qiang chen

You can manually add the freeze codes for backbone and encoder [here](https://github.com/chensnathan/YOLOF/blob/master/yolof/modeling/yolof.py#L133) in the yolof file as we do in the [darknet](https://github.com/chensnathan/YOLOF/blob/master/yolof/modeling/backbone/darknet.py#L341).

Could you provide the full log file?

Which model are you training with? and what command do you use? It will be more clear if you post your training log file in the `log` directory.

It seems that you modify the `BASE_LR`. But it's weird. We have not encountered this error before. Could you try the standard 8 GPUs setting and train YOLOF again?

It seems that there is somewhere the code performances ZeroDivision. Does this error occurs definitely or randomly?

[This](https://github.com/Megvii-BaseDetection/BorderDet/issues/6) may help you.

You may provide more environmental details so that I can help.

We will provide a demo script recently. Before that, you can refer to [the detectron2's demo](https://github.com/facebookresearch/detectron2/blob/master/demo/demo.py).

Hi, For the first question, the indexes are actually selected in each image. Using batch is an implementation to avoid loop calculation. For the second question, you can refer to...

Hi, we did not train tiny models before. But I am happy to help get reasonable results. Could you provide more details about your modification? The backbone file, pre-trained models,...