hyingho

CenterNet mixed-precision training cannot work well with specific cuDNN versions. ## How to reproduce * Revision: 859d9d4a * Branch: `master` * GPUs: V100 * Use `nnabla/nnabla-ext-cuda-multi-gpu:py310-cuda110-mpi3.1.6-v1.34.0` as the base image...

Mixed-precision training cannot work well (got NaN when processing) with specific cuDNN versions. According to our experiments, `cuda=11.0.3, CUDNN_VERSION=8.0.5.39` has this issue and it was solved in a newer cuDNN...