hyperparameters

[berad1ar]

thanks for your work.I have a question need your help.
my dataset has 2000 images for training, and their objects are small.
I need to change hyperparameters to increase P and mAP. In your opinion, which hyperparameters should be changed and approximately how much should I increase or decrease them?

[glenn-jocher]

@berad1ar its impossible to say. Start with the default hyps to establish a benchmark, then use hyperparameter evolution to automate the process.

Evolution command for coco_16img dataset for example:

while true
  do 
    python3 train.py --data data/coco_16img.data --img-size 320 --batch-size 16 --accumulate 1 --cache --evolve
done

See hyperparameter evolution: #392

[glenn-jocher]

@berad1ar closing as duplicate of #392, please post any comments there.