search

MIC-DKFZ / MedNeXt

[MICCAI 2023] MedNeXt is a fully ConvNeXt architecture for 3D medical image segmentation.
https://arxiv.org/pdf/2303.09975
Apache License 2.0
345 stars 26 forks source link

Sharing the precise setup for your experiments #14

Closed Zhack47 closed 6 months ago

Zhack47 commented 10 months ago

Hello @saikat-roy, thank you for open sourcing this work!

I am currently trying to reproduce it on my side, however, using the same conditions as in your supplementary material (i.e.: batch size= 2, patch 128x128x128 and 250 it/epochs wit the Small (S) model and k5 kernel, i get a 200 sec/ epoch runtime, different from your 117 secs.

I have eliminated the other options, so I am now left with setup differences. Could you give us the setup you used ( GPU, pytorch version..) to train your model ?

Regards

Zach

saikat-roy commented 9 months ago

Hey @Zhack47. My first thought is that maybe I used a Tesla A100 40GB for baselining the S model without any checkpointing, and that your GPU is different. If that doesn't help, I can share my remaining setup.