├── requirements.txt ├── requirements-test.txt ├── cache.db ├── requirements-base.txt ├── start.sh ├── .dockerignore ├── test_api.py ├── docker-compose.yml ├── Dockerfile ├── _wsgi.py ├── README.md └── model.py /requirements.txt: -------------------------------------------------------------------------------- 1 | 2 | 3 | -------------------------------------------------------------------------------- /requirements-test.txt: -------------------------------------------------------------------------------- 1 | pytest 2 | pytest-cov -------------------------------------------------------------------------------- /cache.db: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/YiuChoi/label-studio-sam2/master/cache.db -------------------------------------------------------------------------------- /requirements-base.txt: -------------------------------------------------------------------------------- 1 | gunicorn==22.0.0 2 | label-studio-ml @ git+https://github.com/HumanSignal/label-studio-ml-backend.git -------------------------------------------------------------------------------- /start.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | # Execute the gunicorn command 4 | exec gunicorn --bind :${PORT:-9090} --workers ${WORKERS:-1} --threads ${THREADS:-4} --timeout 0 --pythonpath '/app' _wsgi:app 5 | -------------------------------------------------------------------------------- /.dockerignore: -------------------------------------------------------------------------------- 1 | # Exclude everything 2 | ** 3 | 4 | # Include Dockerfile and docker-compose for reference (optional, decide based on your use case) 5 | !Dockerfile 6 | !docker-compose.yml 7 | 8 | # Include Python application files 9 | !*.py 10 | 11 | # Include requirements files 12 | !requirements*.txt 13 | 14 | # Include script 15 | !*.sh 16 | 17 | # Exclude specific requirements if necessary 18 | # requirements-test.txt (Uncomment if you decide to exclude this) 19 | -------------------------------------------------------------------------------- /test_api.py: -------------------------------------------------------------------------------- 1 | """ 2 | This file contains tests for the API of your model. You can run these tests by installing test requirements: 3 | 4 | ```bash 5 | pip install -r requirements-test.txt 6 | ``` 7 | Then execute `pytest` in the directory of this file. 8 | 9 | - Change `NewModel` to the name of the class in your model.py file. 10 | - Change the `request` and `expected_response` variables to match the input and output of your model. 11 | """ 12 | 13 | import pytest 14 | import json 15 | from model import NewModel 16 | 17 | 18 | @pytest.fixture 19 | def client(): 20 | from _wsgi import init_app 21 | app = init_app(model_class=NewModel) 22 | app.config['TESTING'] = True 23 | with app.test_client() as client: 24 | yield client 25 | 26 | 27 | def test_predict(client): 28 | request = { 29 | 'tasks': [{ 30 | 'data': { 31 | # Your input test data here 32 | } 33 | }], 34 | # Your labeling configuration here 35 | 'label_config': '' 36 | } 37 | 38 | expected_response = { 39 | 'results': [{ 40 | # Your expected result here 41 | }] 42 | } 43 | 44 | response = client.post('/predict', data=json.dumps(request), content_type='application/json') 45 | assert response.status_code == 200 46 | response = json.loads(response.data) 47 | assert response == expected_response 48 | -------------------------------------------------------------------------------- /docker-compose.yml: -------------------------------------------------------------------------------- 1 | version: "3.8" 2 | 3 | services: 4 | ml-backend: 5 | container_name: ml-backend 6 | image: humansignal/ml-backend:v0 7 | build: 8 | context: . 9 | args: 10 | TEST_ENV: ${TEST_ENV} 11 | 12 | # deploy: 13 | # resources: 14 | # reservations: 15 | # devices: 16 | # - driver: nvidia 17 | # count: 1 18 | # capabilities: [ gpu ] 19 | 20 | 21 | environment: 22 | # specify these parameters if you want to use basic auth for the model server 23 | - BASIC_AUTH_USER= 24 | - BASIC_AUTH_PASS= 25 | # set the log level for the model server 26 | - LOG_LEVEL=DEBUG 27 | # any other parameters that you want to pass to the model server 28 | - ANY=PARAMETER 29 | # specify the number of workers and threads for the model server 30 | - WORKERS=1 31 | - THREADS=8 32 | # specify the model directory (likely you don't need to change this) 33 | - MODEL_DIR=/data/models 34 | # specify device 35 | - DEVICE=cuda # or 'cpu' (coming soon) 36 | # SAM2 model config 37 | - MODEL_CONFIG=configs/sam2.1/sam2.1_hiera_l.yaml 38 | # SAM2 checkpoint 39 | - MODEL_CHECKPOINT=sam2.1_hiera_large.pt 40 | 41 | # Specify the Label Studio URL and API key to access 42 | # uploaded, local storage and cloud storage files. 43 | # Do not use 'localhost' as it does not work within Docker containers. 44 | # Use prefix 'http://' or 'https://' for the URL always. 45 | # Determine the actual IP using 'ifconfig' (Linux/Mac) or 'ipconfig' (Windows). 46 | - LABEL_STUDIO_URL= 47 | - LABEL_STUDIO_API_KEY= 48 | ports: 49 | - "9090:9090" 50 | volumes: 51 | - "./data/server:/data" 52 | -------------------------------------------------------------------------------- /Dockerfile: -------------------------------------------------------------------------------- 1 | FROM pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime 2 | ARG DEBIAN_FRONTEND=noninteractive 3 | ARG TEST_ENV 4 | 5 | WORKDIR /app 6 | 7 | RUN mamba update conda -y 8 | 9 | RUN --mount=type=cache,target="/var/cache/apt",sharing=locked \ 10 | --mount=type=cache,target="/var/lib/apt/lists",sharing=locked \ 11 | apt-get -y update \ 12 | && apt-get install -y git \ 13 | && apt-get install -y wget \ 14 | && apt-get install -y g++ freeglut3-dev build-essential libx11-dev \ 15 | libxmu-dev libxi-dev libglu1-mesa libglu1-mesa-dev libfreeimage-dev \ 16 | && apt-get -y install ffmpeg libsm6 libxext6 libffi-dev python3-dev python3-pip gcc 17 | 18 | ENV PYTHONUNBUFFERED=1 \ 19 | PYTHONDONTWRITEBYTECODE=1 \ 20 | PIP_CACHE_DIR=/.cache \ 21 | PORT=9090 \ 22 | WORKERS=2 \ 23 | THREADS=4 \ 24 | CUDA_HOME=/usr/local/cuda 25 | 26 | RUN mamba install nvidia/label/cuda-12.4.0::cuda -y 27 | 28 | ENV CUDA_HOME=/opt/conda \ 29 | TORCH_CUDA_ARCH_LIST="6.0;6.1;7.0;7.5;8.0;8.6+PTX;8.9;9.0" 30 | 31 | # install base requirements 32 | COPY requirements-base.txt . 33 | RUN --mount=type=cache,target=${PIP_CACHE_DIR},sharing=locked \ 34 | pip install -r requirements-base.txt 35 | 36 | COPY requirements.txt . 37 | RUN --mount=type=cache,target=${PIP_CACHE_DIR},sharing=locked \ 38 | pip3 install -r requirements.txt 39 | 40 | # install segment-anything-2 41 | RUN cd / && git clone --depth 1 --branch main --single-branch https://github.com/facebookresearch/sam2.git 42 | WORKDIR /sam2 43 | RUN --mount=type=cache,target=${PIP_CACHE_DIR},sharing=locked \ 44 | pip3 install -e . 45 | RUN cd checkpoints && ./download_ckpts.sh 46 | 47 | WORKDIR /app 48 | 49 | # install test requirements if needed 50 | COPY requirements-test.txt . 51 | # build only when TEST_ENV="true" 52 | RUN --mount=type=cache,target=${PIP_CACHE_DIR},sharing=locked \ 53 | if [ "$TEST_ENV" = "true" ]; then \ 54 | pip3 install -r requirements-test.txt; \ 55 | fi 56 | 57 | COPY . ./ 58 | 59 | WORKDIR ../sam2 60 | 61 | CMD ["../app/start.sh"] -------------------------------------------------------------------------------- /_wsgi.py: -------------------------------------------------------------------------------- 1 | import os 2 | import argparse 3 | import json 4 | import logging 5 | import logging.config 6 | 7 | logging.config.dictConfig({ 8 | "version": 1, 9 | "disable_existing_loggers": False, 10 | "formatters": { 11 | "standard": { 12 | "format": "[%(asctime)s] [%(levelname)s] [%(name)s::%(funcName)s::%(lineno)d] %(message)s" 13 | } 14 | }, 15 | "handlers": { 16 | "console": { 17 | "class": "logging.StreamHandler", 18 | "level": os.getenv('LOG_LEVEL'), 19 | "stream": "ext://sys.stdout", 20 | "formatter": "standard" 21 | } 22 | }, 23 | "root": { 24 | "level": os.getenv('LOG_LEVEL'), 25 | "handlers": [ 26 | "console" 27 | ], 28 | "propagate": True 29 | } 30 | }) 31 | 32 | from label_studio_ml.api import init_app 33 | from model import NewModel 34 | 35 | 36 | _DEFAULT_CONFIG_PATH = os.path.join(os.path.dirname(__file__), 'config.json') 37 | 38 | 39 | def get_kwargs_from_config(config_path=_DEFAULT_CONFIG_PATH): 40 | if not os.path.exists(config_path): 41 | return dict() 42 | with open(config_path) as f: 43 | config = json.load(f) 44 | assert isinstance(config, dict) 45 | return config 46 | 47 | 48 | if __name__ == "__main__": 49 | parser = argparse.ArgumentParser(description='Label studio') 50 | parser.add_argument( 51 | '-p', '--port', dest='port', type=int, default=9090, 52 | help='Server port') 53 | parser.add_argument( 54 | '--host', dest='host', type=str, default='0.0.0.0', 55 | help='Server host') 56 | parser.add_argument( 57 | '--kwargs', '--with', dest='kwargs', metavar='KEY=VAL', nargs='+', type=lambda kv: kv.split('='), 58 | help='Additional LabelStudioMLBase model initialization kwargs') 59 | parser.add_argument( 60 | '-d', '--debug', dest='debug', action='store_true', 61 | help='Switch debug mode') 62 | parser.add_argument( 63 | '--log-level', dest='log_level', choices=['DEBUG', 'INFO', 'WARNING', 'ERROR'], default=None, 64 | help='Logging level') 65 | parser.add_argument( 66 | '--model-dir', dest='model_dir', default=os.path.dirname(__file__), 67 | help='Directory where models are stored (relative to the project directory)') 68 | parser.add_argument( 69 | '--check', dest='check', action='store_true', 70 | help='Validate model instance before launching server') 71 | parser.add_argument('--basic-auth-user', 72 | default=os.environ.get('ML_SERVER_BASIC_AUTH_USER', None), 73 | help='Basic auth user') 74 | 75 | parser.add_argument('--basic-auth-pass', 76 | default=os.environ.get('ML_SERVER_BASIC_AUTH_PASS', None), 77 | help='Basic auth pass') 78 | 79 | args = parser.parse_args() 80 | 81 | # setup logging level 82 | if args.log_level: 83 | logging.root.setLevel(args.log_level) 84 | 85 | def isfloat(value): 86 | try: 87 | float(value) 88 | return True 89 | except ValueError: 90 | return False 91 | 92 | def parse_kwargs(): 93 | param = dict() 94 | for k, v in args.kwargs: 95 | if v.isdigit(): 96 | param[k] = int(v) 97 | elif v == 'True' or v == 'true': 98 | param[k] = True 99 | elif v == 'False' or v == 'false': 100 | param[k] = False 101 | elif isfloat(v): 102 | param[k] = float(v) 103 | else: 104 | param[k] = v 105 | return param 106 | 107 | kwargs = get_kwargs_from_config() 108 | 109 | if args.kwargs: 110 | kwargs.update(parse_kwargs()) 111 | 112 | if args.check: 113 | print('Check "' + NewModel.__name__ + '" instance creation..') 114 | model = NewModel(**kwargs) 115 | 116 | app = init_app(model_class=NewModel, basic_auth_user=args.basic_auth_user, basic_auth_pass=args.basic_auth_pass) 117 | 118 | app.run(host=args.host, port=args.port, debug=args.debug) 119 | 120 | else: 121 | # for uWSGI use 122 | app = init_app(model_class=NewModel) 123 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | 18 | 19 | # 使用SAM2自动化标注图片，用于目标检测 20 | 21 | Segment Anything 2, or SAM 2, is a model released by Meta in July 2024. An update to the original Segment Anything Model, 22 | SAM 2 provides even better object segmentation for both images and video. In this guide, we'll show you how to use 23 | SAM 2 for better image labeling with label studio. 24 | 25 | Click on the image below to watch our ML Evangelist Micaela Kaplan explain how to link SAM 2 to your Label Studio Project. 26 | You'll need to follow the instructions below to stand up an instance of SAM2 before you can link your model! 27 | 28 | [![Connecting SAM2 Model to Label Studio for Image Annotation ](https://img.youtube.com/vi/FTg8P8z4RgY/0.jpg)](https://www.youtube.com/watch?v=FTg8P8z4RgY) 29 | 30 | ## Before you begin 31 | 32 | Before you begin, you must install the [Label Studio ML backend](https://github.com/HumanSignal/label-studio-ml-backend?tab=readme-ov-file#quickstart). 33 | 34 | This tutorial uses the [`segment_anything_2_image` example](https://github.com/HumanSignal/label-studio-ml-backend/tree/master/label_studio_ml/examples/segment_anything_2_image). 35 | 36 | Note that as of 8/1/2024, SAM2 only runs on GPU. 37 | 38 | ## Labeling configuration 39 | 40 | The current implementation of the Label Studio SAM2 ML backend works using Interactive mode. The user-guided inputs are: 41 | - `KeypointLabels` 42 | - `RectangleLabels` 43 | 44 | And then SAM2 outputs `RectangleLabels` as a result. 45 | 46 | This means all three control tags should be represented in your labeling configuration: 47 | 48 | ```xml 49 | 50 | 51 | 52 | 53 | 54 | 59 | 60 | 64 | 65 | ``` 66 | 67 | ## Running from source 68 | 69 | 1. To run the ML backend without Docker, you have to clone the repository and install all dependencies using pip: 70 | 71 | ```bash 72 | git clone https://github.com/HumanSignal/label-studio-ml-backend.git 73 | cd label-studio-ml-backend 74 | pip install -e . 75 | cd label_studio_ml/examples/segment_anything_2_image 76 | pip install -r requirements.txt 77 | ``` 78 | 79 | 2. Download [`segment-anything-2` repo](https://github.com/facebookresearch/sam2) into the root directory. Install SegmentAnything model and download checkpoints using [the official Meta documentation](https://github.com/facebookresearch/sam2?tab=readme-ov-file#installation) 80 | You should now have the following folder structure: 81 | 82 | 83 | | root directory 84 | | label-studio-ml-backend 85 | | label-studio-ml 86 | | examples 87 | | segment_anything_2_image 88 | | sam2 89 | | sam2 90 | | checkpoints 91 | 92 | 93 | 3. Then you can start the ML backend on the default port `9090`: 94 | 95 | ```bash 96 | cd ~/sam2 97 | label-studio-ml start ../label-studio-ml-backend/label_studio_ml/examples/segment_anything_2_image 98 | ``` 99 | 100 | Due to breaking changes from Meta [HERE](https://github.com/facebookresearch/sam2/blob/c2ec8e14a185632b0a5d8b161928ceb50197eddc/sam2/build_sam.py#L20), it is CRUCIAL that you run this command from the sam2 directory at your root directory. 101 | 102 | 4. Connect running ML backend server to Label Studio: go to your project `Settings -> Machine Learning -> Add Model` and specify `http://localhost:9090` as a URL. Read more in the official [Label Studio documentation](https://labelstud.io/guide/ml#Connect-the-model-to-Label-Studio). 103 | 104 | ## Running with Docker 105 | 106 | 1. Start Machine Learning backend on `http://localhost:9090` with prebuilt image: 107 | 108 | ```bash 109 | docker-compose up 110 | ``` 111 | 112 | 2. Validate that backend is running 113 | 114 | ```bash 115 | $ curl http://localhost:9090/ 116 | {"status":"UP"} 117 | ``` 118 | 119 | 3. Connect to the backend from Label Studio running on the same host: go to your project `Settings -> Machine Learning -> Add Model` and specify `http://localhost:9090` as a URL. 120 | 121 | 122 | ## Configuration 123 | Parameters can be set in `docker-compose.yml` before running the container. 124 | 125 | 126 | The following common parameters are available: 127 | - `DEVICE` - specify the device for the model server (currently only `cuda` is supported, `cpu` is coming soon) 128 | - `MODEL_CONFIG` - SAM2 model configuration file (`sam2_hiera_l.yaml` by default) 129 | - `MODEL_CHECKPOINT` - SAM2 model checkpoint file (`sam2_hiera_large.pt` by default) 130 | - `BASIC_AUTH_USER` - specify the basic auth user for the model server 131 | - `BASIC_AUTH_PASS` - specify the basic auth password for the model server 132 | - `LOG_LEVEL` - set the log level for the model server 133 | - `WORKERS` - specify the number of workers for the model server 134 | - `THREADS` - specify the number of threads for the model server 135 | 136 | ## Customization 137 | 138 | The ML backend can be customized by adding your own models and logic inside the `./segment_anything_2` directory. 139 | -------------------------------------------------------------------------------- /model.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import numpy as np 3 | import os 4 | import sys 5 | import pathlib 6 | import cv2 7 | from typing import List, Dict, Optional 8 | from uuid import uuid4 9 | from label_studio_ml.model import LabelStudioMLBase 10 | from label_studio_ml.response import ModelResponse 11 | from label_studio_sdk._extensions.label_studio_tools.core.utils.io import get_local_path 12 | from PIL import Image 13 | 14 | ROOT_DIR = os.getcwd() 15 | sys.path.insert(0, ROOT_DIR) 16 | from sam2.build_sam import build_sam2 17 | from sam2.sam2_image_predictor import SAM2ImagePredictor 18 | 19 | # 环境变量配置 20 | DEVICE = os.getenv('DEVICE', 'cuda') 21 | MODEL_CONFIG = os.getenv('MODEL_CONFIG', 'configs/sam2.1/sam2.1_hiera_l.yaml') 22 | MODEL_CHECKPOINT = os.getenv('MODEL_CHECKPOINT', 'sam2.1_hiera_large.pt') 23 | 24 | # CUDA 设置 25 | if DEVICE == 'cuda': 26 | torch.autocast(device_type="cuda", dtype=torch.bfloat16).__enter__() 27 | if torch.cuda.get_device_properties(0).major >= 8: 28 | torch.backends.cuda.matmul.allow_tf32 = True 29 | torch.backends.cudnn.allow_tf32 = True 30 | 31 | # 构建模型检查点路径 32 | sam2_checkpoint = str(os.path.join(ROOT_DIR, "checkpoints", MODEL_CHECKPOINT)) 33 | 34 | # 全局初始化 SAM2 模型 35 | sam2_model = build_sam2(MODEL_CONFIG, sam2_checkpoint, device=DEVICE) 36 | 37 | class NewModel(LabelStudioMLBase): 38 | """Custom ML Backend model for generating rectanglelabels""" 39 | 40 | def __init__(self, **kwargs): 41 | super(NewModel, self).__init__(**kwargs) 42 | # 定义 value、from_name 和 to_name 43 | self.value = kwargs.get('image_key', 'image') # 图像 URL 的键，默认为 'image' 44 | self.from_name = kwargs.get('from_name', 'label') # 标签名称 45 | self.to_name = kwargs.get('to_name', 'image') # 目标对象名称 46 | # 初始化 SAM2 预测器 47 | self.predictor = SAM2ImagePredictor(sam2_model) 48 | 49 | def set_image(self, image_url, task_id): 50 | """加载并设置图像到 SAM2 预测器""" 51 | image_path = get_local_path(image_url, task_id=task_id) 52 | image = Image.open(image_path) 53 | image = np.array(image.convert("RGB")) 54 | self.predictor.set_image(image) 55 | 56 | def _sam_predict(self, img_url, point_coords=None, point_labels=None, input_box=None, task=None): 57 | """使用 SAM2 进行预测""" 58 | self.set_image(img_url, task.get('id')) 59 | point_coords = np.array(point_coords, dtype=np.float32) if point_coords else None 60 | point_labels = np.array(point_labels, dtype=np.float32) if point_labels else None 61 | input_box = np.array(input_box, dtype=np.float32) if input_box else None 62 | 63 | masks, scores, _ = self.predictor.predict( 64 | point_coords=point_coords, 65 | point_labels=point_labels, 66 | box=input_box, 67 | multimask_output=False # 仅返回一个掩码 68 | ) 69 | return { 70 | 'masks': masks, 71 | 'scores': scores 72 | } 73 | 74 | def predict(self, tasks: List[Dict], context: Optional[Dict] = None, **kwargs) -> ModelResponse: 75 | """基于用户交互预测 rectanglelabels""" 76 | results = [] 77 | 78 | # 获取 from_name, to_name 和 value 79 | from_name, to_name, value = self.get_first_tag_occurence('RectangleLabels', 'Image') 80 | 81 | # 如果没有上下文（用户未交互），返回空预测 82 | if not context or not context.get('result'): 83 | return ModelResponse(predictions=[]) 84 | 85 | # 获取图像尺寸 86 | image_width = context['result'][0]['original_width'] 87 | image_height = context['result'][0]['original_height'] 88 | 89 | # 收集用户交互信息 90 | point_coords = [] 91 | point_labels = [] 92 | input_box = None 93 | selected_label = None 94 | for ctx in context['result']: 95 | x = ctx['value']['x'] * image_width / 100 96 | y = ctx['value']['y'] * image_height / 100 97 | ctx_type = ctx['type'] 98 | selected_label = ctx['value'][ctx_type][0] if ctx_type in ctx['value'] else 'object' 99 | if ctx_type == 'keypointlabels': 100 | point_labels.append(int(ctx.get('is_positive', 0))) 101 | point_coords.append([int(x), int(y)]) 102 | elif ctx_type == 'rectanglelabels': 103 | box_width = ctx['value']['width'] * image_width / 100 104 | box_height = ctx['value']['height'] * image_height / 100 105 | input_box = [int(x), int(y), int(box_width + x), int(box_height + y)] 106 | 107 | print(f'Point coords: {point_coords}, Point labels: {point_labels}, Input box: {input_box}') 108 | 109 | # 获取图像 URL 110 | img_url = tasks[0]['data'][value] 111 | 112 | # 使用 SAM2 预测 113 | predictor_results = self._sam_predict( 114 | img_url=img_url, 115 | point_coords=point_coords or None, 116 | point_labels=point_labels or None, 117 | input_box=input_box, 118 | task=tasks[0] 119 | ) 120 | 121 | # 处理预测结果，生成 rectanglelabels 122 | masks = predictor_results['masks'] 123 | scores = predictor_results['scores'] 124 | 125 | for mask, score in zip(masks, scores): 126 | # 将掩码转换为二值图像 127 | mask = mask.astype(np.uint8) * 255 128 | 129 | # 找到掩码的轮廓 130 | contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) 131 | if not contours: 132 | continue 133 | 134 | # 获取最大轮廓 135 | contour = max(contours, key=cv2.contourArea) 136 | 137 | # 计算边界框 138 | x, y, w, h = cv2.boundingRect(contour) 139 | 140 | # 转换为 Label Studio 的百分比格式 141 | x_percent = (x / image_width) * 100 142 | y_percent = (y / image_height) * 100 143 | width_percent = (w / image_width) * 100 144 | height_percent = (h / image_height) * 100 145 | 146 | # 构造 rectanglelabels 结果 147 | result = { 148 | 'id': str(uuid4())[:4], 149 | 'from_name': from_name, 150 | 'to_name': to_name, 151 | 'type': 'rectanglelabels', 152 | 'value': { 153 | 'x': x_percent, 154 | 'y': y_percent, 155 | 'width': width_percent, 156 | 'height': height_percent, 157 | 'rectanglelabels': [selected_label] 158 | }, 159 | 'score': float(score), 160 | 'original_width': image_width, 161 | 'original_height': image_height, 162 | 'image_rotation': 0 163 | } 164 | results.append(result) 165 | 166 | return ModelResponse(predictions=[{ 167 | 'result': results, 168 | 'model_version': '1.0', # 硬编码 model_version 169 | 'score': sum([r['score'] for r in results]) / max(len(results), 1) 170 | }]) --------------------------------------------------------------------------------