{
  "id": 199701,
  "title": "How to work locally in code competitions with private test set?",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/199701",
  "author_name": "",
  "post_date": "2020-11-26T22:58:49.331193100Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I am rather new to kaggle and only effectively participated in one competition, where you simply had to upload your predictions as a CSV-file. Therefore, you could work completely locally.</p>\n<p>This doesn't seem to be possible anymore for code competitions with private test sets if my understanding is correct. Thus, do you now work only in kaggle notebooks? Or do you first work and validate locally, before basically copying the trained model and the inference script to a new kaggle notebook? This seems rather tedious… Also ensembling becomes significantly more tedious now, or?</p>\n<p>Concluding: Any tips on how to work somewhat locally in these kinds of competitions are very welcome. Unfortunately, I didn't find any helpful pages using google. Thanks in advance!</p>",
  "messages": [
    {
      "id": "1092510",
      "postDate": "11/26/2020 22:58:49",
      "content": "<p>I am rather new to kaggle and only effectively participated in one competition, where you simply had to upload your predictions as a CSV-file. Therefore, you could work completely locally.</p>\n<p>This doesn't seem to be possible anymore for code competitions with private test sets if my understanding is correct. Thus, do you now work only in kaggle notebooks? Or do you first work and validate locally, before basically copying the trained model and the inference script to a new kaggle notebook? This seems rather tedious… Also ensembling becomes significantly more tedious now, or?</p>\n<p>Concluding: Any tips on how to work somewhat locally in these kinds of competitions are very welcome. Unfortunately, I didn't find any helpful pages using google. Thanks in advance!</p>",
      "rawMarkdown": "I am rather new to kaggle and only effectively participated in one competition, where you simply had to upload your predictions as a CSV-file. Therefore, you could work completely locally.\n\nThis doesn't seem to be possible anymore for code competitions with private test sets if my understanding is correct. Thus, do you now work only in kaggle notebooks? Or do you first work and validate locally, before basically copying the trained model and the inference script to a new kaggle notebook? This seems rather tedious... Also ensembling becomes significantly more tedious now, or?\n\nConcluding: Any tips on how to work somewhat locally in these kinds of competitions are very welcome. Unfortunately, I didn't find any helpful pages using google. Thanks in advance!",
      "votes": null
    },
    {
      "id": "1092542",
      "postDate": "11/27/2020 00:26:30",
      "content": "<p>You have it basically right.</p>\n<ol>\n<li>Train models either locally, in Kaggle Notebooks or using any cloud provider.</li>\n<li>Load model and weights into a dataset</li>\n<li>Use Kaggle Notebook to do inference using the model and weights in your dataset.</li>\n<li>Ensembling can only be done \"on-the-fly\" within a notebook as you inference each model. You do not have the ability to inference 20 models, save the submission files and then ensemble them (unless you can get it all done within one notebook.</li>\n</ol>",
      "rawMarkdown": "You have it basically right.\n\n1. Train models either locally, in Kaggle Notebooks or using any cloud provider.\n2. Load model and weights into a dataset\n3. Use Kaggle Notebook to do inference using the model and weights in your dataset.\n4. Ensembling can only be done \"on-the-fly\" within a notebook as you inference each model. You do not have the ability to inference 20 models, save the submission files and then ensemble them (unless you can get it all done within one notebook.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1092542,
      "author_name": "richardepstein",
      "author_url": "",
      "post_date": "11/27/2020 00:26:30",
      "content": "<p>You have it basically right.</p>\n<ol>\n<li>Train models either locally, in Kaggle Notebooks or using any cloud provider.</li>\n<li>Load model and weights into a dataset</li>\n<li>Use Kaggle Notebook to do inference using the model and weights in your dataset.</li>\n<li>Ensembling can only be done \"on-the-fly\" within a notebook as you inference each model. You do not have the ability to inference 20 models, save the submission files and then ensemble them (unless you can get it all done within one notebook.</li>\n</ol>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1092510": "I am rather new to kaggle and only effectively participated in one competition, where you simply had to upload your predictions as a CSV-file. Therefore, you could work completely locally.\n\nThis doesn't seem to be possible anymore for code competitions with private test sets if my understanding is correct. Thus, do you now work only in kaggle notebooks? Or do you first work and validate locally, before basically copying the trained model and the inference script to a new kaggle notebook? This seems rather tedious... Also ensembling becomes significantly more tedious now, or?\n\nConcluding: Any tips on how to work somewhat locally in these kinds of competitions are very welcome. Unfortunately, I didn't find any helpful pages using google. Thanks in advance!",
    "1092542": "You have it basically right.\n\n1. Train models either locally, in Kaggle Notebooks or using any cloud provider.\n2. Load model and weights into a dataset\n3. Use Kaggle Notebook to do inference using the model and weights in your dataset.\n4. Ensembling can only be done \"on-the-fly\" within a notebook as you inference each model. You do not have the ability to inference 20 models, save the submission files and then ensemble them (unless you can get it all done within one notebook."
  },
  "source": "meta"
}