Environment problems

[wentao-cv]

Hello Mr. yujjyu，I'm a student in school and I admire your research on wavelet flow. Recently, I ran your published code and encountered the problem of GPU efficiency. I hope to get your help. My environment is tensorflow 1.15.0, Python 3.6.8, CUDA 10.0 and its corresponding cudnn 7.6.0. The GPU is NVIDIA 2080ti. The GPU efficiency of the program is only about 40%. Is that the same for your running efficiency? How did you solve it? If it's convenient, could you tell me the environment in which you run the program? Especially CUDA 、 cudnn and GPU.By the way, I tried to switch CUDA 10.1 and CUDA 10.2. As a result, an error will be reported when running. Only CUDA 10.0 can run.
I would appreciate it if you could help me.

[JasonYuJjyu]

When you say the "GPU efficiency" I assume you mean the GPU utilization shown when running nvidia-smi. I assume you are trying to run the model for high resolution data at 1024x1024. The main bottleneck in that case is the data loading since the included data loader reads in PNGs and has to decompress them which can be slow. One way to speed that up is to just use more CPU cores. Another way is to convert the data into tfrecords and modify the pipeline to read those instead. This is much faster since there is no decompression needed when loading data, but it uses more IO.