{
  "id": 117828,
  "title": "Can we do transfer learning for Kannada MNIST (using original MINST or any other)?",
  "url": "/competitions/Kannada-MNIST/discussion/117828",
  "author_name": "",
  "post_date": "2019-11-18T05:37:08.932633900Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>If we can do transfer learning using original MINST, that will be great, but the original MINST was done for  English numbers(digits), but these are Kannada numbers(digits), there is a lot of difference between the two. Is transfer learning possible from any other appropriate pre-trained weights? What do you think?</p>",
  "messages": [
    {
      "id": "675469",
      "postDate": "11/18/2019 05:37:08",
      "content": "<p>If we can do transfer learning using original MINST, that will be great, but the original MINST was done for  English numbers(digits), but these are Kannada numbers(digits), there is a lot of difference between the two. Is transfer learning possible from any other appropriate pre-trained weights? What do you think?</p>",
      "rawMarkdown": "If we can do transfer learning using original MINST, that will be great, but the original MINST was done for  English numbers(digits), but these are Kannada numbers(digits), there is a lot of difference between the two. Is transfer learning possible from any other appropriate pre-trained weights? What do you think?",
      "votes": null
    },
    {
      "id": "675909",
      "postDate": "11/18/2019 18:53:54",
      "content": "<p>Yes you can try. Digit 0 looks like zero, 3 and 7 look like 2, 4 like 8 ... I suppose the MNIST-pretrained model will know some of the basic shapes.\nI'd guess if you unfreeze the layers after the flattening, it will rearrange the classes pretty good. But I haven't tried it. \nLet us know how it worked out for you.</p>",
      "rawMarkdown": "Yes you can try. Digit 0 looks like zero, 3 and 7 look like 2, 4 like 8 ... I suppose the MNIST-pretrained model will know some of the basic shapes.\nI'd guess if you unfreeze the layers after the flattening, it will rearrange the classes pretty good. But I haven't tried it. \nLet us know how it worked out for you.",
      "votes": null
    },
    {
      "id": "728787",
      "postDate": "01/25/2020 09:02:42",
      "content": "<p>Sorry for the late reply but it turns out that we can do it.</p>",
      "rawMarkdown": "Sorry for the late reply but it turns out that we can do it.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 675909,
      "author_name": "joatom",
      "author_url": "",
      "post_date": "11/18/2019 18:53:54",
      "content": "<p>Yes you can try. Digit 0 looks like zero, 3 and 7 look like 2, 4 like 8 ... I suppose the MNIST-pretrained model will know some of the basic shapes.\nI'd guess if you unfreeze the layers after the flattening, it will rearrange the classes pretty good. But I haven't tried it. \nLet us know how it worked out for you.</p>",
      "votes": null,
      "replies": [
        {
          "id": 728787,
          "author_name": "amitjslearn",
          "author_url": "",
          "post_date": "01/25/2020 09:02:42",
          "content": "<p>Sorry for the late reply but it turns out that we can do it.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "675469": "If we can do transfer learning using original MINST, that will be great, but the original MINST was done for  English numbers(digits), but these are Kannada numbers(digits), there is a lot of difference between the two. Is transfer learning possible from any other appropriate pre-trained weights? What do you think?",
    "675909": "Yes you can try. Digit 0 looks like zero, 3 and 7 look like 2, 4 like 8 ... I suppose the MNIST-pretrained model will know some of the basic shapes.\nI'd guess if you unfreeze the layers after the flattening, it will rearrange the classes pretty good. But I haven't tried it. \nLet us know how it worked out for you.",
    "728787": "Sorry for the late reply but it turns out that we can do it."
  },
  "source": "meta"
}