About the train

[G-Apple1]

Thank you for your nice project!
I called your swint-Detectron2 code to train my own data set. I used the network configuration of retinanet. No error was reported during the training process,iteration=200000 AP50=0.7 or so, and the test set is only AP50=0.5 or so. I have changed pixel_mean according to my dataset.
I would like to know what other parameters should be paid attention to in training, which may be helpful for my training!

[xiaohu2015]

@G-Apple1 the best is to use https://github.com/xiaohu2015/SwinT_detectron2/blob/main/configs/SwinT/retinanet_swint_T_FPN_3x.yaml
for your dataset, I think you can adjust some params (lr, train size, and so on)to try to get better result.