Closed liujunyan691 closed 2 weeks ago
Hoping for your answer! Thanks again
Please, use the provided weights. They give the same results as in the paper.
Please, use the provided weights. They give the same results as in the paper.
Could you please provide an evaluation script that supports the new data format, where the labels are only in the labels
folder as JSON files?
Ivtmetrics supports the new data format, check documentation for more details
On Fri, 20 Sept 2024, 23:29 Jiajie Li, @.***> wrote:
Please, use the provided weights. They give the same results as in the paper.
Could you please provide an evaluation script that supports the new data format, where the labels are only in the labels folder as JSON files?
— Reply to this email directly, view it on GitHub https://github.com/CAMMA-public/rendezvous/issues/24#issuecomment-2364642139, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEOK5UXDUMKSNVWPTTDZ6F3ZXSHT3AVCNFSM6AAAAABJ7Z2TOWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGNRUGY2DEMJTHE . You are receiving this because you modified the open/close state.Message ID: @.***>
Thank you for your work and sharing. I have read your papers on dataset splits and RDV, but I have found that the default training method in the code is different from the training method described in the paper. I have trained using both methods and have not been able to reproduce the results on cholect50 using RDV split (29.9 on pytorch). Could you please provide the model weights corresponding to the 29.9 result on pytorch and describe the corresponding training strategy? (I saw in the previous answer that one of the reasons is the use of different augmentation style per epoch, which is different from the simplified dataloader in the git repo. But this description is too simplistic to reproduce it as a baseline for me). Can you please share your hyperparameters and describe which training strategy corresponds to 29.9?