SAM-记录一下自己在linux下部署SAM的过程_linux 部署 sam教程

SAM部署

1.检查所需的环境：`python>=3.8``pytorch>=1.7``torchvision>=0.8`

2.安装Anaconda：

    1.先去官网找到对应的下载地址，然后在终端执行

    wget https://repo.anaconda.com/archive/Anaconda3-5.3.1-Linux-x86_64.sh

    2.运行.sh文件

    bash Anaconda3-5.3.1-Linux-x86_64.sh

3.pytorch的下载

    1.使用conda创建一个虚拟环境（python）=3.8

    conda create -n envName_pytorch python=3.8

    2.进入虚拟环境

    source activate envName_pytorch

    3.conda下载pytorch

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

    4.进行检验

    python ，import torch ，torch.rand(3,3)

    但是输入import torch报错

    Traceback (most recent call last):
          File "`<stdin>`", line 1, in `<module>`
        ModuleNotFoundError: No module named 'torch'

    解决方法：重新下载torch

    conda install pytorch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 pytorch-cuda=11.8 -c pytorch -c nvidia

    但是无法下载：Solving environment: failed

    PackagesNotFoundError: The following packages are not available from current channels:

    解决方法：重新去https://anaconda.org 这个网址，在上方的搜索条上搜索你要安装这个包的其他channel

    5.更新依赖库

    进行换源失败，换回conda源： conda config --remove-key channels

4.下载SAM:

    1.从github仓库本地克隆存储库：git clone https : // github.com / facebookresearch / segment-anything.git

    拉取失败：fatal: unable to access 'https://github.com/facebookresearch/segment-anything.git/': GnuTLS recv error (-110): The TLS connection was non-properly terminated.

    解决方案：git config --global --unset https.proxy

    2.安装SAM：

    cd segment-anything

    pip install -e .

    3.安装其他依赖

    pip install opencv-python pycocotools matplotlib onnxruntime onnx

    4.下载model checkpoints：[https://github.com/facebookresearch/segment-anything#model-checkpoints](https://github.com/facebookresearch/segment-anything#model-checkpoints "https://github.com/facebookresearch/segment-anything#model-checkpoints")

5.运行SAM进行模型训练

    1. python scripts/amg.py --checkpoint /home/xtyg/workspace-gm/segment-anything/notebooks/sam_vit_h_4b8939.pth --model-type vit_h --input /home/xtyg/workspace-gm/segment-anything/test/input --output /hpme/xtyg/workspace-gm/segment-anything/test/output

   但报错：raise AssertionError("Torch not compiled with CUDA enabled")
   AssertionError: Torch not compiled with CUDA enabled

   解决方案：卸载torch：conda uninstall libtorch

   2.解决后，再次运行出现新的报错

   Traceback (most recent call last):
   File "scripts/amg.py", line 7, in `<module>`
   import cv2  # type: ignore
   ModuleNotFoundError: No module named 'cv2'

   解决方案：使用pip安装cv2：pip install[opencv](https://so.csdn.net/so/search?q=opencv&spm=1001.2101.3001.7020)-python

   使用python import cv2检验

   3.解决后运行模型出现：

   ModuleNotFoundError: No module named 'segment_anything'

   解决方案，重新pip下载SAM

   4.运行模型分割的代码！！！！！！！！！！！！！！生成掩码

   python scripts/amg.py --checkpoint /home/xtyg/workspace-gm/segment-anything1/sam_vit_h_4b8939.pth --model-type vit_h --input /home/xtyg/workspace-gm/segment-anything1/test/input --output /home/xtyg/workspace-gm/segment-anything1/test/output

   成功实现分割！
   5.扣取图像中的所有物体
        1.导入依赖库和函数
        from segment_anything import sam_model_registry, SamAutomaticMaskGenerator, SamPredictor
        import cv2
        import matplotlib.pyplot as plt
        import numpy as np
        
        def show_anns(anns):
            if len(anns) == 0:
                return
            sorted_anns = sorted(anns, key=(lambda x: x['area']), reverse=True)
            ax = plt.gca()
            ax.set_autoscale_on(False)
            polygons = []
            color = []
            for ann in sorted_anns:
                m = ann['segmentation']
                img = np.ones((m.shape[0], m.shape[1], 3))
                color_mask = np.random.random((1, 3)).tolist()[0]
                for i in range(3):
                    img[:,:,i] = color_mask[i]
                ax.imshow(np.dstack((img, m*0.35)))
        2.读取图片
        image = cv2.imread(r"F:\Dataset\Tomato_Appearance\Tomato_Xishi\images\xs_1.jpg")
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        3.实例化模型：
        sam_checkpoint = "F:\sam_vit_h_4b8939.pth"
        model_type = "default"
        device = "cuda"
        
        sam = sam_model_registry[model_type](checkpoint=sam_checkpoint)
        sam.to(device=device)
        4.分割并展示
            mask_generator = SamAutomaticMaskGenerator(sam)
            masks = mask_generator.generate(image)
            
            plt.figure(figsize=(20,20))
            plt.imshow(image)
            show_anns(masks)
            plt.axis('off')
            plt.show()
6.退出虚拟环境：conda deactivate

7.ONNX导出：

    1.下载ONNX模块：pip install onnx

    2.python scripts/export_onnx_model.py --checkpoint /home/xtyg/workspace-gm/segment-anything1/sam_vit_h_4b8939.pth --model-type vit_h  --output /home/xtyg/workspace-gm/segment-anything1/onnxout.onnx

    使用上面命令进行导出
8.web的演示

   1.下载yarn：

        1.导入软件源的 GPG key 并且添加 Yarn APT 软件源到你的系统

    curl -sS https://dl.yarnpkg.com/debian/pubkey.gpg | sudo apt-key add -

    echo "deb https://dl.yarnpkg.com/debian/ stable main" | sudo tee /etc/apt/sources.list.d/yarn.list

        2.下载yarn：sudo apt install yarn

        启动yarn报错：ERROR: There are no scenarios; must have at least one.”

        解决方案：sudo apt remove cmdtest
        sudo apt remove yarn
        curl -sS https://dl.yarnpkg.com/debian/pubkey.gpg | sudo apt-key add -
        echo "deb https://dl.yarnpkg.com/debian/ stable main" | sudo tee /etc/apt/sources.list.d/yarn.list
        sudo apt update
        sudo apt install yarn

        3.进入demo ：cd demo

        运行：yarn && yarn start

        报错：error npyjs@0.4.0: The engine "node" is incompatible with this module. Expected version "^12.20.0 || >=14.13.1". Got     "10.19.0"
        error Found incompatible module.
        info Visit https://yarnpkg.com/en/docs/cli/install for documentation about this command.
        
        解决方案：yarn config set ignore-engines true
        
        运行yarn && yarn start
        
        报错：Read error (Connection reset by peer) in headers.
    
        解决方案：yarn依赖包没有安装完全，yarn没有真正搭建运行起来
        卸载：sudo npm uninstall yarn -g
        重新安装：sudo npm install -g yarn
        依次输入以下命令
        yarn -v
        yarn config get registry
        yarn config set registry https://registry.npm.taobao.org
        yarn global add
        yarn global list
        yarn init
        yarn list
        安装好后继续报错：info node: --openssl-legacy-provider is not allowed in NODE_OPTIONS
        解决方案：重新安装nodejs

 2. 运行yarn start显示ERROR in unable to locate '/home/xtyg/workspace-gm/segment-anything1/demo/model' glob

        解决方案：上述的原因是模型没有导出，无法在该路径下找到onnx模型
        1.首先导出onnx的模型
            import torch
            import numpy as np
            import cv2
            import matplotlib.pyplot as plt
            from segment_anything import sam_model_registry, SamPredictor
            from segment_anything.utils.onnx import SamOnnxModel

            import onnxruntime
            from onnxruntime.quantization import QuantType
            from onnxruntime.quantization.quantize import quantize_dynamic

            # 我本地存在checkpoints/sam_vit_h_4b8939.pth
            checkpoint =  "sam_vit_h_4b8939.pth"
            model_type = "vit_h"

            sam = sam_model_registry[model_type](checkpoint=checkpoint)

            onnx_model_path = None  # Set to use an already exported model, then skip to the next section.

            import warnings

            onnx_model_path = "sam_onnx_example.onnx"

            onnx_model = SamOnnxModel(sam, return_single_mask=True)

            dynamic_axes = {
                "point_coords": {1: "num_points"},
                "point_labels": {1: "num_points"},
            }

            embed_dim = sam.prompt_encoder.embed_dim
            embed_size = sam.prompt_encoder.image_embedding_size
            mask_input_size = [4 * x for x in embed_size]
            dummy_inputs = {
                "image_embeddings": torch.randn(1, embed_dim, *embed_size, dtype=torch.float),
                "point_coords": torch.randint(low=0, high=1024, size=(1, 5, 2), dtype=torch.float),
                "point_labels": torch.randint(low=0, high=4, size=(1, 5), dtype=torch.float),
                "mask_input": torch.randn(1, 1, *mask_input_size, dtype=torch.float),
                "has_mask_input": torch.tensor([1], dtype=torch.float),
                "orig_im_size": torch.tensor([1500, 2250], dtype=torch.float),
            }
            output_names = ["masks", "iou_predictions", "low_res_masks"]

            with warnings.catch_warnings():
                warnings.filterwarnings("ignore", category=torch.jit.TracerWarning)
                warnings.filterwarnings("ignore", category=UserWarning)
                # 这里导出了sam_onnx_example.onnx
                with open(onnx_model_path, "wb") as f:
                    torch.onnx.export(
                        onnx_model,
                        tuple(dummy_inputs.values()),
                        f,
                        export_params=True,
                        verbose=False,
                        opset_version=17,
                        do_constant_folding=True,
                        input_names=list(dummy_inputs.keys()),
                        output_names=output_names,
                        dynamic_axes=dynamic_axes,
                    )   
            onnx_model_quantized_path = "sam_onnx_quantized_example.onnx"
            quantize_dynamic(
                model_input=onnx_model_path,
                model_output=onnx_model_quantized_path,
                per_channel=False,
                reduce_range=False,
                weight_type=QuantType.QUInt8,
            )
            onnx_model_path = onnx_model_quantized_path        
        2.通过上面的代码会生成sam_onnx_example.onnx，和sam_onnx_quantized_example.onnx模型
        3.运行下面的代码对jpg图像进行编码，输出.npy的编码文件
            checkpoint = "../checkpoints/sam_vit_h_4b8939.pth"
            model_type = "vit_h"
            sam = sam_model_registry[model_type](checkpoint=checkpoint)
            sam.to(device='cuda')
            predictor = SamPredictor(sam)

            image = cv2.imread('../demo/src/assets/data/dogs.jpg')
            predictor.set_image(image)
            image_embedding = predictor.get_image_embedding().cpu().numpy()
            np.save("dogs_embedding.npy", image_embedding)
            type(image_embedding),image_embedding.shape

        4.按照demo/src/App.tsx规定的路径放置文件：
            const IMAGE_PATH = "/assets/data/dogs.jpg";
            const IMAGE_EMBEDDING = "/assets/data/dogs_embedding.npy";
            const MODEL_DIR = "/model/sam_onnx_quantized_example.onnx";
   3.接下来进入 demo ：cd demo
        运行yarn && yarn starts
    在电脑的浏览器上输入http://192.168.1.183:8080/可以成功连上服务器