{
  "id": 53722,
  "title": "probe the LB test data",
  "url": "/competitions/data-science-bowl-2018/discussion/53722",
  "author_name": "",
  "post_date": "2018-04-04T08:41:05.550238700Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>the results:</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/308879/8931/probe.png\" alt=\"enter image description here\"></p>",
  "messages": [
    {
      "id": "308879",
      "postDate": "04/04/2018 08:41:05",
      "content": "<p>the results:</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/308879/8931/probe.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "the results:\n\n  ![enter image description here][1]\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/308879/8931/probe.png",
      "votes": null
    },
    {
      "id": "308888",
      "postDate": "04/04/2018 09:01:55",
      "content": "<p>From your result, I get two things:</p>\n\n<ol>\n<li><p>Gray image is more easy segmentation than color (due to color set so small)==&gt; Need to balancing them</p></li>\n<li><p>small number of cells in an image got better performance ==&gt; score based on image not object as you mentioned</p></li>\n</ol>\n\n<p>Question: How we can make a balancing training set? As I remember, you will train the network to obtain a checkpoint, apply training set using the checkpoint and measure the LB score. Sort the LB score in decreasing. Then duplicate these images have low LB. Is it correct?</p>",
      "rawMarkdown": "From your result, I get two things:\n\n1. Gray image is more easy segmentation than color (due to color set so small)==&gt; Need to balancing them\n\n2. small number of cells in an image got better performance ==&gt; score based on image not object as you mentioned\n\nQuestion: How we can make a balancing training set? As I remember, you will train the network to obtain a checkpoint, apply training set using the checkpoint and measure the LB score. Sort the LB score in decreasing. Then duplicate these images have low LB. Is it correct?",
      "votes": null
    },
    {
      "id": "308903",
      "postDate": "04/04/2018 09:47:18",
      "content": "<p>one can compare results of unet and mask-rcnn on validation set. Maybe one would would better than the other at different nuclei density. such information can help you to decide how to ensmble</p>",
      "rawMarkdown": "one can compare results of unet and mask-rcnn on validation set. Maybe one would would better than the other at different nuclei density. such information can help you to decide how to ensmble",
      "votes": null
    },
    {
      "id": "309362",
      "postDate": "04/05/2018 06:17:13",
      "content": "<p>How do I read that?</p>",
      "rawMarkdown": "How do I read that?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 308888,
      "author_name": "john1231983",
      "author_url": "",
      "post_date": "04/04/2018 09:01:55",
      "content": "<p>From your result, I get two things:</p>\n\n<ol>\n<li><p>Gray image is more easy segmentation than color (due to color set so small)==&gt; Need to balancing them</p></li>\n<li><p>small number of cells in an image got better performance ==&gt; score based on image not object as you mentioned</p></li>\n</ol>\n\n<p>Question: How we can make a balancing training set? As I remember, you will train the network to obtain a checkpoint, apply training set using the checkpoint and measure the LB score. Sort the LB score in decreasing. Then duplicate these images have low LB. Is it correct?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 308903,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "04/04/2018 09:47:18",
      "content": "<p>one can compare results of unet and mask-rcnn on validation set. Maybe one would would better than the other at different nuclei density. such information can help you to decide how to ensmble</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 309362,
      "author_name": "pointyointment",
      "author_url": "",
      "post_date": "04/05/2018 06:17:13",
      "content": "<p>How do I read that?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "308879": "the results:\n\n  ![enter image description here][1]\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/308879/8931/probe.png",
    "308888": "From your result, I get two things:\n\n1. Gray image is more easy segmentation than color (due to color set so small)==&gt; Need to balancing them\n\n2. small number of cells in an image got better performance ==&gt; score based on image not object as you mentioned\n\nQuestion: How we can make a balancing training set? As I remember, you will train the network to obtain a checkpoint, apply training set using the checkpoint and measure the LB score. Sort the LB score in decreasing. Then duplicate these images have low LB. Is it correct?",
    "308903": "one can compare results of unet and mask-rcnn on validation set. Maybe one would would better than the other at different nuclei density. such information can help you to decide how to ensmble",
    "309362": "How do I read that?"
  },
  "source": "meta"
}