search

ZiweiWangTHU / BiDet

This is the official pytorch implementation for paper: BiDet: An Efficient Binarized Object Detector, which is accepted by CVPR2020.
MIT License
173 stars 34 forks source link

Question on pretrained models. #9

Closed killawhale2 closed 4 years ago

killawhale2 commented 4 years ago

I've sent an email to the authors roughly 2 weeks ago about the pretrained models but haven't got a response yet (maybe the email didn't go?) so I'm re-iterating the question here.

For obtaining the pretrained weights of ResNet18/VGG16, do you train the networks as a floating point network or do you binarize the networks like in XNOR-Net/Bi-Real Net and then train them on ImageNet to obtain the pretrained weights?

I'm trying to use different backbone networks and an answer to this question would help me in obtaining the pretrained weights for my networks.

Wuziyi616 commented 4 years ago

Hi! Yes, I didn't get the email you mentioned, maybe you sent it to the wrong person I guess. As for your question, yes we trained binarized backbone networks on ImageNet for our detector. For example, we follow Bi-Real Net's training strategy to train our Bi-Real-VGG/ResNet (first real-valued with relu activation, then clipping activation, finally binary activation & weights). The training takes a long time (about 2 days for one model), and I use a large batch size on 4 GPUs to stabilize the training of the binary backbones. Hope my answer can help you!

killawhale2 commented 4 years ago

Thank you for your quick reply!