs.get("oss_bucket", None),- scope_map=kwargs.get("scope_map", "module.,none"),+ scope_map=kwargs.get("scope_map", []),excludes=kwargs.get("excludes", None),)else:diff --git a/funasr/models/llm_asr_nar/model.py b/funasr/models/llm_asr_nar/model.pyindex a61190c1d..6a4eccefe 100644--- a/funasr/models/llm_asr_nar/model.py+++ b/funasr/models/llm_asr_nar/model.py@@ -315,7 +315,7 @@ class LLMASRNAR(nn.Module):model_outputs = self.llm(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=None)preds = torch.argmax(model_outputs.logits, -1)text = tokenizer.batch_decode(preds, add_special_tokens=False, skip_special_tokens=True)- text = text.split(': "\n')[-1]+ text = text[0].split(': \n')[-1]# preds = torch.argmax(model_outputs.logits, -1)ibest_writer = Nonediff --git a/funasr/train_utils/load_pretrained_model.py b/funasr/train_utils/load_pretrained_model.pyindex 23a6ef504..84c632027 100644--- a/funasr/train_utils/load_pretrained_model.py+++ b/funasr/train_utils/load_pretrained_model.py@@ -38,52 +38,17 @@ def filter_state_dict()return match_state-def assigment_scope_map(dst_state: dict, src_state: dict, scope_map: str=None):- """Compute the union of the current variables and checkpoint variables."""- import collections- import re-- # current model variables- name_to_variable = collections.OrderedDict()- for name, var in dst_state.items():- name_to_variable[name] = var-- scope_map_num = 0- if scope_map is not None:- scope_map = scope_map.split(",")- scope_map_num = len(scope_map) // 2- for scope_map_idx in range(scope_map_num):- scope_map_id = scope_map_idx * 2- logging.info('assignment_map from scope {} to {}'.format(scope_map[scope_map_id], scope_map[scope_map_id+1]))-- assignment_map = {}- for name, var in src_state.items():-- if scope_map:- for scope_map_idx in range(scope_map_num):- scope_map_id = scope_map_idx * 2- try:- idx = name.index(scope_map[scope_map_id])- new_name = scope_map[scope_map_id+1] + name[idx + len(scope_map[scope_map_id]):]- if new_name in name_to_variable:- assignment_map[name] = var- except:- continue- else:- if name in name_to_variable:- assignment_map[name] = var-- return assignment_map-def load_pretrained_model(path: str,model: torch.nn.Module,- ignore_init_mismatch: bool,+ ignore_init_mismatch: bool=True,map_location: str = "cpu",oss_bucket=None,- scope_map="module.:none",+ scope_map=[],excludes=None,+ ignore_mismatch=False,+ **kwargs,):"""Load a model state and set it to the model.@@ -110,12 +75,10 @@ def load_pretrained_model(if isinstance(scope_map, str):scope_map = scope_map.split(",")+ scope_map += ["module.", "None"]for k in dst_state.keys():- # if not k.startswith("module.") and "module." + k in src_state.keys():- # k_ddp = "module." + k- # else:- # k_ddp = k+k_src = kif scope_map is not None:@@ -124,66 +87,25 @@ def load_pretrained_model(for i in range(0, len(scope_map), 2):src_prefix = scope_map[i] if scope_map[i].lower() != "none" else ""dst_prefix = scope_map[i+1] if scope_map[i+1].lower() != "none" else ""-- if k.startswith(dst_prefix) and k.replace(dst_prefix, src_prefix) in src_state.keys():- k_src = k.replace(dst_prefix, src_prefix)- print(f"init param, map: {k} from {k_src} in ckpt")++ if dst_prefix == "" and (src_prefix + k) in src_state.keys():+ k_src = src_prefix + k+ if not k_src.startswith("module."):+ print(f"init param, map: {k} from {k_src} in ckpt")+ elif k.startswith(dst_prefix) and k.replace(dst_prefix, src_prefix, 1) in src_state.keys():+ k_src = k.replace(dst_prefix, src_prefix, 1)+ if not k_src.startswith("module."):+ print(f"init param, map: {k} from {k_src} in ckpt")if k_src in src_state.keys():- dst_state[k] = src_state[k_src]-- # if k_ddp.startswith("audio_encoder"):- # if k_ddp.replace("audio_encoder", "encoder.model") in src_state.keys():- # k_ddp = k_ddp.replace("audio_encoder", "encoder.model")- # if k_ddp.startswith("adaptor"):- # if k_ddp.replace("adaptor", "encoder_projector") in src_state.keys():- # k_ddp = k_ddp.replace("adaptor", "encoder_projector")- # if k_ddp in src_state:- # dst_state[k] = src_state[k_ddp]+ if ignore_init_mismatch and dst_state[k].shape != src_state[k_src].shape:+ print(f"ignore_mismatch:{ignore_mismatch}, dst: {k, dst_state[k].shape}, src: {k_src, src_state[k_src].shape}")+ else:+ dst_state[k] = src_state[k_src]++else:print(f"Warning, miss key in ckpt: {k}, mapped: {k_src}")- flag = obj.load_state_dict(dst_state, strict=False)+ flag = obj.load_state_dict(dst_state, strict=True)# print(flag)--# def load_pretrained_model(-# path: str,-# model: torch.nn.Module,-# ignore_init_mismatch: bool,-# map_location: str = "cpu",-# oss_bucket=None,-# scope_map=None,-# excludes=None,-# ):-# """Load a model state and set it to the model.-#-# Args:-# init_param: <file_path>:<src_key>:<dst_key>:<exclude_Keys>-#-# Examples:-#-# """-#-# obj = model-#-# if oss_bucket is None:-# src_state = torch.load(path, map_location=map_location)-# else:-# buffer = BytesIO(oss_bucket.get_object(path).read())-# src_state = torch.load(buffer, map_location=map_location)-# src_state = src_state["model"] if "model" in src_state else src_state-#-# if excludes is not None:-# for e in excludes.split(","):-# src_state = {k: v for k, v in src_state.items() if not k.startswith(e)}-#-# dst_state = obj.state_dict()-# src_state = assigment_scope_map(dst_state, src_state, scope_map)-#-# if ignore_init_mismatch:-# src_state = filter_state_dict(dst_state, src_state)-#-# logging.debug("Loaded src_state keys: {}".format(src_state.keys()))-# logging.debug("Loaded dst_state keys: {}".format(dst_state.keys()))-# dst_state.update(src_state)-# obj.load_state_dict(dst_state, strict=True)