{
  "id": 418313,
  "title": "my tricks to improve the scatter",
  "url": "/competitions/benetech-making-graphs-accessible/discussion/418313",
  "author_name": "",
  "post_date": "2023-06-20T00:48:11.577112400Z",
  "votes": 11,
  "comment_count": 3,
  "views": 0,
  "content": "<ol>\n<li>Add more scatter data(about 30,000 records) to train a Deplot model.  </li>\n<li>Train a cnn model to predict the x/y axis data point number of scatter.  A cnn model can get 92.6% accuracy ,better than 73%  of Deplot model.  We did a post-process after Deplot model prediction ,used the cnn model prediction to improve scatter prediction.  Before the cnn model , Deplot model got 0.47 (pb).  After the cnn model, we got 0.52(pb).  A big improvement!</li>\n</ol>",
  "messages": [
    {
      "id": "2309885",
      "postDate": "06/20/2023 00:48:11",
      "content": "<ol>\n<li>Add more scatter data(about 30,000 records) to train a Deplot model.  </li>\n<li>Train a cnn model to predict the x/y axis data point number of scatter.  A cnn model can get 92.6% accuracy ,better than 73%  of Deplot model.  We did a post-process after Deplot model prediction ,used the cnn model prediction to improve scatter prediction.  Before the cnn model , Deplot model got 0.47 (pb).  After the cnn model, we got 0.52(pb).  A big improvement!</li>\n</ol>",
      "rawMarkdown": "1.  Add more scatter data(about 30,000 records) to train a Deplot model.  \n2. Train a cnn model to predict the x/y axis data point number of scatter.  A cnn model can get 92.6% accuracy ,better than 73%  of Deplot model.  We did a post-process after Deplot model prediction ,used the cnn model prediction to improve scatter prediction.  Before the cnn model , Deplot model got 0.47 (pb).  After the cnn model, we got 0.52(pb).  A big improvement!",
      "votes": null
    },
    {
      "id": "2310099",
      "postDate": "06/20/2023 05:37:06",
      "content": "<p><a href=\"https://www.kaggle.com/qinhui1999\" target=\"_blank\">@qinhui1999</a> thanks for you ideas for improve code</p>",
      "rawMarkdown": "qinhui1999 thanks for you ideas for improve code",
      "votes": null
    },
    {
      "id": "2310561",
      "postDate": "06/20/2023 13:15:25",
      "content": "<p>It's an interesting idea to predict the number of points!<br>\nPlease share, what is the architecture of the network?</p>",
      "rawMarkdown": "It's an interesting idea to predict the number of points!\nPlease share, what is the architecture of the network?",
      "votes": null
    },
    {
      "id": "2310794",
      "postDate": "06/20/2023 15:55:13",
      "content": "<p>Interesting approach. I tried to implement a counting model for scatter plots, but could not get this to work.. </p>\n<p>With the CNN, did you treat it a classification or regression problem? Or did you use a segmentation based approach?</p>",
      "rawMarkdown": "Interesting approach. I tried to implement a counting model for scatter plots, but could not get this to work.. \n\nWith the CNN, did you treat it a classification or regression problem? Or did you use a segmentation based approach?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2310099,
      "author_name": "serangu",
      "author_url": "",
      "post_date": "06/20/2023 05:37:06",
      "content": "<p><a href=\"https://www.kaggle.com/qinhui1999\" target=\"_blank\">@qinhui1999</a> thanks for you ideas for improve code</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2310561,
      "author_name": "anuarsenk",
      "author_url": "",
      "post_date": "06/20/2023 13:15:25",
      "content": "<p>It's an interesting idea to predict the number of points!<br>\nPlease share, what is the architecture of the network?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2310794,
      "author_name": "brendanartley",
      "author_url": "",
      "post_date": "06/20/2023 15:55:13",
      "content": "<p>Interesting approach. I tried to implement a counting model for scatter plots, but could not get this to work.. </p>\n<p>With the CNN, did you treat it a classification or regression problem? Or did you use a segmentation based approach?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2309885": "1.  Add more scatter data(about 30,000 records) to train a Deplot model.  \n2. Train a cnn model to predict the x/y axis data point number of scatter.  A cnn model can get 92.6% accuracy ,better than 73%  of Deplot model.  We did a post-process after Deplot model prediction ,used the cnn model prediction to improve scatter prediction.  Before the cnn model , Deplot model got 0.47 (pb).  After the cnn model, we got 0.52(pb).  A big improvement!",
    "2310099": "qinhui1999 thanks for you ideas for improve code",
    "2310561": "It's an interesting idea to predict the number of points!\nPlease share, what is the architecture of the network?",
    "2310794": "Interesting approach. I tried to implement a counting model for scatter plots, but could not get this to work.. \n\nWith the CNN, did you treat it a classification or regression problem? Or did you use a segmentation based approach?"
  },
  "source": "meta"
}