{
  "id": 189928,
  "title": "Model training: synchronous or saved/used?",
  "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/189928",
  "author_name": "learner76",
  "post_date": "2020-10-09T10:46:30.918000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Apologies in advance if this is too basic and has been addressed - I could not find good answers. Is it a good practice for your submitted kernel to plan for training your model from scratch while being evaluated? An alternative would be to train your model, save the results of the training and mostly use the submission kernel only to use saved model for predictions. For the former approach, I wonder if that will cross Notebook time limits and for the latter approach, I wonder if the model will be as trained as possible if we let it train on private test data (which it presumably will if the submitted kernel was training from scratch).</p>",
  "messages": [
    {
      "id": 1043932,
      "postDate": "2020-10-09T10:46:30.920Z",
      "content": "<p>Apologies in advance if this is too basic and has been addressed - I could not find good answers. Is it a good practice for your submitted kernel to plan for training your model from scratch while being evaluated? An alternative would be to train your model, save the results of the training and mostly use the submission kernel only to use saved model for predictions. For the former approach, I wonder if that will cross Notebook time limits and for the latter approach, I wonder if the model will be as trained as possible if we let it train on private test data (which it presumably will if the submitted kernel was training from scratch).</p>",
      "rawMarkdown": "Apologies in advance if this is too basic and has been addressed - I could not find good answers. Is it a good practice for your submitted kernel to plan for training your model from scratch while being evaluated? An alternative would be to train your model, save the results of the training and mostly use the submission kernel only to use saved model for predictions. For the former approach, I wonder if that will cross Notebook time limits and for the latter approach, I wonder if the model will be as trained as possible if we let it train on private test data (which it presumably will if the submitted kernel was training from scratch).",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1043932": "Apologies in advance if this is too basic and has been addressed - I could not find good answers. Is it a good practice for your submitted kernel to plan for training your model from scratch while being evaluated? An alternative would be to train your model, save the results of the training and mostly use the submission kernel only to use saved model for predictions. For the former approach, I wonder if that will cross Notebook time limits and for the latter approach, I wonder if the model will be as trained as possible if we let it train on private test data (which it presumably will if the submitted kernel was training from scratch)."
  }
}