Issue collection

[kuangliu]

Hey guys, I'm super busy the two weeks. Finally I get some time working on this. For now, let's fix the issue one by one.

@njtuzzy:

@kuangliu @Mendel1 In the encoder file, the output of "get_anchor_boxes" is the "xcenter,ycenter, xwidth, ywidth" format, it seems that it does not need to change to xxwh(I guess you mean xywh) using change_box_order function?

anchor_boxes is ordered as xywh, boxes is changed from xyxy to xywh with change_box_order:

boxes = change_box_order(boxes, 'xyxy2xywh')

Now they are both xywh. Any problems?

[c0nn3r]

I do you think you should look at my original list of issues, but I do understand too many comments were made.

[Mendel1]

@kuangliu Can you upload your test result?Could your net predict correctly?

[kuangliu]

No. I know the loss is not working. I would like to know is there any bug in the encoder first.

[Mendel1]

I think decode is wrong.At least it's not the way the paper indicates.

[njtuzzy]

@kuangliu yes you are correct !

[kuangliu]

Cool. I update the loss function. I'm still testing it. The loss vibrates. Not sure what's going wrong.

[njtuzzy]

@kuangliu may i ask the class label is from 0 to 20 for voc? 0 means background?

[kuangliu]

Yeah. Just as Faster RCNN.

[njtuzzy]

@kuangliu did you initalize the bias of final cov-layer as section 4.1 mentioned?

[philokey]

@kuangliu the latest code will get nan in training processing.

[xueeinstein]

@kuangliu the loss explodes while training like this:

If it works for you now, please publish a detection demo, if possible, with a available pytorch model. Thanks in advance!

[kobehuai]

the function encoder.decode is work? I try to run the test.py after training but find the score is null after score, labels = cls_preds.sigmoid().max(1), I guess the gt is encoded and the cls_preds and loc_preds don't need decode when training, so it works well. But something go wrong in the test process when the preds output need decode, are there something wrong in the encoder.decode func?

[dami23]

@kuangliu Dear Kuangliu, where can I download the pretrained model metioned in the code? Could you give the address of the pretrained model? Thank you very much!!

[kehaozhe]

@kobehuai hi, have you solve this problems? I am working with the same question. thank you if you can share you solution.

[kobehuai]

@kehaozhe I find there are some nan samples in the cls_preds, so before the sigmoid, I set all the nan to -1: cls_preds[torch.isnan(cls_preds)] = -1 But the fundamental problem is that the network should not output nan, I still have not solved this problem.