{
  "id": 306573,
  "title": "Can I use an imported model to make predictions in a submission entry?",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/306573",
  "author_name": "",
  "post_date": "2022-02-10T01:15:51.593794100Z",
  "votes": 3,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Say I train a model and save it to the working directory.  Then if I start a new notebook and import the trained model can I just use that for predictions? </p>\n<p>Doing this seems a way to save a lot of time but is it allowed for competition submissions? </p>\n<p>Or, when submitting, is something happening behind the scenes?  For example, is the original notebook that creates and trains the model re-run anyway? So in this case there would be no time advantage to importing an existing model?</p>",
  "messages": [
    {
      "id": "1683672",
      "postDate": "02/10/2022 01:15:51",
      "content": "<p>Say I train a model and save it to the working directory.  Then if I start a new notebook and import the trained model can I just use that for predictions? </p>\n<p>Doing this seems a way to save a lot of time but is it allowed for competition submissions? </p>\n<p>Or, when submitting, is something happening behind the scenes?  For example, is the original notebook that creates and trains the model re-run anyway? So in this case there would be no time advantage to importing an existing model?</p>",
      "rawMarkdown": "Say I train a model and save it to the working directory.  Then if I start a new notebook and import the trained model can I just use that for predictions? \n\nDoing this seems a way to save a lot of time but is it allowed for competition submissions? \n\nOr, when submitting, is something happening behind the scenes?  For example, is the original notebook that creates and trains the model re-run anyway? So in this case there would be no time advantage to importing an existing model?",
      "votes": null
    },
    {
      "id": "1683724",
      "postDate": "02/10/2022 02:26:34",
      "content": "<p>Yes, you can train your model in one notebook, load it as a dataset and import it into a different notebook for inference.</p>",
      "rawMarkdown": "Yes, you can train your model in one notebook, load it as a dataset and import it into a different notebook for inference.",
      "votes": null
    },
    {
      "id": "1683751",
      "postDate": "02/10/2022 03:12:42",
      "content": "<p>Thank you.</p>",
      "rawMarkdown": "Thank you.",
      "votes": null
    },
    {
      "id": "1683877",
      "postDate": "02/10/2022 05:35:03",
      "content": "<p>As far as I know, everyone uses trained model imported into the submission notebook. Most high res model cannot be trained in kaggle due to the GPU's memory limitation.</p>\n<p>The original training notebook is not reran. It's not involved. Only your submission notebook is ran.</p>",
      "rawMarkdown": "As far as I know, everyone uses trained model imported into the submission notebook. Most high res model cannot be trained in kaggle due to the GPU's memory limitation.\n\nThe original training notebook is not reran. It's not involved. Only your submission notebook is ran.",
      "votes": null
    },
    {
      "id": "1684808",
      "postDate": "02/10/2022 18:21:48",
      "content": "<p>yes, the 9hrs are just for inference time, meaning that it doesnt include the time needed to train</p>",
      "rawMarkdown": "yes, the 9hrs are just for inference time, meaning that it doesnt include the time needed to train",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1683724,
      "author_name": "davidbroberts",
      "author_url": "",
      "post_date": "02/10/2022 02:26:34",
      "content": "<p>Yes, you can train your model in one notebook, load it as a dataset and import it into a different notebook for inference.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1683751,
      "author_name": "melaniebrennan",
      "author_url": "",
      "post_date": "02/10/2022 03:12:42",
      "content": "<p>Thank you.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1683877,
      "author_name": "liminchen1",
      "author_url": "",
      "post_date": "02/10/2022 05:35:03",
      "content": "<p>As far as I know, everyone uses trained model imported into the submission notebook. Most high res model cannot be trained in kaggle due to the GPU's memory limitation.</p>\n<p>The original training notebook is not reran. It's not involved. Only your submission notebook is ran.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1684808,
      "author_name": "kennyxie",
      "author_url": "",
      "post_date": "02/10/2022 18:21:48",
      "content": "<p>yes, the 9hrs are just for inference time, meaning that it doesnt include the time needed to train</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1683672": "Say I train a model and save it to the working directory.  Then if I start a new notebook and import the trained model can I just use that for predictions? \n\nDoing this seems a way to save a lot of time but is it allowed for competition submissions? \n\nOr, when submitting, is something happening behind the scenes?  For example, is the original notebook that creates and trains the model re-run anyway? So in this case there would be no time advantage to importing an existing model?",
    "1683724": "Yes, you can train your model in one notebook, load it as a dataset and import it into a different notebook for inference.",
    "1683751": "Thank you.",
    "1683877": "As far as I know, everyone uses trained model imported into the submission notebook. Most high res model cannot be trained in kaggle due to the GPU's memory limitation.\n\nThe original training notebook is not reran. It's not involved. Only your submission notebook is ran.",
    "1684808": "yes, the 9hrs are just for inference time, meaning that it doesnt include the time needed to train"
  },
  "source": "meta"
}