Cannot reproduce the results of faster-rcnn

[jonbakerfish]

For bugs or installation issues, please provide the following information.
The more information you provide, the more likely people will be able to help you.

Environment info

Operating System: Debian GNU/Linux 8 (jessie)

Compiler: gcc & g++ (4.9.2)

Package used (Python/R/Scala/Julia): Python

MXNet version: 0.11.1

Or if installed from source:

MXNet commit hash (git rev-parse HEAD): 2372518

If you are using python package, please provide

Python version and distribution: 2.7 / anaconda

Error Message:

Cannot reproduce the mAPs of faster-rcnn with resnet101+voc2017+voc2012.
The mAP I got is ~0.69 which is far less than the claimed 0.79.

If I use the pre-trained model from https://github.com/precedenceguo/mx-rcnn, the mAPs is correct. Please also see this issue.

Minimum reproducible example

in the rcnn folder, script/resnet_voc0712.sh 0

[solin319]

I got mAP 0.69 ,too.
Why all the gamma and beta were fixed in config?
Why "use_global_stats" in "symbol_resnet.py" was always true in train and test?

[szha]

@apache/mxnet-committers: This issue has been inactive for the past 90 days. It has no label and needs triage.

For general "how-to" questions, our user forum (and Chinese version) is a good place to get help.

[rahul003]

@cjolivier01 Could you please label this with question, performance

[stu1130]

@jonbakerfish @solin319 confirmed that the example still works. got mean AP= 0.70, 0.74 and 0.78 on vgg voc2007, vgg voc2007+2012, resnet voc2007+2012.
You can use python3 train.py --dataset voc --imageset 2007_trainval+2012_trainval --network vgg16 --pretrained model/vgg16-0000.params --gpus 0,1 to do training. Once the PR is merged the issue will be closed, feel free to reopen it if you have further question