Use faster-rcnn training DPN network by alt

Question

Use faster-rcnn training DPN network by alt

zhujiapeng opened this issue 7 years ago · 5 comments

Hi,
I am trying to training dpn network using faster-rcnn by 4-steps method, the training process is normal, both the rpn loss and fast-rcnn loss are reducing. But during the test, the result is very poor, Have you used 4-steps to training dpn network?

Answer 1 · 2017-10-31T13:45:13.000Z

I am sorry, I only used the end2end training procedure.

Answer 2 · 2017-11-01T08:23:06.000Z

Another question, I am trying to use end2end method to training DPN networks, when I use the dpn92-merge.caffemodel to initial the network, the output of some loss are very large(in my case are loss_bbox=210.396, loss_cls=87.0391, rpn_cls_loss=0.7288, rpn_loss_bbox=0.8088), So It made me can't training the network, but when I use the dpn92.caffemodel to initial the network, the loss is normal and decreasing by the time. But the test result is poor... Did you meet such question in your training ? And I think this is also the answer of my first question, in which I also initial the network using dpn92.caffemodel

Answer 3 · 2017-11-01T11:06:14.000Z

I think you should check that your train.prototxt is also merge bn-scale or not?

Answer 4 · 2017-11-01T12:02:16.000Z

@soeaver I used the network on your github 'det/faster_rcnn/models/pascal_voc/dpn92/faster_voc_dpn92-merge.prototxt' initial weights are 'dpn92-merge.caffemodel' to training using end2end method. Got the loss I described in last question.

Answer 5 · 2017-11-01T13:48:00.000Z

Sorry, so that nothing I can help.