Questions about detection pretrained weights

[Coldfire93]

I want to make sure that the file mv2-r50.pth in the detection pretrained weights you provided contains both teacher's and student's weights.

Thank you!

[akuxcw]

Yes, I just checked the mv2-r50.pth contains both teacher and student.

[Coldfire93]

Hi @akuxcw , The size of the mv2-r50.pth is 266MB. The teacher model is about 160MB ( I'm not sure. I used torchvision's pretrained weights before. It is about 160MB. Your teacher's weights come from Detectron2's pretrained detector. Was there a particular reason why you chose to use Detectron2 instead of torchvision?). So, the student model's size is about 100MB? Is that right? Thank you!

[akuxcw]

We use the teacher model from Detectron2 because it provides pretrained detectors and we train our student based on Detectron2. I think there are no pretrained detectors in torchvision, torchvision only provides pretrained weights on ImageNet. Which model do you use before? And we use this model as teacher, which is about 230M, so student's model is about 30M.

[Coldfire93]

Hi @akuxcw , We used this model as teacher. It provides detectors too.

And we use this model as teacher, which is about 230M

Is the backbone of the link you gave above ResNet101? Could you please send me the link of model with ResNet50 as backbone? Because the mv2-r50.pth - teacher model's backbone is ResNet50.

Thank you!

[akuxcw]

Sorry, faster-rcnn-50 is available here. And you are right about the model size. I really don't know torchvision provides pretrained detectors. Thank you for telling me that! And we use detectron2's pretrained weights just because we implement our method based on detectron2.

[Coldfire93]

Hi @akuxcw , I see. Thank you~

[twmht]

@akuxcw

Did you also switch teacher to training mode for BN when training detector?