From 9fb473bc89f64e5e084379a235b8fe1dde1e3ee3 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 23 二月 2024 16:46:50 +0800
Subject: [PATCH] update
---
funasr/auto/auto_model.py | 72 +++++++++++++++++++++++-------------
1 files changed, 46 insertions(+), 26 deletions(-)
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index e4a154d..48a983c 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -1,14 +1,13 @@
import json
import time
+import copy
import torch
-import hydra
import random
import string
import logging
import os.path
import numpy as np
from tqdm import tqdm
-from omegaconf import DictConfig, OmegaConf, ListConfig
from funasr.register import tables
from funasr.utils.load_utils import load_bytes
@@ -17,7 +16,7 @@
from funasr.utils.vad_utils import slice_padding_audio_samples
from funasr.train_utils.set_all_random_seed import set_all_random_seed
from funasr.train_utils.load_pretrained_model import load_pretrained_model
-from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
+from funasr.utils.load_utils import load_audio_text_image_video
from funasr.utils.timestamp_tools import timestamp_sentence
from funasr.models.campplus.utils import sv_chunk, postprocess, distribute_spk
try:
@@ -158,8 +157,10 @@
tokenizer_class = tables.tokenizer_classes.get(tokenizer)
tokenizer = tokenizer_class(**kwargs["tokenizer_conf"])
kwargs["tokenizer"] = tokenizer
- kwargs["token_list"] = tokenizer.token_list
- vocab_size = len(tokenizer.token_list)
+
+ kwargs["token_list"] = tokenizer.token_list if hasattr(tokenizer, "token_list") else None
+ kwargs["token_list"] = tokenizer.get_vocab() if hasattr(tokenizer, "get_vocab") else kwargs["token_list"]
+ vocab_size = len(kwargs["token_list"])
else:
vocab_size = -1
@@ -180,15 +181,18 @@
# init_param
init_param = kwargs.get("init_param", None)
if init_param is not None:
- logging.info(f"Loading pretrained params from {init_param}")
- load_pretrained_model(
- model=model,
- 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", None),
- excludes=kwargs.get("excludes", None),
- )
+ if os.path.exists(init_param):
+ logging.info(f"Loading pretrained params from {init_param}")
+ load_pretrained_model(
+ model=model,
+ 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", None),
+ excludes=kwargs.get("excludes", None),
+ )
+ else:
+ print(f"error, init_param does not exist!: {init_param}")
return model, kwargs
@@ -380,16 +384,22 @@
result[k] = restored_data[j][k]
else:
result[k] += restored_data[j][k]
-
+
+ return_raw_text = kwargs.get('return_raw_text', False)
# step.3 compute punc model
if self.punc_model is not None:
self.punc_kwargs.update(cfg)
punc_res = self.inference(result["text"], model=self.punc_model, kwargs=self.punc_kwargs, disable_pbar=True, **cfg)
- import copy; raw_text = copy.copy(result["text"])
+ raw_text = copy.copy(result["text"])
+ if return_raw_text: result['raw_text'] = raw_text
result["text"] = punc_res[0]["text"]
+ else:
+ raw_text = None
# speaker embedding cluster after resorted
if self.spk_model is not None and kwargs.get('return_spk_res', True):
+ if raw_text is None:
+ logging.error("Missing punc_model, which is required by spk_model.")
all_segments = sorted(all_segments, key=lambda x: x[0])
spk_embedding = result['spk_embedding']
labels = self.cb_model(spk_embedding.cpu(), oracle_num=kwargs.get('preset_spk_num', None))
@@ -398,20 +408,30 @@
if self.spk_mode == 'vad_segment': # recover sentence_list
sentence_list = []
for res, vadsegment in zip(restored_data, vadsegments):
- sentence_list.append({"start": vadsegment[0],\
- "end": vadsegment[1],
- "sentence": res['text'],
- "timestamp": res['timestamp']})
+ if 'timestamp' not in res:
+ logging.error("Only 'iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch' \
+ and 'iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch'\
+ can predict timestamp, and speaker diarization relies on timestamps.")
+ sentence_list.append({"start": vadsegment[0],
+ "end": vadsegment[1],
+ "sentence": res['text'],
+ "timestamp": res['timestamp']})
elif self.spk_mode == 'punc_segment':
- sentence_list = timestamp_sentence(punc_res[0]['punc_array'], \
- result['timestamp'], \
- result['text'])
+ if 'timestamp' not in result:
+ logging.error("Only 'iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch' \
+ and 'iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch'\
+ can predict timestamp, and speaker diarization relies on timestamps.")
+ sentence_list = timestamp_sentence(punc_res[0]['punc_array'],
+ result['timestamp'],
+ raw_text,
+ return_raw_text=return_raw_text)
distribute_spk(sentence_list, sv_output)
result['sentence_info'] = sentence_list
elif kwargs.get("sentence_timestamp", False):
- sentence_list = timestamp_sentence(punc_res[0]['punc_array'], \
- result['timestamp'], \
- result['text'])
+ sentence_list = timestamp_sentence(punc_res[0]['punc_array'],
+ result['timestamp'],
+ raw_text,
+ return_raw_text=return_raw_text)
result['sentence_info'] = sentence_list
if "spk_embedding" in result: del result['spk_embedding']
--
Gitblit v1.9.1