{
  "id": 286058,
  "title": "Evaluating 26 models to split the problem.",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/286058",
  "author_name": "",
  "post_date": "2021-11-07T18:28:27.400649800Z",
  "votes": 8,
  "comment_count": 2,
  "views": 0,
  "content": "<p>As a beginner, I'm not even sure I got the problem we are facing.</p>\n<p>What I understood so far is that we are getting images with only one type of cells (shsy5y , cortical neurons and astrocytes) and predict instance segmentations on these.</p>\n<p>If this is right, maybe splitting the problem into three different problems could make the thing easier. It seems that masks are a bit different depending of the type of cells we find, therefore selecting the type of cells in the first place and then apply three different models to find the masks could be a good approach.</p>\n<p>This is why I tried to create a notebook to find the better and faster models that can solve this first stage.</p>\n<p>I shared a notebook to evaluate 26 different pre-trained models to check if it is possible to split the input images into three different models. </p>\n<p>My questions are:</p>\n<ul>\n<li>Does this approach make any sense?</li>\n<li>It's true my believe that every image has only one type of cells?</li>\n</ul>\n<p>This is the notebook: <a href=\"https://www.kaggle.com/kriyeng/sartorious-eval-best-model-to-split-the-problem\" target=\"_blank\">Sartorious Eval best model to split the problem</a> </p>\n<p>Commentaries are very welcome!</p>",
  "messages": [
    {
      "id": "1574639",
      "postDate": "11/07/2021 18:28:27",
      "content": "<p>As a beginner, I'm not even sure I got the problem we are facing.</p>\n<p>What I understood so far is that we are getting images with only one type of cells (shsy5y , cortical neurons and astrocytes) and predict instance segmentations on these.</p>\n<p>If this is right, maybe splitting the problem into three different problems could make the thing easier. It seems that masks are a bit different depending of the type of cells we find, therefore selecting the type of cells in the first place and then apply three different models to find the masks could be a good approach.</p>\n<p>This is why I tried to create a notebook to find the better and faster models that can solve this first stage.</p>\n<p>I shared a notebook to evaluate 26 different pre-trained models to check if it is possible to split the input images into three different models. </p>\n<p>My questions are:</p>\n<ul>\n<li>Does this approach make any sense?</li>\n<li>It's true my believe that every image has only one type of cells?</li>\n</ul>\n<p>This is the notebook: <a href=\"https://www.kaggle.com/kriyeng/sartorious-eval-best-model-to-split-the-problem\" target=\"_blank\">Sartorious Eval best model to split the problem</a> </p>\n<p>Commentaries are very welcome!</p>",
      "rawMarkdown": "As a beginner, I'm not even sure I got the problem we are facing.\n\nWhat I understood so far is that we are getting images with only one type of cells (shsy5y , cortical neurons and astrocytes) and predict instance segmentations on these.\n\nIf this is right, maybe splitting the problem into three different problems could make the thing easier. It seems that masks are a bit different depending of the type of cells we find, therefore selecting the type of cells in the first place and then apply three different models to find the masks could be a good approach.\n\nThis is why I tried to create a notebook to find the better and faster models that can solve this first stage.\n\nI shared a notebook to evaluate 26 different pre-trained models to check if it is possible to split the input images into three different models. \n\nMy questions are:\n- Does this approach make any sense?\n- It's true my believe that every image has only one type of cells?\n\nThis is the notebook: [Sartorious Eval best model to split the problem](https://www.kaggle.com/kriyeng/sartorious-eval-best-model-to-split-the-problem) \n\nCommentaries are very welcome!",
      "votes": null
    },
    {
      "id": "1575303",
      "postDate": "11/08/2021 10:05:55",
      "content": "<blockquote>\n  <ul>\n  <li>Does this approach make any sense?</li>\n  </ul>\n</blockquote>\n<p>It's unclear whether splitting models can help with the score. This was explored by the contest organizers in <a href=\"https://www.nature.com/articles/s41592-021-01249-6\" target=\"_blank\">LIVECell paper</a> there they had 8 types of cells and got better results by training all of them together rather than each separately. However I did see a post (can't find it now) from someone who got an improvement in this competition.<br>\nFrom a practical point of view my opinion is that it's not worth it as the first thing to work on. You still need to do segmentation either way. And doing a single type isn't any simpler in terms of code you need to write, so you are just adding work for yourself.</p>\n<blockquote>\n  <ul>\n  <li>It's true my believe that every image has only one type of cells?</li>\n  </ul>\n</blockquote>\n<p>I have verified that in the train set we were given every image has only a single type of cells annotated. I haven't seen any word about the test set, but I work under assumption that it's the same.</p>",
      "rawMarkdown": "> - Does this approach make any sense?\n\nIt's unclear whether splitting models can help with the score. This was explored by the contest organizers in [LIVECell paper](https://www.nature.com/articles/s41592-021-01249-6) there they had 8 types of cells and got better results by training all of them together rather than each separately. However I did see a post (can't find it now) from someone who got an improvement in this competition.\nFrom a practical point of view my opinion is that it's not worth it as the first thing to work on. You still need to do segmentation either way. And doing a single type isn't any simpler in terms of code you need to write, so you are just adding work for yourself.\n\n> - It's true my believe that every image has only one type of cells?\n\nI have verified that in the train set we were given every image has only a single type of cells annotated. I haven't seen any word about the test set, but I work under assumption that it's the same.",
      "votes": null
    },
    {
      "id": "1575371",
      "postDate": "11/08/2021 11:00:35",
      "content": "<p>Thank you <a href=\"https://www.kaggle.com/slawekbiel\" target=\"_blank\">@slawekbiel</a> for your detailed explanations and for your great work on the notebooks you shared.<br>\nGood point to start creating the whole solution and think later to split or not.<br>\nThanks!</p>",
      "rawMarkdown": "Thank you @slawekbiel for your detailed explanations and for your great work on the notebooks you shared.\nGood point to start creating the whole solution and think later to split or not.\nThanks!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1575303,
      "author_name": "slawekbiel",
      "author_url": "",
      "post_date": "11/08/2021 10:05:55",
      "content": "<blockquote>\n  <ul>\n  <li>Does this approach make any sense?</li>\n  </ul>\n</blockquote>\n<p>It's unclear whether splitting models can help with the score. This was explored by the contest organizers in <a href=\"https://www.nature.com/articles/s41592-021-01249-6\" target=\"_blank\">LIVECell paper</a> there they had 8 types of cells and got better results by training all of them together rather than each separately. However I did see a post (can't find it now) from someone who got an improvement in this competition.<br>\nFrom a practical point of view my opinion is that it's not worth it as the first thing to work on. You still need to do segmentation either way. And doing a single type isn't any simpler in terms of code you need to write, so you are just adding work for yourself.</p>\n<blockquote>\n  <ul>\n  <li>It's true my believe that every image has only one type of cells?</li>\n  </ul>\n</blockquote>\n<p>I have verified that in the train set we were given every image has only a single type of cells annotated. I haven't seen any word about the test set, but I work under assumption that it's the same.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1575371,
          "author_name": "kriyeng",
          "author_url": "",
          "post_date": "11/08/2021 11:00:35",
          "content": "<p>Thank you <a href=\"https://www.kaggle.com/slawekbiel\" target=\"_blank\">@slawekbiel</a> for your detailed explanations and for your great work on the notebooks you shared.<br>\nGood point to start creating the whole solution and think later to split or not.<br>\nThanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1574639": "As a beginner, I'm not even sure I got the problem we are facing.\n\nWhat I understood so far is that we are getting images with only one type of cells (shsy5y , cortical neurons and astrocytes) and predict instance segmentations on these.\n\nIf this is right, maybe splitting the problem into three different problems could make the thing easier. It seems that masks are a bit different depending of the type of cells we find, therefore selecting the type of cells in the first place and then apply three different models to find the masks could be a good approach.\n\nThis is why I tried to create a notebook to find the better and faster models that can solve this first stage.\n\nI shared a notebook to evaluate 26 different pre-trained models to check if it is possible to split the input images into three different models. \n\nMy questions are:\n- Does this approach make any sense?\n- It's true my believe that every image has only one type of cells?\n\nThis is the notebook: [Sartorious Eval best model to split the problem](https://www.kaggle.com/kriyeng/sartorious-eval-best-model-to-split-the-problem) \n\nCommentaries are very welcome!",
    "1575303": "> - Does this approach make any sense?\n\nIt's unclear whether splitting models can help with the score. This was explored by the contest organizers in [LIVECell paper](https://www.nature.com/articles/s41592-021-01249-6) there they had 8 types of cells and got better results by training all of them together rather than each separately. However I did see a post (can't find it now) from someone who got an improvement in this competition.\nFrom a practical point of view my opinion is that it's not worth it as the first thing to work on. You still need to do segmentation either way. And doing a single type isn't any simpler in terms of code you need to write, so you are just adding work for yourself.\n\n> - It's true my believe that every image has only one type of cells?\n\nI have verified that in the train set we were given every image has only a single type of cells annotated. I haven't seen any word about the test set, but I work under assumption that it's the same.",
    "1575371": "Thank you @slawekbiel for your detailed explanations and for your great work on the notebooks you shared.\nGood point to start creating the whole solution and think later to split or not.\nThanks!"
  },
  "source": "meta"
}