Epoch Discrepancies and Training Conditions in Baseline Method Replication

[Yaoyongchen]

Thank you for your open-source work. I have a small question: why did you use 12 epochs for the baseline method you proposed but 36 epochs for the subsequent modifications? Also, could you please share the training conditions you used? My replication of the baseline method on DOTA1.0 only achieved a maximum of 0.75.

[YangWei231]

My replication of the model 'configs/rhino/rhino-4scale_r50_2xb2-12e_dota.py' also yields a result of 0.75, using a batch size of 2*4 (with four 4090 GPUs). Could I kindly ask the author if they could share the training method used?

[YangWei231]

My replication of the model 'configs/rhino/rhino_phc_haus-4scale_r50_2xb2-36e_dota.py' achieved a performance of 78.11, using bs 2*4 (with four 4090 GPUs), which is close to the reported result of 78.5. More attempts may help reach the performance reported in the paper.