tf-gpu版本为1.12，训练的时候报错Failed to get convolution algorithm. This is probably because cuDNN failed to initialize，请问这种问题应该如何解决？谢谢~~

MaybeShewill-CV / CRNN_Tensorflow

Convolutional Recurrent Neural Networks(CRNN) for Scene Text Recognition

MIT License

1.03k stars 388 forks source link

Closed Leedorado closed 5 years ago

MaybeShewill-CV commented 5 years ago

@Leedorado 查看你的cudnn, cuda, cuda driver, tensorflow是不是匹配吧:)

Leedorado commented 5 years ago

@MaybeShewill-CV 请问你的cuda、tf、cudnn版本是多少？如果是版本不匹配的原因话，是不是连一开始的write_tfrecords.py都跑不起来啊

Leedorado commented 5 years ago

我现在用的版本是tf-gpu 1.12 cuda 9 cudnn 7

MaybeShewill-CV commented 5 years ago

@Leedorado 跟你的一样cudnn 我是7.5

Leedorado commented 5 years ago

@MaybeShewill-CV 请问训练的时候是对显存占用做了限制吗？我提高batch_size到64时的显存占用与16时的显存占用都是21%左右