search

fabiopoiesi / dip

Project page of the paper "Distinctive 3D local deep descriptors" accepted in IEEE International Conference on Pattern Recognition 2020.
64 stars 4 forks source link

How did you get the final_chkpt.pth? #13

Closed wenli4313 closed 2 years ago

wenli4313 commented 2 years ago

Hi,I'm very interested in your work. How did you get the final_chkpt.pth? I am looking forward to your reply.Thank you very much.

fabiopoiesi commented 2 years ago

I used train.py of this repository. I had problems with DataParallel, and I could not reproduce the results. Try not use DataParallel and run a training only with one GPU. It will take longer but you should be able to reproduce a similar checkpoint as mine.

wenli4313 commented 2 years ago

I ran train. Py and used the trained model ckpt_e39_i14400_dim32.pth to test benchmark_eth_pre.py. The result and the parameter setting is as shown in the attachment. I have been unable to find the reason, please help me to check whether there is a problem with the parameter setting.Thank you very much.

------------------ 原始邮件 ------------------ 发件人: "fabiopoiesi/dip" @.>; 发送时间: 2021年11月23日(星期二) 凌晨4:05 @.>; @.**@.>; 主题: Re: [fabiopoiesi/dip] How did you get the final_chkpt.pth? (Issue #13)

I used train.py of this repository. I had problems with DataParallel, and I could not reproduce the results. Try not use DataParallel and run a training only with one GPU. It will take longer but you should be able to reproduce a similar checkpoint as mine.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

wenli4313 commented 2 years ago

I am very interested in your work. Please reply when you see the message. Thank you very much

fabiopoiesi commented 2 years ago

This problem was addressed by setting Shuffle to false in the dataloader of the training function and by using only one GPU for training.