Problem with First Stage Training

[Ze-Yang]

Hi, intuitively, I first download the model you provide and finetune on the few-shot dataset so as to reproduce your results after finetuning. Basically, I can reproduce your results with the pretrained model you provide. Afterwards, I would like to train the base model on my own. In the first stage, I can get similar base mAP with yours, e.g., 81.02 for split1. However, when I perform the second stage finetuning with the model trained by myself, the novel mAP will drop quite a lot, e.g., from 55.89 to 51.91 on split1 5shot case, compared to directly use the model you release. Is there any trick that maybe omitted in the code you released for the first stage training? Thanks.

[davesie]

What dataset did you use?

[MartinYYYYan]

how many gpus are you using?