I installed tensorflow 2.0.0 and Cuda 10.2, I have a new gtx 1660 ti and I’m trying to train a convolutional neurale network but I received this error: “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. [Op:Conv2D]”. How can I fix this?
Please see this related issue for a couple debug tips: https://devtalk.nvidia.com/default/topic/1068031/cudnn/geforce-gtx-1660-super-cuda-not-working-in-anaconda
It might just be an OOM error, meaning your GPU might not have enough memory for your model.
No, its not a memory error
It may well just be a mismatched environment configuration issue then as mentioned in #3 here: https://stackoverflow.com/a/56511889/10993413
I noticed you mention TF2.0 and CUDA10.2, but you don’t mention a CUDNN version. Do you have CUDNN installed? If so, what version?
According to TF docs, they haven’t tested a CUDA 10.2 configuration yet: https://www.tensorflow.org/install/source#linux. You could try downgrading to CUDA 10.0 since that’s tested according to the docs mentioned.
Alternatively, you could try using our NGC TF container which removes the pain of host-side dependencies: https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow
According to the release notes (https://docs.nvidia.com/deeplearning/frameworks/tensorflow-release-notes/rel_19.11.html#rel_19.11), nvcr.io/nvidia/tensorflow:19.11-tf2-py3 has been built with TF2.0 + CUDA 10.2
my bad, the CudaNN version is 7.6.4 cuda10.0_0.
Thank you for replying me
I am facing the same issue. This is on Windows 10 env. The CudaNN version is 7.6.5.
Please see this thread, it should be very similar to your issue: https://devtalk.nvidia.com/default/topic/1068031/cudnn/geforce-gtx-1660-super-cuda-not-working-in-anaconda