快速方便地下载huggingface的模型库和数据集
旋转的油纸伞 2024-08-30 10:33:01 阅读 91
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 工具特点Usage
方法二：模型下载【个人使用记录】保持目录结构数据集下载不足之处
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 <code>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"code>
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'code>
GREEN='\033[0;32m'code>
YELLOW='\033[1;33m'code>
NC='\033[0m' # No Colorcode>
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=$1
shift
# Default values
TOOL="aria2c"code>
THREADS=4
HF_ENDPOINT=${ HF_ENDPOINT:-"https://hf-mirror.com"}
while [[ $# -gt 0 ]]; do
case $1 in
--include) INCLUDE_PATTERN="$2"; shift 2 ;;code>
--exclude) EXCLUDE_PATTERN="$2"; shift 2 ;;code>
--hf_username) HF_USERNAME="$2"; shift 2 ;;code>
--hf_token) HF_TOKEN="$2"; shift 2 ;;code>
--tool) TOOL="$2"; shift 2 ;;code>
-x) THREADS="$2"; shift 2 ;;code>
--dataset) DATASET=1; shift ;;
--local-dir) LOCAL_DIR="$2"; shift 2 ;;code>
*) 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#*/}"code>
fi
if [[ "$DATASET" == 1 ]]; then
MODEL_ID="datasets/$MODEL_ID"code>
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"code>
GIT_REFS_URL="${REPO_URL}/info/refs?service=git-upload-pack"code>
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"code>
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/$file"code>
file_dir=$(dirname "$file")
mkdir -p "$file_dir"
if [[ "$TOOL" == "wget" ]]; then
download_cmd="wget -c \"$url\" -O \"$file\""code>
[[ -n "$HF_TOKEN" ]] && download_cmd="wget --header=\"Authorization: Bearer ${HF_TOKEN}\" -c \"$url\" -O \"$file\""code>
else
download_cmd="aria2c --console-log-level=error --file-allocation=none -x $THREADS -s $THREADS -k 1M -c \"$url\" -d \"$file_dir\" -o \"$(basename "$file")\""code>
[[ -n "$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 "$file")\""code>
fi
[[ -n "$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"code>
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 "$file"code>
else
[[ -n "$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")"code>
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",code>
revision='main',code>
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",code>
revision='main',code>
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',code>
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。
创作不易，观众老爷们请留步… 动起可爱的小手，点个赞再走呗 (๑◕ܫ￩๑)
欢迎大家关注笔者，你的关注是我持续更博的最大动力
原创文章，转载告知，盗版必究
♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠ ⊕ ♠
声明
本文内容仅代表作者观点，或转载于其他网站，本站不以此文作为商业用途
如有涉及侵权，请联系本站进行删除
转载本站原创文章，请注明来源及作者。