Tune for COCO

[Flova]

Proposed changes

• tune hyperparams
• scale based on mini batch size
• linear instead of step wise decay
• use sgd
• only log once per optimizer step

Related issues

• very low mAP on coco val2014 when training from scratch #818
• how to adjust learning rates or something to be better when using a pretrained model to train my dataset?please give me some advice #820

Necessary checks

• Update poetry package version semantically
• Write documentation
• Create issues for future work
• Test on your machine

[J-LINC]

hi ，Regarding what you mentioned:linear instead of step wise decay, I have reviewed your code and found that there is only one learning rate decay strategy, which is to adjust according to steps. How can I perform linear decay or other decay methods
as for using sgd ,i think you are right for most trains and i successed when i using sgd in trianing bdd100k