search

davidiommi / Pytorch--3D-Medical-Images-Segmentation--SALMON

Segmentation deep learning ALgorithm based on MONai toolbox: single and multi-label segmentation software developed by QIMP team-Vienna.
https://zmpbmt.meduniwien.ac.at/wissenschaft-forschung/qimp/
MIT License
120 stars 26 forks source link

About patch size #6

Closed Lnan1615 closed 3 years ago

Lnan1615 commented 3 years ago

When I tried to train my own model to segment the airway with CT images, the loss was consistently above 0.9. I tried to adjust the size of patch, but the loss was still not reduced. Are there any other parameters I need to modify?

davidiommi commented 3 years ago

You can try to change the learning rate, the optimizer (SDG to Adam), and the size of the patches.

It can be also that your patch size is too small compared to the full image size. You can resample the image and the patches can include more information with a lower resolution.

Every case is different, you have to play with it and experiment.

MIC-MRENet commented 3 years ago

Hi~, the step of （for batch_data in train_loader）is too slow. Do you have this problem? Thanks~

davidiommi commented 3 years ago

Try to set pin memory TRUE https://discuss.pytorch.org/t/when-to-set-pin-memory-to-true/19723