OTB configs / results don't match

[amoskalev]

Hi,

I trained models with siamrpn_r50_l234_dwxcorr_8gpu / siamrpn_r50_l234_dwxcorr_otb_8gpu configs and tested on OTB. The highest score I got after hyper-parameter tuning is 0.687 / 0.676 resp.

I trained siamrpn_r50_l234_dwxcorr_8gpu on 1 1080Ti and siamrpn_r50_l234_dwxcorr_otb_8gpu on 4 1080Ti's.

I am a bit confused as 0.696 is reported in Model zoo for OTB. Did I miss something during the training? Is training results are dependent on # of gpus?

[ZhiyuanChen]

Have you adjusted batch size, learning rate, etc.?

[amoskalev]

@zyc-ai No, I didn't modify configs

[ZhiyuanChen]

@zyc-ai No, I didn't modify configs

Then you probably should do it.

[amoskalev]

Can you tell what should I adjust in configs?

[amoskalev]

Hi @zyc-ai , can you help?

[amoskalev]

@zyc-ai @lb1100 @StrangerZhang Hi, how should I modify config to match 0.696 score on OTB?

[ZhiyuanChen]

It would take a lot of efforts to reproduce, just try.