UnetR training script - Githubissues

sct-pipeline / contrast-agnostic-softseg-spinalcord

Contrast-agnostic spinal cord segmentation project with softseg

MIT License

4 stars 3 forks source link

UnetR training script #52

Closed louisfb01 closed 1 year ago

louisfb01 commented 1 year ago

Build the training script for spine generic with the no_crop dataset using the UnetR model with MONAI.

louisfb01 commented 1 year ago

In progress on the lfb/unetr_multi_contrast_training branch.

[x] Prepare dataset for MONAI.
[x] Setup correct environment and update README.
[x] Training script adapting @naga-karthik's work with Unet3D but replacing it with UnetR.
[x] Adapt training script to CA script based on @ebadrian current training script.
[ ] Add comparison with other ivadomed models.

jcohenadad commented 1 year ago

[ ] I suggest to do a script vs. a function for reproducibility (ie: we need to hard-code the params for training the model). See eg this training script.
[ ] Add a link to wandb training experiments. This will help compare the performance of models.
[ ] On wandb, display the image/seg and output pred. See example script in this post.
[ ] Add a "test script" to be used for inference on single file. Example here.

naga-karthik commented 1 year ago

Closing this we are not trianing UNetR for this version of the contrast-agnostic model. A new, more targeted issue will be opened in the future if we decide to use for the next versions of the model on aggregated datasets