Training nnUNetv2 Key Error: 'conv_kernel_sizes'

[milutinovicmarija]

I encountered the following error while trying to train the model nnUNetv2_train DATASET_ID 3d_fullres 0:

../nunetv2/utilities/plans_handling/plans_handler.py", line 116, in conv_kernel_sizes
return self.configuration['conv_kernel_sizes']
KeyError: 'conv_kernel_sizes'

nnUNetv2_plan_and_preprocess -d DATASET_ID was successfuly completed and nnUNetPlans.json file is present in the nnUNet_preprocessed folder containing different configurations.

There is a key kernel_sizes in nnUnetPlans.json file, but there is no key conv_kernel_sizes.

The following keys are present in the configuration file that is being processed by plans_handler.py:

data_identifier
preprocessor_name
batch_size
patch_size
median_image_size_in_voxels
spacing
normalization_schemes
use_mask_for_norm
resampling_fn_data
resampling_fn_seg
resampling_fn_data_kwargs
resampling_fn_seg_kwargs
resampling_fn_probabilities
resampling_fn_probabilities_kwargs
architecture
batch_dice

[ken011528]

It maybe because of the version of nnunetv2.

[FabianIsensee]

Please update to the latest version and rerun nnUNetv2_plan_and_preprocess. Then try again

[jeb2112]

I had the same error for a completely different yet similar reason, but in a trickier sort of way. It didn't occur during training, but in prediction after training was complete. The situation is that I had been running training on a AWS cloud instance, and was downloading the result to run predictions on my local computer. While setting up the cloud instance, I was installing nnUNetv2 fresh, but on the local computer I was using nnUNetv2 installed almost a year ago. After the first training run, I was able to run predictions locally with the older copy of the nnUNetv2 code from last year. But after the second training run, I was getting the above error. The sole change between the two training runs is that I ran the first one with fold=all, and the second one with 5-fold cross-validation. After I created a new local env with a freshly installed nnUNetv2 on my computer, prediction on the 5-fold validation trained models now worked.