How to run our own 3D model in Clara Train

Hello Everyone,

We are trying to have our own 3D model in Clara Train. Can you review my steps below and let me know whether this is correct and have I missed any steps?

  1. Create a model directory like annotation_spleen and put my model files (ckpt files). I say only ckpt files (,model.ckpt.index, model.ckpt.meta) because when I download models from NGC, I only see these files.

  2. later convert my model to model.fzn.pb and model.trt.pb using tlt-export command. Am I right? I mean convert it into trt-optimized form

  3. Create a config file (like spleen.json). But what I would like to know is all your pretrained_models have the same input_nodes (cond/Merge) and output_nodes (NV_output_label). From where do you get these details? Should we retain the same values in a config file for our model as well except the model path?

  4. Include the model details in Engine_pool.json file. But will the class (Dextr3dCroppedEngine) and module details (medical.aas.dextr3dcropped) remain the same for our model as well?

  5. Another question regarding your pretrained models is, how is the label name used? I mean in engine_pool.json, you have a label value ex: “Tumor core” for brain annotation model. But after launching MITK with AIAA, I wasn’t able to see this label name “Tumor core” in MITK drop down values for segmentation label column whereas I was able to find for “spleen”. Is it really any problem or it can be anything?

Request your help with the above questions

1 Like