阅读量:0
快速方便地下载huggingface的模型库和数据集
方法一：用于使用 aria2/wget+git 下载 Huggingface 模型和数据集的 CLI 工具
来自https://gist.github.com/padeoe/697678ab8e528b85a2a7bddafea1fa4f。
使用方法：将hfd.sh拷贝过去，然后参考下面的参考命令，下载数据集或者模型
🤗Huggingface 模型下载器
考虑到官方 huggingface-cli 缺乏多线程下载支持，以及错误处理不足在 hf_transfer 中，这个命令行工具巧妙地利用 wget 或 aria2 来处理 LFS 文件，并使用 git clone 来处理其余文件。
特点
- ⏯️ 从断点恢复：您可以随时重新运行它或按 Ctrl+C。
- 🚀 多线程下载：利用多线程加速下载过程。
- 🚫 文件排除：使用
--exclude或--include跳过或指定文件，为具有重复格式的模型（例如，*.bin或*.safetensors）节省时间）。 - 🔐 身份验证支持：对于需要 Huggingface 登录的门控模型，请使用
--hf_username和--hf_token进行身份验证。 - 🪞 镜像站点支持：使用“HF_ENDPOINT”环境变量进行设置。
- 🌍代理支持：使用“HTTPS_PROXY”环境变量进行设置。
- 📦 简单：仅依赖
git、aria2c/wget。
Usage
首先，下载 hfd.sh 或克隆此存储库，然后授予脚本执行权限。
chmod a+x hfd.sh 为了方便起见，您可以创建一个别名
alias hfd="$PWD/hfd.sh" 使用说明：
$ ./hfd.sh -h Usage: hfd <repo_id> [--include include_pattern] [--exclude exclude_pattern] [--hf_username username] [--hf_token token] [--tool aria2c|wget] [-x threads] [--dataset] [--local-dir path] Description: Downloads a model or dataset from Hugging Face using the provided repo ID. Parameters: repo_id The Hugging Face repo ID in the format 'org/repo_name'. --include (Optional) Flag to specify a string pattern to include files for downloading. --exclude (Optional) Flag to specify a string pattern to exclude files from downloading. include/exclude_pattern The pattern to match against filenames, supports wildcard characters. e.g., '--exclude *.safetensor', '--include vae/*'. --hf_username (Optional) Hugging Face username for authentication. **NOT EMAIL**. --hf_token (Optional) Hugging Face token for authentication. --tool (Optional) Download tool to use. Can be aria2c (default) or wget. -x (Optional) Number of download threads for aria2c. Defaults to 4. --dataset (Optional) Flag to indicate downloading a dataset. --local-dir (Optional) Local directory path where the model or dataset will be stored. Example: hfd bigscience/bloom-560m --exclude *.safetensors hfd meta-llama/Llama-2-7b --hf_username myuser --hf_token mytoken -x 4 hfd lavita/medical-qa-shared-task-v1-toy --dataset 下载模型：
hfd bigscience/bloom-560m 下载模型需要登录
从https://huggingface.co/settings/tokens获取huggingface令牌，然后
hfd meta-llama/Llama-2-7b --hf_username YOUR_HF_USERNAME_NOT_EMAIL --hf_token YOUR_HF_TOKEN 下载模型并排除某些文件（例如.safetensors）：
hfd bigscience/bloom-560m --exclude *.safetensors 使用 aria2c 和多线程下载：
hfd bigscience/bloom-560m 输出：
下载过程中，将显示文件 URL：
$ hfd bigscience/bloom-560m --tool wget --exclude *.safetensors ... Start Downloading lfs files, bash script: wget -c https://huggingface.co/bigscience/bloom-560m/resolve/main/flax_model.msgpack # wget -c https://huggingface.co/bigscience/bloom-560m/resolve/main/model.safetensors wget -c https://huggingface.co/bigscience/bloom-560m/resolve/main/onnx/decoder_model.onnx ... # 安装包 apt update apt-get install aria2 apt-get install iftop apt-get install git-lfs #参考命令 bash /xxx/xxx/hfd.sh mmaaz60/ActivityNet-QA-Test-Videos --tool aria2c -x 16 --dataset --local-dir /xxx/xxx/ActivityNet hfd.sh
#!/usr/bin/env bash # Color definitions RED='\033[0;31m' GREEN='\033[0;32m' YELLOW='\033[1;33m' NC='\033[0m' # No Color trap 'printf "${YELLOW}\nDownload interrupted. If you re-run the command, you can resume the download from the breakpoint.\n${NC}"; exit 1' INT display_help() { cat << EOF Usage: hfd <repo_id> [--include include_pattern] [--exclude exclude_pattern] [--hf_username username] [--hf_token token] [--tool aria2c|wget] [-x threads] [--dataset] [--local-dir path] Description: Downloads a model or dataset from Hugging Face using the provided repo ID. Parameters: repo_id The Hugging Face repo ID in the format 'org/repo_name'. --include (Optional) Flag to specify a string pattern to include files for downloading. --exclude (Optional) Flag to specify a string pattern to exclude files from downloading. include/exclude_pattern The pattern to match against filenames, supports wildcard characters. e.g., '--exclude *.safetensor', '--include vae/*'. --hf_username (Optional) Hugging Face username for authentication. **NOT EMAIL**. --hf_token (Optional) Hugging Face token for authentication. --tool (Optional) Download tool to use. Can be aria2c (default) or wget. -x (Optional) Number of download threads for aria2c. Defaults to 4. --dataset (Optional) Flag to indicate downloading a dataset. --local-dir (Optional) Local directory path where the model or dataset will be stored. Example: hfd bigscience/bloom-560m --exclude *.safetensors hfd meta-llama/Llama-2-7b --hf_username myuser --hf_token mytoken -x 4 hfd lavita/medical-qa-shared-task-v1-toy --dataset EOF exit 1 } MODEL_ID=$1 shift # Default values TOOL="aria2c" THREADS=4 HF_ENDPOINT=${HF_ENDPOINT:-"https://hf-mirror.com"} while [[ $# -gt 0 ]]; do case $1 in --include) INCLUDE_PATTERN="$2"; shift 2 ;; --exclude) EXCLUDE_PATTERN="$2"; shift 2 ;; --hf_username) HF_USERNAME="$2"; shift 2 ;; --hf_token) HF_TOKEN="$2"; shift 2 ;; --tool) TOOL="$2"; shift 2 ;; -x) THREADS="$2"; shift 2 ;; --dataset) DATASET=1; shift ;; --local-dir) LOCAL_DIR="$2"; shift 2 ;; *) shift ;; esac done # Check if aria2, wget, curl, git, and git-lfs are installed check_command() { if ! command -v $1 &>/dev/null; then echo -e "${RED}$1 is not installed. Please install it first.${NC}" exit 1 fi } # Mark current repo safe when using shared file system like samba or nfs ensure_ownership() { if git status 2>&1 | grep "fatal: detected dubious ownership in repository at" > /dev/null; then git config --global --add safe.directory "${PWD}" printf "${YELLOW}Detected dubious ownership in repository, mark ${PWD} safe using git, edit ~/.gitconfig if you want to reverse this.\n${NC}" fi } [[ "$TOOL" == "aria2c" ]] && check_command aria2c [[ "$TOOL" == "wget" ]] && check_command wget check_command curl; check_command git; check_command git-lfs [[ -z "$MODEL_ID" || "$MODEL_ID" =~ ^-h ]] && display_help if [[ -z "$LOCAL_DIR" ]]; then LOCAL_DIR="${MODEL_ID#*/}" fi if [[ "$DATASET" == 1 ]]; then MODEL_ID="datasets/$MODEL_ID" fi echo "Downloading to $LOCAL_DIR" if [ -d "$LOCAL_DIR/.git" ]; then printf "${YELLOW}%s exists, Skip Clone.\n${NC}" "$LOCAL_DIR" cd "$LOCAL_DIR" && ensure_ownership && GIT_LFS_SKIP_SMUDGE=1 git pull || { printf "${RED}Git pull failed.${NC}\n"; exit 1; } else REPO_URL="$HF_ENDPOINT/$MODEL_ID" GIT_REFS_URL="${REPO_URL}/info/refs?service=git-upload-pack" echo "Testing GIT_REFS_URL: $GIT_REFS_URL" response=$(curl -s -o /dev/null -w "%{http_code}" "$GIT_REFS_URL") if [ "$response" == "401" ] || [ "$response" == "403" ]; then if [[ -z "$HF_USERNAME" || -z "$HF_TOKEN" ]]; then printf "${RED}HTTP Status Code: $response.\nThe repository requires authentication, but --hf_username and --hf_token is not passed. Please get token from https://huggingface.co/settings/tokens.\nExiting.\n${NC}" exit 1 fi REPO_URL="https://$HF_USERNAME:$HF_TOKEN@${HF_ENDPOINT#https://}/$MODEL_ID" elif [ "$response" != "200" ]; then printf "${RED}Unexpected HTTP Status Code: $response\n${NC}" printf "${YELLOW}Executing debug command: curl -v %s\nOutput:${NC}\n" "$GIT_REFS_URL" curl -v "$GIT_REFS_URL"; printf "\n${RED}Git clone failed.\n${NC}"; exit 1 fi echo "GIT_LFS_SKIP_SMUDGE=1 git clone $REPO_URL $LOCAL_DIR" GIT_LFS_SKIP_SMUDGE=1 git clone $REPO_URL $LOCAL_DIR && cd "$LOCAL_DIR" || { printf "${RED}Git clone failed.\n${NC}"; exit 1; } ensure_ownership while IFS= read -r file; do truncate -s 0 "$file" done <<< $(git lfs ls-files | cut -d ' ' -f 3-) fi printf "\nStart Downloading lfs files, bash script:\ncd $LOCAL_DIR\n" files=$(git lfs ls-files | cut -d ' ' -f 3-) declare -a urls while IFS= read -r file; do url="$HF_ENDPOINT/$MODEL_ID/resolve/main/$file" file_dir=$(dirname "$file") mkdir -p "$file_dir" if [[ "$TOOL" == "wget" ]]; then download_cmd="wget -c \"$url\" -O \"$file\"" [[ -n "$HF_TOKEN" ]] && download_cmd="wget --header=\"Authorization: Bearer ${HF_TOKEN}\" -c \"$url\" -O \"$file\"" else download_cmd="aria2c --console-log-level=error --file-allocation=none -x $THREADS -s $THREADS -k 1M -c \"$url\" -d \"$file_dir\" -o \"$(basename "$file")\"" [[ -n "$HF_TOKEN" ]] && download_cmd="aria2c --header=\"Authorization: Bearer ${HF_TOKEN}\" --console-log-level=error --file-allocation=none -x $THREADS -s $THREADS -k 1M -c \"$url\" -d \"$file_dir\" -o \"$(basename "$file")\"" fi [[ -n "$INCLUDE_PATTERN" && ! "$file" == $INCLUDE_PATTERN ]] && printf "# %s\n" "$download_cmd" && continue [[ -n "$EXCLUDE_PATTERN" && "$file" == $EXCLUDE_PATTERN ]] && printf "# %s\n" "$download_cmd" && continue printf "%s\n" "$download_cmd" urls+=("$url|$file") done <<< "$files" for url_file in "${urls[@]}"; do IFS='|' read -r url file <<< "$url_file" printf "${YELLOW}Start downloading ${file}.\n${NC}" file_dir=$(dirname "$file") if [[ "$TOOL" == "wget" ]]; then [[ -n "$HF_TOKEN" ]] && wget --header="Authorization: Bearer ${HF_TOKEN}" -c "$url" -O "$file" || wget -c "$url" -O "$file" else [[ -n "$HF_TOKEN" ]] && aria2c --header="Authorization: Bearer ${HF_TOKEN}" --console-log-level=error --file-allocation=none -x $THREADS -s $THREADS -k 1M -c "$url" -d "$file_dir" -o "$(basename "$file")" || aria2c --console-log-level=error --file-allocation=none -x $THREADS -s $THREADS -k 1M -c "$url" -d "$file_dir" -o "$(basename "$file")" fi [[ $? -eq 0 ]] && printf "Downloaded %s successfully.\n" "$url" || { printf "${RED}Failed to download %s.\n${NC}" "$url"; exit 1; } done printf "${GREEN}Download completed successfully.\n${NC}" 方法二：模型下载【个人使用记录】
这个代码不能保持目录结构，见下面的改进版
import datetime import os import threading from huggingface_hub import hf_hub_url from huggingface_hub.hf_api import HfApi from huggingface_hub.utils import filter_repo_objects # 执行命令 def execCmd(cmd): print("命令%s开始运行%s" % (cmd, datetime.datetime.now())) os.system(cmd) print("命令%s结束运行%s" % (cmd, datetime.datetime.now())) if __name__ == '__main__': # 需下载的hf库名称 repo_id = "Salesforce/blip2-opt-2.7b" # 本地存储路径 save_path = './blip2-opt-2.7b' # 获取项目信息 _api = HfApi() repo_info = _api.repo_info( repo_id=repo_id, repo_type="model", revision='main', token=None, ) # 获取文件信息 filtered_repo_files = list( filter_repo_objects( items=[f.rfilename for f in repo_info.siblings], allow_patterns=None, ignore_patterns=None, ) ) cmds = [] threads = [] # 需要执行的命令列表 for file in filtered_repo_files: # 获取路径 url = hf_hub_url(repo_id=repo_id, filename=file) # 断点下载指令 cmds.append(f'wget -c {url} -P {save_path}') print(cmds) print("程序开始%s" % datetime.datetime.now()) for cmd in cmds: th = threading.Thread(target=execCmd, args=(cmd,)) th.start() threads.append(th) for th in threads: th.join() print("程序结束%s" % datetime.datetime.now()) 保持目录结构
import datetime import os import threading from pathlib import Path from huggingface_hub import hf_hub_url from huggingface_hub.hf_api import HfApi from huggingface_hub.utils import filter_repo_objects # 执行命令 def execCmd(cmd): print("命令%s开始运行%s" % (cmd, datetime.datetime.now())) os.system(cmd) print("命令%s结束运行%s" % (cmd, datetime.datetime.now())) if __name__ == '__main__': # 需下载的hf库名称 repo_id = "Salesforce/blip2-opt-2.7b" # 本地存储路径 save_path = './blip2-opt-2.7b' # 创建本地保存目录 Path(save_path).mkdir(parents=True, exist_ok=True) # 获取项目信息 _api = HfApi() repo_info = _api.repo_info( repo_id=repo_id, repo_type="model", revision='main', token=None, ) # 获取文件信息 filtered_repo_files = list( filter_repo_objects( items=[f.rfilename for f in repo_info.siblings], allow_patterns=None, ignore_patterns=None, ) ) cmds = [] threads = [] # 需要执行的命令列表 for file in filtered_repo_files: # 获取路径 url = hf_hub_url(repo_id=repo_id, filename=file) # 在本地创建子目录 local_file = os.path.join(save_path, file) local_dir = os.path.dirname(local_file) Path(local_dir).mkdir(parents=True, exist_ok=True) # 断点下载指令 cmds.append(f'wget -c {url} -P {local_dir}') print(cmds) print("程序开始%s" % datetime.datetime.now()) for cmd in cmds: th = threading.Thread(target=execCmd, args=(cmd,)) th.start() threads.append(th) for th in threads: th.join() print("程序结束%s" % datetime.datetime.now()) 数据集下载
import datetime import os import threading from pathlib import Path from huggingface_hub import HfApi from huggingface_hub.utils import filter_repo_objects # 执行命令 def execCmd(cmd): print("命令%s开始运行%s" % (cmd, datetime.datetime.now())) os.system(cmd) print("命令%s结束运行%s" % (cmd, datetime.datetime.now())) if __name__ == '__main__': # 需下载的数据集ID dataset_id = "openai/webtext" # 本地存储路径 save_path = './webtext' # 创建本地保存目录 Path(save_path).mkdir(parents=True, exist_ok=True) # 获取数据集信息 _api = HfApi() dataset_info = _api.dataset_info( dataset_id=dataset_id, revision='main', token=None, ) # 获取文件信息 filtered_dataset_files = list( filter_repo_objects( items=[f.rfilename for f in dataset_info.siblings], allow_patterns=None, ignore_patterns=None, ) ) cmds = [] threads = [] # 需要执行的命令列表 for file in filtered_dataset_files: # 获取路径 url = dataset_info.get_file_url(file) # 在本地创建子目录 local_file = os.path.join(save_path, file) local_dir = os.path.dirname(local_file) Path(local_dir).mkdir(parents=True, exist_ok=True) # 断点下载指令 cmds.append(f'wget -c {url} -P {local_dir}') print(cmds) print("程序开始%s" % datetime.datetime.now()) for cmd in cmds: th = threading.Thread(target=execCmd, args=(cmd,)) th.start() threads.append(th) for th in threads: th.join() print("程序结束%s" % datetime.datetime.now()) 不足之处
不支持需要授权的库。
文件太多可能会开很多线程。
♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠
