windows10下tensorflow-gpu运行问题记录:
- Error polling for event status: failed to query event: CUDA ERROR ILLEGAL INSTRUCTION
- could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR
文章目录
运行环境
最近在做遥感影像CNN语义分割的实验,在windows10上运行的,为了快速实现原型,采用了高层API框架Keras,后端采用了Tensorflow。实验过程中出现了不少问题,为防止再次踩坑,故记录下来。我的实验运行环境配置如下:
- Keras 2.2
- TensorFlow-gpu 1.8
- Python 3.6
- CUDA 9.0
- cuDNN 7.4
- OS: Windows 10
- GPU: GeForce GTX 1070 (8 G)
问题1. Error polling for event status: failed to query event: CUDA ERROR ILLEGAL INSTRUCTION
实验计算量比较大,本来是设置的训练期间运行30个epoch,但总会出现遍历几轮后报错退出的问题,错误信息为
Error polling for event status: failed to query event: CUDA ERROR ILLEGAL INSTRUCTI