about the precision

[Hello526]

Hello, I used my own dataset to train and found that the precison is always low, about 60%, and recall is about 96%. I don't know which parameter is related to this?

[glenn-jocher]

If you increase the iou threshold hyperparameter in train.py this will increase precision and lower recall.

[Hello526]

THX!

[Hello526]

I added cls in train.py and found that precsion increased by 10%.

[Hello526]

[Hello526]

75.6% to 85.2% for precision

[glenn-jocher]

@Hello526 ah, this is a huge jump in precision. What change did you make exactly?

[Xiwang-pge]

Which Threshold parameter did you change in train.py? iou_t?

[Hello526]

@glenn-jocher I have been using a fixed parameter set before, and later set evolve=True, precision has improved a lot.

[glenn-jocher]

@Hello526 well that's good news. Yes you can use --evolve for hyper parameter evolution. For coco we've been doing something like this to set the hyper parameters, and then letting it run for about 100 trials:

!python3 train.py --data data/coco.data --img-size 320 --single-scale --epochs 1 --batch-size 32 --accumulate 2 --giou --evolve