zilunzhang
commented
9 months ago Thank you very much for your outstanding work！ I have a question that I haven't quite understood. When fine-tuning your RS5M model on RSICD or RSITMD using the methods outlined in the paper (infoNCE, lr=1e-6), I did not achieve the expected performance. Taking RSICD as an example, the paper and the weights you provided for RS5M RET-2 version result in an accuracy around 38, but when I fine-tuned using my own RS5M VitB32 version, the result was around 34. Could you provide more details on fine-tuning RET-2 or RSICD so that I can better replicate the process? Thank you very much.
Hi
We used the ITRA codebase to tune the model.
The RET-2 data was made by: https://github.com/ChenDelong1999/RemoteCLIP/issues/13#issuecomment-1836218269
Hyperparams for RET-2 model:
--lr: 5e-06 --weight_decay: 0.5 --batch_size: 600 --epochs: 7 --warmup 100 --max_grad_norm: 50.0
Best,
Zilun
mcx-mcx
lemyx
Thank you very much for your outstanding work！ I have a question that I haven't quite understood. When fine-tuning your RS5M model on RSICD or RSITMD using the methods outlined in the paper (infoNCE, lr=1e-6), I did not achieve the expected performance. Taking RSICD as an example, the paper and the weights you provided for RS5M RET-2 version result in an accuracy around 38, but when I fine-tuned using my own RS5M VitB32 version, the result was around 34. Could you provide more details on fine-tuning RET-2 or RSICD so that I can better replicate the process? Thank you very much.