cuDNN failed to initialize

Question

tensorflow.python.framework.errors_impl.UnknownError: 2 root error(s) found.
  (0) Unknown: 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.
     [[{{node conv2d_1/convolution}}]]
     [[norm_dense_1/Softmax/_3605]]
  (1) Unknown: 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.
     [[{{node conv2d_1/convolution}}]]

This error was encountered when trying to run a project from Github.I have tried reinstalling all the programs from scratch in a new conda environment. Can't seem to figure out whats the problem.

GPU Zotac gtx 1070 ti

My current software installation is as follows:-

• Python 3.6
• Tensorflow GPU 1.15
• CUDA 10
• cuDNN 7.4

Answer 1

Ensure you have CUDA Toolkit 7.5 Please open a new session then just try to load tensorflow

Are you able to load tensorflow without any error? Make sure all the ddl cudard64_101.ddl and others are loading. If there are no issues with tensorflow then run the following

physical_devices = tf.config.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(physical_devices[0], True)

Once done try running code again

Answer 2

I had the same output. The problem for me was that i had accidentally set two different environment variables for cuda. The PATH had 2 locations for cuda folder and that caused the problem for me. Removing one of them solved the problem for me.