search

Nischaydnk / RSNA-2023-1st-place-solution

Code and Documentation for the first place solution in 2023 Abdominal Trauma Detection Competition hosted by RSNA on Kaggle.
Apache License 2.0
47 stars 14 forks source link

Normalization for segmentation #3

Open powerLEO101 opened 10 months ago

powerLEO101 commented 10 months ago

Hello!

Thank you again for posting the solution and answering my previosu questions.

As I am trying to recreate and learn from your solution, I am confused by the normalizatoin of the prediction of 3d segmentatoin model.

When training, the normalization is done in the dice loss function through smp, but when predicting, no normalizatoin was done and the outputs were still in logits.

I understand that, if after sigmoid, it is OK to turn array to binary through final_outputs>0.5. But when there is no normalization, should the line be, instead, final_output > 0?

I wonder if this is a mistake or intentionally?

https://github.com/Nischaydnk/RSNA-2023-1st-place-solution/blob/77ed554f80c626195d0e18adcd8a739b2a96ef86/Datasets/make_info_data.py#L179

Best

powerLEO101 commented 10 months ago

And may I ask what is the final loss for 3d segmentation? Without TOTAL SEGMENTATOR data, I am stuck around 0.3.