{
  "id": 120555,
  "title": "A different way to ensemble: single digit models",
  "url": "/competitions/Kannada-MNIST/discussion/120555",
  "author_name": "",
  "post_date": "2019-12-06T23:15:06.716274500Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>hi Kagglers,</p>\n\n<p>Here's a fun kernel I created where instead of training one 10 digit classifier, I trained 10 single digit classifiers:\n<a href=\"https://www.kaggle.com/melissarajaram/model-ensembling-and-transfer-learning\">https://www.kaggle.com/melissarajaram/model-ensembling-and-transfer-learning</a></p>\n\n<p>While the ensembled accuracy wasn't as high as I might have hoped (0.978), the kernel contains:\n- example of how to do your own transfer learning by pretraining a 'backbone' model on 10 digits, and then using the pretrained convolution portion for single digit identification\n- an explanation of how to ensemble the output from single digit models into a 10 category prediction</p>\n\n<p>This kernel is beginner-friendly, as I explain every step of the process.</p>",
  "messages": [
    {
      "id": "689430",
      "postDate": "12/06/2019 23:15:06",
      "content": "<p>hi Kagglers,</p>\n\n<p>Here's a fun kernel I created where instead of training one 10 digit classifier, I trained 10 single digit classifiers:\n<a href=\"https://www.kaggle.com/melissarajaram/model-ensembling-and-transfer-learning\">https://www.kaggle.com/melissarajaram/model-ensembling-and-transfer-learning</a></p>\n\n<p>While the ensembled accuracy wasn't as high as I might have hoped (0.978), the kernel contains:\n- example of how to do your own transfer learning by pretraining a 'backbone' model on 10 digits, and then using the pretrained convolution portion for single digit identification\n- an explanation of how to ensemble the output from single digit models into a 10 category prediction</p>\n\n<p>This kernel is beginner-friendly, as I explain every step of the process.</p>",
      "rawMarkdown": "hi Kagglers,\n\nHere's a fun kernel I created where instead of training one 10 digit classifier, I trained 10 single digit classifiers:\nhttps://www.kaggle.com/melissarajaram/model-ensembling-and-transfer-learning\n\nWhile the ensembled accuracy wasn't as high as I might have hoped (0.978), the kernel contains:\n- example of how to do your own transfer learning by pretraining a 'backbone' model on 10 digits, and then using the pretrained convolution portion for single digit identification\n- an explanation of how to ensemble the output from single digit models into a 10 category prediction\n\nThis kernel is beginner-friendly, as I explain every step of the process.",
      "votes": null
    },
    {
      "id": "690003",
      "postDate": "12/07/2019 20:21:09",
      "content": "<p>Hi there! (Dataset creator here)\nThank you so very much for this contribution. It is quite amazing what this community has done with this dataset.\nI am curating a list of useful notebooks that I'll add to my main repo on github and was wondering if I had your consent to share yours?\nKind regards</p>",
      "rawMarkdown": "Hi there! (Dataset creator here)\nThank you so very much for this contribution. It is quite amazing what this community has done with this dataset.\nI am curating a list of useful notebooks that I'll add to my main repo on github and was wondering if I had your consent to share yours?\nKind regards",
      "votes": null
    },
    {
      "id": "690529",
      "postDate": "12/08/2019 17:59:10",
      "content": "<p>Wow, thanks! I'd be honored, and you have my consent to share the notebook.</p>",
      "rawMarkdown": "Wow, thanks! I'd be honored, and you have my consent to share the notebook.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 690003,
      "author_name": "higgstachyon",
      "author_url": "",
      "post_date": "12/07/2019 20:21:09",
      "content": "<p>Hi there! (Dataset creator here)\nThank you so very much for this contribution. It is quite amazing what this community has done with this dataset.\nI am curating a list of useful notebooks that I'll add to my main repo on github and was wondering if I had your consent to share yours?\nKind regards</p>",
      "votes": null,
      "replies": [
        {
          "id": 690529,
          "author_name": "melissarajaram",
          "author_url": "",
          "post_date": "12/08/2019 17:59:10",
          "content": "<p>Wow, thanks! I'd be honored, and you have my consent to share the notebook.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "689430": "hi Kagglers,\n\nHere's a fun kernel I created where instead of training one 10 digit classifier, I trained 10 single digit classifiers:\nhttps://www.kaggle.com/melissarajaram/model-ensembling-and-transfer-learning\n\nWhile the ensembled accuracy wasn't as high as I might have hoped (0.978), the kernel contains:\n- example of how to do your own transfer learning by pretraining a 'backbone' model on 10 digits, and then using the pretrained convolution portion for single digit identification\n- an explanation of how to ensemble the output from single digit models into a 10 category prediction\n\nThis kernel is beginner-friendly, as I explain every step of the process.",
    "690003": "Hi there! (Dataset creator here)\nThank you so very much for this contribution. It is quite amazing what this community has done with this dataset.\nI am curating a list of useful notebooks that I'll add to my main repo on github and was wondering if I had your consent to share yours?\nKind regards",
    "690529": "Wow, thanks! I'd be honored, and you have my consent to share the notebook."
  },
  "source": "meta"
}