{
  "id": 277605,
  "title": "Digit Rcognizer 100% with ArcFace",
  "url": "/competitions/digit-recognizer/discussion/277605",
  "author_name": "Nguyễn Sỹ Đức",
  "post_date": "2021-10-10T09:35:25.469000",
  "votes": 2,
  "comment_count": 4,
  "views": null,
  "content": "<p>I've just found a way to get 100%. I used the ideas in ArcFace to train my base model, which outputs a 512-D vector. Then, I performed similarity searching to predict digits.<br>\nHere is my github. You can find some useful information there.<br>\n<a href=\"https://github.com/tonyD1999/Handwritten-Digit-Recognition-with-ARCFace\" target=\"_blank\">https://github.com/tonyD1999/Handwritten-Digit-Recognition-with-ARCFace</a></p>",
  "messages": [
    {
      "id": 1570579,
      "postDate": "2021-11-04T08:43:42.143Z",
      "content": "<p>I think nearly any CNN would reach 1.0 if you use MNIST dataset in training, because it includes all the samples from the testing set, so you are free to overfit.</p>",
      "rawMarkdown": "I think nearly any CNN would reach 1.0 if you use MNIST dataset in training, because it includes all the samples from the testing set, so you are free to overfit.",
      "votes": 4
    },
    {
      "id": 1540239,
      "postDate": "2021-10-10T09:35:25.470Z",
      "content": "<p>I've just found a way to get 100%. I used the ideas in ArcFace to train my base model, which outputs a 512-D vector. Then, I performed similarity searching to predict digits.<br>\nHere is my github. You can find some useful information there.<br>\n<a href=\"https://github.com/tonyD1999/Handwritten-Digit-Recognition-with-ARCFace\" target=\"_blank\">https://github.com/tonyD1999/Handwritten-Digit-Recognition-with-ARCFace</a></p>",
      "rawMarkdown": "I've just found a way to get 100%. I used the ideas in ArcFace to train my base model, which outputs a 512-D vector. Then, I performed similarity searching to predict digits.\nHere is my github. You can find some useful information there.\nhttps://github.com/tonyD1999/Handwritten-Digit-Recognition-with-ARCFace",
      "votes": 2
    },
    {
      "id": 1750138,
      "postDate": "2022-04-09T11:05:36.593Z",
      "content": "<p>Very interesting viewpoint, nice work!</p>",
      "rawMarkdown": "Very interesting viewpoint, nice work!"
    },
    {
      "id": 1641157,
      "postDate": "2022-01-07T06:37:35.427Z",
      "content": "<p>Thanks for your advice, I'm going to learn and try it.</p>",
      "rawMarkdown": "Thanks for your advice, I'm going to learn and try it."
    },
    {
      "id": 1598534,
      "postDate": "2021-11-28T14:57:12.270Z",
      "content": "<p>Thanks for this new topic to learn and explore!</p>",
      "rawMarkdown": "Thanks for this new topic to learn and explore!"
    }
  ],
  "comments": [
    {
      "id": 1570579,
      "author_name": "Sergey Bochenkov",
      "author_url": "",
      "post_date": "2021-11-04T08:43:42.143000",
      "content": "<p>I think nearly any CNN would reach 1.0 if you use MNIST dataset in training, because it includes all the samples from the testing set, so you are free to overfit.</p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 1750138,
      "author_name": "Joanne Mon",
      "author_url": "",
      "post_date": "2022-04-09T11:05:36.593000",
      "content": "<p>Very interesting viewpoint, nice work!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1641157,
      "author_name": "Minglei",
      "author_url": "",
      "post_date": "2022-01-07T06:37:35.427000",
      "content": "<p>Thanks for your advice, I'm going to learn and try it.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1598534,
      "author_name": "Gururaj Bhat",
      "author_url": "",
      "post_date": "2021-11-28T14:57:12.270000",
      "content": "<p>Thanks for this new topic to learn and explore!</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1570579": "I think nearly any CNN would reach 1.0 if you use MNIST dataset in training, because it includes all the samples from the testing set, so you are free to overfit.",
    "1540239": "I've just found a way to get 100%. I used the ideas in ArcFace to train my base model, which outputs a 512-D vector. Then, I performed similarity searching to predict digits.\nHere is my github. You can find some useful information there.\nhttps://github.com/tonyD1999/Handwritten-Digit-Recognition-with-ARCFace",
    "1750138": "Very interesting viewpoint, nice work!",
    "1641157": "Thanks for your advice, I'm going to learn and try it.",
    "1598534": "Thanks for this new topic to learn and explore!"
  }
}