I've checked the pre-trained models provided by garrickbrazil could reproduce mAP with this docker environment.
I've not tested the models trained with the docker could reproduce the results.
Reports from someone tried training are more than welcome.
I used the same versions for pytorch, torchvision, and other libraries (see docker/requirements.txt) so I hope it reproduces training as well.
Hi @garrickbrazil, thank you for making the code public.
I added Dockerfile to this repo. I hope this helps someone who cannot prepare old CUDA environment in local.
For usage, please see docker/README.md.