search

sanghoon / pva-faster-rcnn

Demo code for PVANet
https://arxiv.org/abs/1611.08588
Other
651 stars 241 forks source link

Fail to convert full/original.model to full/test.model #15

Closed John-HW-Cao closed 7 years ago

John-HW-Cao commented 7 years ago

Hi

I am trying to use the method in https://github.com/sanghoon/pva-faster-rcnn/issues/5 to merge conv, bn, scale layers in original.model to a single conv layer in test.model.

But I failed. The file I obtained is different from full/test.model

And I check the full/original.model find that there are many zeros in original.model, which causes the failure.

So could anyone please tell me that if it is possible to do such a conversion by that method? Is there another trick in the conversion?

sanghoon commented 7 years ago

Although we've done it with our own script (which wouldn't be generalized well enough), there're some good discussions about it at #5. I'm going to repeat the procedure with the script mentioned at #5

sanghoon commented 7 years ago

I've committed a simple script to merge 'Conv-BN-Scale' layers into a single Conv layer. Please checkout 39570aa.