The seg-pretraining issues about alveolar_canal

[Key-lei]

Thank you for your answer! Currently I changed the split in the dataset of the inference function under the generation.py file in the experiments folder from splits=['train', 'val', 'test'] to splits='synthetic', ran gen-inference I don't know if this principle is causing my problem, but when I run the splits under the original inference, only 91 dense annotations of synthetic are generated. Thanks again for your answer, and I hope to hear from you again if I have any follow-up questions. have a nice life!

[LucaLumetti]

Hi @Key-lei,
looks like line 204 of experiments.py is empty. I would assume the error is thrown by line 216 of the same file.
If all the data inside the generated dataset is empty, it probably means that something during the Deep Label Expansion (or Generation) phase has gone wrong. If the inference has been performed using checkpoints calculated by you, please check the metrics you obtained in this phase and check that the network didn't overfit on predicting always empty volumes. As I've noticed that you have added/deleted some lines of code, be sure you didn't break anything.

synthetic_loader should be a set of SubjectImage containing the CBCT data of volume contained in the synthetic split together with their generated dense label (as they do not have dense annotations produced by medical experts).

[Key-lei]

Thank you for your reply! Also, I noticed that after executing gen-training.yaml and executing gen-inference.yaml some of the patients under the sparse file generated generated.npy files but still most of the patients did not have generated.npy is this the cause of the error? I kept all the code the same as you provided and executed gen-training.yaml and gen-inference.yaml in order and it worked, but when I executed seg-pretraining.yaml I got an error.

[LucaLumetti]

I noticed that after executing gen-training.yaml and executing gen-inference.yaml some of the patients under the sparse file generated generated.npy files but still most of the patients did not have generated.npy

All the patients in the synthetic split must have a generated.npy file before the execution of seg-pretraining.py. Patients in the other three sets (val, train, test) do not require a generated.npy file as they all have dense annotations produced by medical experts (thus better than the one the networks generate).

Please report me the logs generated during the execution of gen-training.yaml and gen-inference.yaml, as well as the final metrics obtained during the gen-training.yaml

[puppy2000]

Thank you for your answer! Currently I changed the split in the dataset of the inference function under the generation.py file in the experiments folder from splits=['train', 'val', 'test'] to splits='synthetic', ran gen-inference I don't know if this principle is causing my problem, but when I run the splits under the original inference, only 91 dense annotations of synthetic are generated. Thanks again for your answer, and I hope to hear from you again if I have any follow-up questions. have a nice life!

Are you still doing well in the subsequent training process?