如何在Linux上让Face-recognition脚本调用独立NVIDIA显卡

我正在使用face_recognition库，调用CNN模型进行人脸检测：

locations = face_recognition.face_locations(frame, model='cnn')

想让脚本用独立显卡而非集成显卡运行，但尝试optirun python main.py和primusrun python main.py都没用，执行nvidia-smi查看显卡使用情况时，只有Xorg在占用显存：

nvidia-smi
...

|=======================================================================================|
|    0   N/A  N/A      7626      G   /usr/lib/Xorg                                 5MiB |
+---------------------------------------------------------------------------------------+

显然脚本没用到独立显卡。我的Linux基础薄弱，但想深入了解系统，麻烦帮忙解决。

我的脚本如下：

def detected():

    # data = pickle.loads(open('user_1_encodings.pickle', 'rb').read())

    cams = cv2.VideoCapture(0)
    not_identify = []

    while True:
        ret, frame = cams.read()

        locations = face_recognition.face_locations(frame, model='cnn')
        encodings = face_recognition.face_encodings(frame, locations)

        for face_encoding, face_location in zip(encodings, locations):
            # user_encodings = user_data['encodings']
            # result = face_recognition.compare_faces(user_encodings, face_encoding)
            result = False
            green = [0, 255, 0]
            blue = [255, 0, 0]
            red = [0, 0, 255]

            if any(result):
                print(result)
                match = user_data['name']
                left_top = (face_location[3], face_location[0])
                right_bottom = (face_location[1], face_location[2])
                cv2.rectangle(frame, left_top, right_bottom, green, 2)

                text_x = face_location[3]
                text_y = face_location[0] - 10  
                font_scale = 0.5
                font_thickness = 1
                font_face = cv2.FONT_HERSHEY_COMPLEX_SMALL
                text_size, _ = cv2.getTextSize(match, font_face, font_scale, font_thickness)
                text_width, text_height = text_size


                rect_left_top = (text_x-1, text_y+5)
                rect_right_bottom = (text_x + text_width+5, text_y - text_height-5)
                cv2.rectangle(frame, rect_left_top, rect_right_bottom, (255, 255, 255), -1)


                cv2.putText(frame, match, (text_x, text_y), font_face, font_scale, (255, 0, 0), font_thickness)

            else:
                print('NOT OK')
                print(face_location)
                match = 'Unknown'

                top, right, bottom, left = face_location
                face = frame[top:bottom, left:right]
                pil_image = Image.fromarray(face)

                print(pil_image)
                not_identify.append(pil_image)
                print(not_identify)
                if len(not_identify) > 3:
                    #
                    if result is False:
                        not_identify = []
                    else:
                        data = result
                        not_identify = []


                left_top = face_location[3], face_location[0]
                right_bottom = (face_location[1], face_location[2])

                cv2.rectangle(frame, left_top, right_bottom, red, 2)


            text_x = face_location[3]
            text_y = face_location[0] - 10  

            font_scale = 0.5
            font_thickness = 1
            font_face = cv2.FONT_HERSHEY_COMPLEX_SMALL
            text_size, _ = cv2.getTextSize(match, font_face, font_scale, font_thickness)
            text_width, text_height = text_size

            rect_left_top = (text_x-1, text_y+5)
            rect_right_bottom = (text_x + text_width+5, text_y - text_height-5)
            cv2.rectangle(frame, rect_left_top, rect_right_bottom, (255, 255, 255), -1)

            cv2.putText(frame, match, (text_x, text_y), font_face, font_scale, (255, 0, 0), font_thickness)

    cv2.imshow('scan', frame)
    k = cv2.waitKey(20)
    if k == ord('q'):
        break


if __name__ == '__main__':
    detected()

基准测试程序用optirun运行时能正常调用独立显卡，说明驱动安装没问题：

+---------------------------------------------------------------------------------------+
| Processes:                                                                            |
|  GPU   GI   CI        PID   Type   Process name                            GPU Memory |
|        ID   ID                                                             Usage      |
|=======================================================================================|
|    0   N/A  N/A      6351      G   /usr/lib/Xorg                                24MiB |
|    0   N/A  N/A      6355      G   glxspheres64                                  2MiB |
+---------------------------------------------------------------------------------------+

1. 检查dlib是否支持CUDA #

face_recognition底层依赖dlib，CNN人脸检测需要dlib编译时启用CUDA支持。先运行以下代码确认：

import dlib
print(dlib.DLIB_USE_CUDA)

如果输出False，说明当前dlib没有CUDA支持，必须重新编译安装。

2. 重新编译安装带CUDA支持的dlib和face_recognition #

• 先卸载现有版本：

pip uninstall -y dlib face_recognition

• 安装编译依赖（Ubuntu/Debian系）：

sudo apt install build-essential cmake libopenblas-dev liblapack-dev libx11-dev libgtk-3-dev python3-dev

• 编译安装dlib（确保CUDA已正确安装，执行nvcc --version能输出版本信息）：

pip install dlib --no-binary :all:

编译过程会自动检测CUDA环境，只要CUDA正常，就会启用CUDA支持。

• 重新安装face_recognition：

pip install face_recognition

3. 优化optirun/primusrun的运行方式 #

• 先进入optirun的shell环境再运行脚本，确保环境变量正确传递：

optirun bash
python main.py

• 或者指定Python解释器的完整路径，避免调用到错误的版本（比如虚拟环境内的Python）：

optirun /usr/bin/python3 main.py
# 虚拟环境示例
optirun ~/venvs/face_detect/bin/python main.py

4. 验证GPU使用情况 #

运行脚本后，再次执行nvidia-smi，如果看到python进程占用GPU显存，就说明成功调用了独立显卡。也可以用nvidia-smi dmon实时监控GPU使用率变化。

内容的提问来源于stack exchange，提问作者Roman