I am practicing building a TensorFlow transfer learning pipeline for EfficientDetD0 in tensorflow.applications. However, whenever I try running my program, I encounter the following errors:

Epoch 1/3
2023-06-27 15:54:39.937126: E tensorflow/compiler/xla/stream_executor/cuda/] Could not create cudnn handle: CUDNN_STATUS_NOT_INITIALIZED
2023-06-27 15:54:39.937170: E tensorflow/compiler/xla/stream_executor/cuda/] Memory usage: 5111808 bytes free, 8325824512 bytes total.
2023-06-27 15:54:39.937224: E tensorflow/compiler/xla/stream_executor/cuda/] Possibly insufficient driver version: 525.116.4
2023-06-27 15:54:39.937245: W tensorflow/core/framework/] OP_REQUIRES failed at conv_ops_impl.h:770 : UNIMPLEMENTED: DNN library is not found.

I am not sure if this error is because of my cudnn in my local machine or in tensorflow.applications model itself.

Node: 'model/efficientnetb0/stem_conv/Conv2D'
DNN library is not found.
	 [[{{node model/efficientnetb0/stem_conv/Conv2D}}]] [Op:__inference_train_function_17233]


Welcome to the Tensorflow Forum!

The error CUDNN_STATUS_NOT_INITIALIZED caused by either an incorrect GPU driver with an insufficient version or an incompatible version of cuDNN in w.r.t CUDA.

You can refer below image for officially tested build configuration for Linux

You can download the required GPU driver from NVIDIA Driver Downloads.

Additionally, for the most up-to-date information on software compatibility, including the operating system, NVIDIA CUDA, the CUDA driver and NVIDIA hardware, you can refer to the NVIDIA cuDNN Support Matrix.

Thank you!