how to get the best dice checkpoint during training?

KidsWithTokens / MedSegDiff

Medical Image Segmentation with Diffusion Model

MIT License

976 stars 145 forks source link

how to get the best dice checkpoint during training? #136

Open LiuTingWed opened 9 months ago

LiuTingWed commented 9 months ago

Good job！ But how can i get the best dice checkpoint during training? It seem there is no relative eval code in the project during training.

maomaocun commented 8 months ago

In my work,isic dataset got the best presentation in checkpoint 80000iterations batch_size =8 MedSegdiff-v1 basic,iou=0.8197,dice=0.8981

LiuTingWed commented 8 months ago

Well，sir，could you tell me how could you get the best result？ just load each checkpint and eval each .pt one by one？

maomaocun commented 8 months ago

yes,just sample from about 60000iterations 65000iteration ...... to 11500iterations,and you can see the 80000iteration did the best ,my gpu is burnning!!!!!!!!!!!!!!

maomaocun commented 8 months ago

can I communicate with you in Wechat? 19162076385

wzx0720 commented 6 months ago

can I communicate with you in Wechat? 19162076385

@maomaocun Hi, I also do my research on MedSegDiff-v1, but I met some problems while sampling. May I communicate with you on WeChat, too?