Question about the inference time. #65
Description
Hello.
Fantastic work on pointcloud self-supervised learning.
However, I'm quite confused about the inference time during dvae pretraining. It seems that the inference time is far longer than the training time. (For 2x NVIDIA-A100 40GB, the training time for one epoch is ~3 minues while the inference time is ~10 minutes.) The nvidia-smi command shows that the GPU utilization is ~1% during inference, while the CPUs are running at full utilization.
When I dig deep into the code, I discovered that the inference batch size is 1. When I adjust the inference batch size to the same as the training batch size, the inference time is still far longer than training and the utilization of CPUs and GPUs barely change.
I just wonder whether changing the inference batch size would affect the performance. Also, why does inference take far longer than training?