Created by: wang-kangkang

我的需求是在reader中使用inference_model进行预测，reader放在multiprocess_reader中。使用了两种实现方式都报错：Error: cudaSetDevice failed in paddle::platform::SetDeviced, error code : 3, Please see detail in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.htm l#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430e0038: initialization error at (/paddle/paddle/fluid/platform/gpu_info.cc:240) 第一种是：在reader中定义CUDAPlace和exe；第二种是在reader中使用python API预测。其中第二种方法在pool构造的多进程中没问题，在本场景的multiprocess_reader中有问题。 最小可复现代码：paddle1.7 post97

import os
os.environ["CUDA_VISIBLE_DEVICES"]="2"
import paddle
import paddle.fluid as fluid
from paddle.fluid.core import PaddleTensor
from paddle.fluid.core import AnalysisConfig
from paddle.fluid.core import create_paddle_predictor
def reader():
    landmark_config = AnalysisConfig('lmk_model/model','lmk_model/params')
    landmark_config.switch_use_feed_fetch_ops(False)
    #landmark_config.disable_gpu() 
    landmark_config.enable_use_gpu(100, 0)  
    landmark_predictor = create_paddle_predictor(landmark_config) 
    for i in range(10):
        yield i
if __name__=="__main__":
    place = fluid.CUDAPlace(0)
    exe = fluid.Executor(place)
    train_reader = paddle.reader.multiprocess_reader([reader, reader])
    for data in train_reader():
        print(data)