Question about the model training

[kevinLee1990]

Hi thank you for your excellent work. I meet a problem when re-run your experiments.

When I re-train the Instance segmentation model with R-50 on COCO dataset, the results are:
43.5, 23.0, 47.0, 65.1
43.2, 22.7, 46.4, 64.8
which are a bit lower with your reported number:
43.7, 23.4, 47.2, 64.8

I use the standard configuration file to run the experiments without any modification, and run on 4/8 V-100 cards. I do not know whether it's just a common scenario, or did you meet the same problem during training?

[bowenc0221]

I have not systematically evaluated the variance on the COCO dataset, but the gap (+/- 0.2 AP) seems reasonable.

[kevinLee1990]

@bowenc0221 Thx for your replying. I re-train your model again in my side, and it achieves 43.3% AP, which is still lower than the reported number (0.4% lower). More specifically, for AP_s, the variance is even larger in my side (-0.4, -0.7, -0.6). Do u have any idea or suggestion to address this issue?