ppengtang/pcl.pytorch

How to using Ens

zhengshangdong opened this issue · 2 comments

zhengshangdong commented

Thanks so much for your great job！
I have ran your code and mAP is about 50.05%~50.90% in my experiments.
I suppose the 54.1%mAP in your experiments can be gained by Ens.?
So, how to realize the Ens. operation or have the Ens. operation contained in your code(where).

Looking forward to your reply！

HMCCMH commented

I want to know it, too.

ppengtang commented

Hi, sorry for the late reply. The 54.1%mAP is obtained by a single VGG16 based model (no Ens.) using vgg16_voc2007_more.yaml.
It seems that vgg16_voc2007_more.yaml is not very stable and you could try vgg16_voc2007.yaml to get more stable results. Or you could use my trained models here.

I did ensemble for caffe version models only and haven't tried it for PyTorch version models. To do ensemble, I trained a VGG_M based PCL model and a VGG16 based PCL model, and computed the average score from these two models for each proposal.