Huggingface

2024-04-11

源切换

镜像源：https://hf-mirror.com/

安装hfd.sh依赖以下包

apt install git-lfs
apt install aria2

工程推荐配置

# ~/.bashrc
export HF_ENDPOINT=https://hf-mirror.com
export HF_HUB_ENABLE_HF_TRANSFER=1
export HF_HUB_DOWNLOAD_TIMEOUT=600
export TRANSFORMERS_CACHE=/mnt/hf_cache/hub    # Transformers 库的模型和分词器缓存
export HF_HOME=/mnt/hf_cache                   # 所有 Hugging Face 库（Transformers、Datasets、Evaluate等）的通用根缓存目录

强制离线环境配置

export HF_HUB_OFFLINE=1
export TRANSFORMERS_OFFLINE=1
export HF_DATASETS_OFFLINE=1
export TRANSFORMERS_CACHE=/mnt/hf_cache/hub    # Transformers 库的模型和分词器缓存

在线模型依赖切换到本地模型

下载模型

./hfd.sh BELLE-2/Belle-whisper-large-v3-zh --tool aria2c -x 4

下载模块

pip install transformers

from transformers import pipeline

transcriber = pipeline(
  "automatic-speech-recognition", 
  model="BELLE-2/Belle-whisper-large-v3-zh"
)

transcriber.model.config.forced_decoder_ids = (
  transcriber.tokenizer.get_decoder_prompt_ids(
    language="zh", 
    task="transcribe"
  )
)

transcription = transcriber("my_audio.wav")

print(transcription)

切换到本地模型

transcriber = pipeline(
  "automatic-speech-recognition", 
  model=r"c:/01_cloud/hf/Belle-whisper-large-v3-zh"
)

Transformers

https://huggingface.co/docs/transformers/main/en/main_classes/pipelines