I can not get the same results in split 1 and shot 10? I just have one gpu 3090.

[kike-0304]

I just have one gpu 3090. And I do not change config, like lr and batchsize, how can I get the same result with paper?
My nAP50 = 60.8 (paper = 63.4)
final model (http://dl.yf.io/fs-det/model/)
datasets (http://dl.yf.io/fs-det/datasets/vocsplit/*.txt)

[qindengda]

Hello, how did you compile the code on the 3090 using cuda 11.0.

[kike-0304]

Hello, how did you compile the code on the 3090 using cuda 11.0.

I can not compile the code on the 3090, but you can run it with detectron2.

[RuoyuChen10]

Hello, how did you compile the code on the 3090 using cuda 11.0.

I can not compile the code on the 3090, but you can run it with detectron2.

I have reproduced the same results, I think there are several reasons:

• the best model is not the final results, you can try some other checkpoints saved by different steps.
• try 2 GPUs or more, I have tried, and the results with a single GPU are pretty lower than multi GPUs (61 -> 64).

I use the official compilation, not the detectron2 provided by the author.

[qindengda]

收到，谢谢！