search

seannz / svr

[CVPR2024] Fully convolutional slice-to-volume reconstruction for single-stack MRI
https://seaniyoung.com/publications.html
MIT License
27 stars 7 forks source link

GPU requirements in README #5

Open tomjf opened 3 months ago

tomjf commented 3 months ago

Hello, thank you for sharing the code.

What size GPU was used in training? I couldn't find this in the paper. I am running into OOM issues after 100 epochs on a 24GB RTX A5000. It would be helpful to state the GPU requirements in the README, especially if they are large. Or perhaps I have made a mistake.

seannz commented 3 months ago

Hello Tom,

Thanks for your interest in our work — we trained slightly different models for adult brain SVR and fetal. IIRC both were trained on an RTX 8000 with 48gb of ram. Since the publication of the paper, we have made some architectural tweaks to our models and are in the process of training smaller and more accurate models right now, but in the mean time please feel free to train a smaller model listed in flow_SNet4.py. Let me know if you have any questions —Sean

tomjf commented 3 months ago

Thanks Sean, I tried one of the smaller models and that works. Thanks again. Is the splat-slice U-Net in the paper flow_SNet3d0_192?