NN-UNet

[iffthomas]

Hi, Thank you for implementing an nnUnet api for monai.

I'm currently playing around with your implementation.
I'm curious if there exist a possiblity to change the trainer_class_name to nnUNetTrainer_250epoch while only training for the 2D case on all folds I have (I've got 5).
what would be a suitable approach to achieve this?
I've tried adding --config "2d" that my command would look like this
python -m monai.apps.nnunet nnUNetV2Runner run --input_config='./input.yaml' --trainer_class_name nnUNetTrainer_1epoch --config "2d"

[KumoLiu]

Hi @iffthomas, for how to run nnunet, you can refer to this tutorial.

MONAI/monai/apps/nnunet/nnunetv2_runner.py

Line 957 in 2658b00

self.train()

MONAI/monai/apps/nnunet/nnunetv2_runner.py

Line 559 in 2658b00

configs: tuple | str = (M.N_3D_FULLRES, M.N_2D, M.N_3D_LOWRES, M.N_3D_CASCADE_FULLRES),

If you use the run command, it will run all the configs by default.

Hope it helps, thanks!

[iffthomas]

Also What I figured out,
python -m monai.apps.nnunet nnUNetV2Runner train --input_config='./input.yaml' --trainer_class_name nnUNetTrainer_250epoch --config "2d" get me an error message, is this intended?

below my logs
nnunet_trainer = get_trainer_from_args(dataset_name_or_id, configuration, fold, trainer_class_name,
File "/cluster/home/ifft/miniconda3/envs/venv-spine/lib/python3.10/site-packages/nnunetv2/run/run_training.py", line 42, in get_trainer_from_args
raise RuntimeError(f'Could not find requested nnunet trainer {trainer_name} in '
RuntimeError: Could not find requested nnunet trainer nnUNetTrainer_250epoch in nnunetv2.training.nnUNetTrainer (/cluster/home/ifft/miniconda3/envs/venv-spine/lib/python3.10/site-packages/nnunetv2/training/nnUNetTrainer). If it is located somewhere else, please move it there.

[KumoLiu]

Hi @iffthomas,

• If you want to use the one-line command run to start the nnU-Net workflow, you can modify the train here manually by specify the "configs='2d'".
• And for the error, thanks for pointing out the typo here, I will create a PR to fix it.
  Thanks!

[iffthomas]

Sorry, I'm rather new to programming in general,

• If you want to use the one-line command run to start the nnU-Net workflow, you can modify the train here manually by specify the "configs='2d'".

what the best approach in doing this?
Go to my downloaded files where I've installed Monai NN-Unet and then changeing it there?

[KumoLiu]

Yes, but it's not always a good habit to modify the source code, so instead of using the one-line command, you could run it separately, step by step, so that you can specify the config to be used during the train.

MONAI/monai/apps/nnunet/nnunetv2_runner.py

Line 117 in fbbb64b

--trainer_class_name "nnUNetTrainer_5epochs" \\

Perhaps we could add more flexibility to this one-line command? What do you think? cc @dongyang0122

MONAI/monai/apps/nnunet/nnunetv2_runner.py

Line 932 in fbbb64b

def run(

[iffthomas]

Ah yes, currently I'm doing this with just running the train stage with this command prior to plan & preprocessing,
Thank you very much for taking your time answering my questions