多进程调用GPU出现UnknownError (see above for traceback): Failed to get convolution algorithm._error reason: assert expression: <unknown> message-优快云博客

本文链接：https://blog.youkuaiyun.com/qq_25620121/article/details/118408904

本文探讨了如何在使用TensorFlow和Keras进行目标检测时，为多个摄像头进行多进程并行调用GPU，避免了内存初始化错误。通过将类实例化移至函数内部，针对每个进程独立创建YOLO对象，解决了UnknownError问题，确保了高效资源利用。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

在利用tensorflow和keras进行目标检测时，针对多个摄像头进行多进程调用GPU，在执行以下代码

import tensorflow as tf

os.environ["CUDA_VISIBLE_DEVICES"] = '0'
# '''使用tf.GPUOptions限制GPU资源的使用方法一'''
config = tf.ConfigProto()
# config.gpu_options.allow_growth = True  # 动态申请显存
config.gpu_options.per_process_gpu_memory_fraction = 0.2  # 占用20%显存

出现UnknownError (see above for traceback): Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.的情况

解决办法是将类的实例化过程移到函数内部，针对每个摄像头分别实例化。举例

# yolo = YOLO()  
# 类的实例化如果在这里执行，会出现错误
# UnknownError (see above for traceback): Failed to get convolution algorithm. 
# 将此行代码移到函数process内部，针对每个进程分别进行类的实例化即可。

def process(position):
    yolo = YOLO()  # 类的实例化
    image = yolo.detect_image(image)
    return image

if __name__ == '__main__':
    mp.freeze_support()
    processAll = []
    processAll.append(mp.Process(target=process, args=('1')))
    processAll.append(mp.Process(target=process, args=('2')))
    processAll.append(mp.Process(target=process, args=('3')))
    processAll.append(mp.Process(target=process, args=('4')))

    [setattr(p, "daemon", True) for p in processAll]
    [p.start() for p in processAll]
    [p.join() for p in processAll]