search

chengyangfu / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
169 stars 47 forks source link

About the "use_global_stats" parameter of BN layer setting in your train.prototxt #15

Open ghost opened 6 years ago

ghost commented 6 years ago

Hi, Thanks for your sharing. Emm..., I try to realize the results you put in DSSD paper which I have cited in my last ICME paper, but I have failed many times, not only on PASCAL VOC, but COCO. When I check your train.prototxt, I find that you set the parameter of "use_global_stats" as true, it makes me confused, according to the realization of BN layer in Caffe, is not it should be "false" for training and "true" for testing? I have tried many times to run your codes, check many times, still failed... Could you give me some advise ? I will be very appreciate it!! Best

ghost commented 6 years ago

I run the SSD321 with ResNet101, I only got 60+ in VOC 2007 test set...

chengyangfu commented 6 years ago

Not sure which part is wrong. But I think when trainning SSD321 with ResNet101, the "use_global_states" should be false. I turn that to True when training DSSD321.