TIO-IKIM / CellViT

CellViT: Vision Transformers for Precise Cell Segmentation and Classification
https://doi.org/10.1016/j.media.2024.103143
Other
236 stars 41 forks source link

Question regarding the batch size #74

Open won5830 opened 3 days ago

won5830 commented 3 days ago

Hi thank you for your wonderful contribution. I'm currently trying to retrain the model using your data with mixed precision. However, I have a question about your batch size regarding CellViT initiated with SAM-B, H, and L. What was your batch size for the training ? was it all 1 ? I found that you used 16 for in your supplementary data but I found that GPU run out of its memory even with 80 GB memory. Is there any more things to pay attention other than mixed precision for training such large model ?