SparseInst icon indicating copy to clipboard operation
SparseInst copied to clipboard

Published 20 hours ago • verified publisher icon hustvl

mAP equals 0 after 10 epochs of training

Open RocketFlash opened this issue 3 years ago • 2 comments

Hi! Thank you for amazing work!

I am trying to train r50vd_giam_aug model on my dataset in COCO format. In categories I have only 1 class with id=1 and each segment has category_id=1. I changed NUM_CLASSES parameter to 1 and registered dataset in train_net.py script. Also I am training on 1 GPU with batch size = 8 with AMP (reduced LR to 0.000005). After 10 epochs of training mAP is still equals 0. What am I doing wrong?

Screenshot 2022-09-16 at 16 10 36

RocketFlash avatar Sep 16 '22 10:09 RocketFlash

Tried to train the model without AMP and mAP start growing. I've read other issues and it seems something wrong with AMP training

RocketFlash avatar Sep 16 '22 14:09 RocketFlash

Hi @RocketFlash, thanks for your interest in SparseInst. The learning rate is too small and you can use the default learning rate for an attempt. As for AMP, we suggest you use the soft version of SparseInst to avoid the nan error.

wondervictor avatar Sep 17 '22 12:09 wondervictor