Guidelines to deploy our model - Clara Train

Hello,

We followed your tutorial for spleen and trying to adapt it to our model. Here are few questions regarding that

  1. Our input data for skin lesion annotation is 2D jpeg images. Guess we don’t have to convert/resample our input data and output labels to isotropic format as shown in spleen tutorial. Is my understanding correct?

  2. As we are trying to have our own skin lesion segmentation model for clara train demo, we can copy our model in (checkpoint form with 3 files model.ckpt.data-00000-of-00001, model.ckpt.index, model.ckpt.meta) from docker host to container. we don’t really have to upload in NGC at this point of time. Am I right?

  3. So, I guess we may not need engine_pool.json and model_config.json (like spleen.json) for our model. Only when we have our model in NGC and trying to use MITK, we might need those config files. As our input data is 2d jpeg images, MITK interface isn’t possible and our model is also stored manually in container by us. I mean we have already trained locally at our site and like to generate inference based on our model. ,I understand that we may not need it. Can you correct me if I am wrong?

  4. Once my model files are in the container, I can then make use of tlt-export format to produce frozen and trt model files. Am I right?

  5. Similarly, I can use tlt-infer command to generate inference from our model and the end-end flow is complete. Am I right? But do I have to prepare these infer_config and eval_config files? Are they necessary?

  6. I see that all your existing models follow a prescribed format. I mean you have files like train_config, data_list which are in json format. Am I right to understand that it is only for easy comprehension and reusability and may not necessarily be applicable for our model at this time?

  7. If we would like to have our model uploaded to NGC, what is the prerequisite? So that, it can be used by everyone. I guess then I have to follow all the formatting options. Am I right?

  8. Is there anyway to have a model which can be used in both clara deploy and clara train?

1 Like