Slow Training and High Losses

[Sanjay1911]

Hi @qfgaohao & everyone!

I'm currently implementing an anonymization pipeline for privacy specific data as a part of my mini-thesis and is going to delpoy it in a jetson xavier. I use Widerfaces and Large License plate dataset as a multi class object detection problem.

When I start training with train_ssd.py , my training for each epoch is super slow even though I use L4 GPU from Colab. And my log says "Using CUDA" but GPU is not utilized at all. Is there any specific reason or am I missing something?

Also my loss values are high. I would really appreciate any help/suggestion regarding this issue.