Hyper-parameters setting

[jie-li-hust]

Hi! Thank you for the great work.
I’m fine-tuning this model, and using the checkpoint trained with focal seems to lower performance on COCO and RefCOCOg. Could the batch size be influencing this as well?
I am training on 4 GPUs, the setting is "TRAIN.BATCH_SIZE_TOTAL 20 \ TRAIN.BATCH_SIZE_PER_GPU 5 \ DATALOADER_NUM_WORKERS 4" ,and results on RefCOCOg (~62-63% cIoU),COCO(~38.1% mAP, ~60.7% mIoU).
Would increasing the epoch parameter, reducing the LR_MULTIPLIER for the backbone (to 0.05), or lowering WARMUP_ITERS（to 5） be helpful?

[jie-li-hust]

By the way, the demo code shows davit.py, but how can we calculate the parameters using davitd5_unicl_lang_v1.yaml? Also, I was wondering why the BACKBONE_NAME is set to 'davit' instead of 'davitd5'?