{
  "id": 212493,
  "title": "How to improve the accuracy of the model ? any other suggestion ?",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/212493",
  "author_name": "",
  "post_date": "2021-01-19T05:01:48.367655200Z",
  "votes": 8,
  "comment_count": 18,
  "views": 0,
  "content": "<ul>\n<li>Collect more data.</li>\n<li>Collect a more diverse training set.</li>\n<li>Train algorithm longer with gradient descent.</li>\n<li>Try different optimization algorithm </li>\n<li>Try a bigger or a smaller network.</li>\n<li>Add L2 regularization or dropout.</li>\n<li>Change network architecture (activation functions, hidden units, etc.)</li>\n</ul>\n<p>from <a href=\"https://www.kaggle.com/saurabhshahane\" target=\"_blank\">@saurabhshahane</a> </p>\n<ul>\n<li>Feeding good quality of data</li>\n<li>Reducing the noise</li>\n</ul>\n<p>from <a href=\"https://www.kaggle.com/tamilselvanmoorthy\" target=\"_blank\">@tamilselvanmoorthy</a> </p>\n<ul>\n<li>use tampered loss functions as there are many mislabelling.</li>\n</ul>",
  "messages": [
    {
      "id": "1159183",
      "postDate": "01/19/2021 05:01:48",
      "content": "<ul>\n<li>Collect more data.</li>\n<li>Collect a more diverse training set.</li>\n<li>Train algorithm longer with gradient descent.</li>\n<li>Try different optimization algorithm </li>\n<li>Try a bigger or a smaller network.</li>\n<li>Add L2 regularization or dropout.</li>\n<li>Change network architecture (activation functions, hidden units, etc.)</li>\n</ul>\n<p>from <a href=\"https://www.kaggle.com/saurabhshahane\" target=\"_blank\">@saurabhshahane</a> </p>\n<ul>\n<li>Feeding good quality of data</li>\n<li>Reducing the noise</li>\n</ul>\n<p>from <a href=\"https://www.kaggle.com/tamilselvanmoorthy\" target=\"_blank\">@tamilselvanmoorthy</a> </p>\n<ul>\n<li>use tampered loss functions as there are many mislabelling.</li>\n</ul>",
      "rawMarkdown": "Collect more data.\n- Collect a more diverse training set.\n- Train algorithm longer with gradient descent.\n- Try different optimization algorithm \n- Try a bigger or a smaller network.\n- Add L2 regularization or dropout.\n- Change network architecture (activation functions, hidden units, etc.)\n\nfrom @saurabhshahane \n- Feeding good quality of data\n- Reducing the noise\n\nfrom @tamilselvanmoorthy \n- use tampered loss functions as there are many mislabelling.",
      "votes": null
    },
    {
      "id": "1159190",
      "postDate": "01/19/2021 05:09:55",
      "content": "<p>I am doing multiple model ensemble, it may not improve the LB score, but the robustness of the model may be useful for private data sets…</p>",
      "rawMarkdown": "I am doing multiple model ensemble, it may not improve the LB score, but the robustness of the model may be useful for private data sets...",
      "votes": null
    },
    {
      "id": "1159225",
      "postDate": "01/19/2021 05:38:41",
      "content": "<p>thanks <a href=\"https://www.kaggle.com/zhangeng\" target=\"_blank\">@zhangeng</a>, can you cite some resources?</p>",
      "rawMarkdown": "thanks @zhangeng, can you cite some resources?",
      "votes": null
    },
    {
      "id": "1159235",
      "postDate": "01/19/2021 05:45:34",
      "content": "<p><a href=\"https://www.kaggle.com/zhangeng\" target=\"_blank\">@zhangeng</a> I've also tried ensembling my models but didn't see any improvements on the lb, only on cv.</p>",
      "rawMarkdown": "zhangeng I've also tried ensembling my models but didn't see any improvements on the lb, only on cv.",
      "votes": null
    },
    {
      "id": "1159256",
      "postDate": "01/19/2021 06:17:58",
      "content": "<p>For example, I am using vit, B4, and resnet networks for ensemble, and the model is being trained…</p>",
      "rawMarkdown": "For example, I am using vit, B4, and resnet networks for ensemble, and the model is being trained...",
      "votes": null
    },
    {
      "id": "1159290",
      "postDate": "01/19/2021 07:12:17",
      "content": "<ol>\n<li>Feeding good quality of data</li>\n<li>Reducing the noise</li>\n</ol>",
      "rawMarkdown": "1. Feeding good quality of data\n2. Reducing the noise",
      "votes": null
    },
    {
      "id": "1159293",
      "postDate": "01/19/2021 07:14:00",
      "content": "<p>thanks <a href=\"https://www.kaggle.com/saurabhshahane\" target=\"_blank\">@saurabhshahane</a> 👍</p>",
      "rawMarkdown": "thanks @saurabhshahane 👍",
      "votes": null
    },
    {
      "id": "1159304",
      "postDate": "01/19/2021 07:20:30",
      "content": "<p>The problem is that the test set is also very noisy…</p>",
      "rawMarkdown": "The problem is that the test set is also very noisy...",
      "votes": null
    },
    {
      "id": "1159342",
      "postDate": "01/19/2021 07:45:07",
      "content": "<p>I agree with you <a href=\"https://www.kaggle.com/zhangeng\" target=\"_blank\">@zhangeng</a> . I'm saying in general.</p>",
      "rawMarkdown": "I agree with you @zhangeng . I'm saying in general.",
      "votes": null
    },
    {
      "id": "1159718",
      "postDate": "01/19/2021 12:27:20",
      "content": "<ol>\n<li>try different augmentation techniques. </li>\n<li>efficient net gave good accuracy. </li>\n<li>use tampered loss functions as there are many mislabelling. </li>\n</ol>",
      "rawMarkdown": "1. try different augmentation techniques. \n2. efficient net gave good accuracy. \n3. use tampered loss functions as there are many mislabelling.",
      "votes": null
    },
    {
      "id": "1159732",
      "postDate": "01/19/2021 12:37:59",
      "content": "<p>thanks, <a href=\"https://www.kaggle.com/tamilselvanmoorthy\" target=\"_blank\">@tamilselvanmoorthy</a> I didn't know about tampered function, I will look and try it</p>",
      "rawMarkdown": "thanks, @tamilselvanmoorthy I didn't know about tampered function, I will look and try it",
      "votes": null
    },
    {
      "id": "1160365",
      "postDate": "01/19/2021 21:22:02",
      "content": "<p>There's lots of methods to improve accuracy: search \"training in noisy labels\" there's lots of different ideas</p>",
      "rawMarkdown": "There's lots of methods to improve accuracy: search \"training in noisy labels\" there's lots of different ideas",
      "votes": null
    },
    {
      "id": "1160367",
      "postDate": "01/19/2021 21:23:58",
      "content": "<p>thanks <a href=\"https://www.kaggle.com/capiru\" target=\"_blank\">@capiru</a> I am trying💪 </p>",
      "rawMarkdown": "thanks @capiru I am trying💪",
      "votes": null
    },
    {
      "id": "1160480",
      "postDate": "01/20/2021 00:32:36",
      "content": "<p>If you search the discussion for: cleanlab is what i'm trying right now for cleaning labels</p>",
      "rawMarkdown": "If you search the discussion for: cleanlab is what i'm trying right now for cleaning labels",
      "votes": null
    },
    {
      "id": "1160851",
      "postDate": "01/20/2021 07:10:59",
      "content": "<p>ok thanks, I will try </p>",
      "rawMarkdown": "ok thanks, I will try",
      "votes": null
    },
    {
      "id": "1168122",
      "postDate": "01/24/2021 18:22:50",
      "content": "<p>use tampered loss functions</p>",
      "rawMarkdown": "use tampered loss functions",
      "votes": null
    },
    {
      "id": "1168528",
      "postDate": "01/25/2021 03:00:51",
      "content": "<p>It's true that ensemble sometimes increases lb, sometimes decreases lb, so it's necessary to know that many ensemble methods, such as weighted average or simple average, etc.. Ensemble makes me higher than 90.0. Now I'm going to use all the data for training. Although it's not easy to master the training times and find the extreme value, I think it's worth trying</p>",
      "rawMarkdown": "It's true that ensemble sometimes increases lb, sometimes decreases lb, so it's necessary to know that many ensemble methods, such as weighted average or simple average, etc.. Ensemble makes me higher than 90.0. Now I'm going to use all the data for training. Although it's not easy to master the training times and find the extreme value, I think it's worth trying",
      "votes": null
    },
    {
      "id": "1168615",
      "postDate": "01/25/2021 04:41:36",
      "content": "<p>thanks <a href=\"https://www.kaggle.com/pratt3000\" target=\"_blank\">@pratt3000</a> it was said :) </p>",
      "rawMarkdown": "thanks @pratt3000 it was said :)",
      "votes": null
    },
    {
      "id": "1168632",
      "postDate": "01/25/2021 04:46:08",
      "content": "<p>thanks, <a href=\"https://www.kaggle.com/zhangeng\" target=\"_blank\">@zhangeng</a>  I will try ensembling, can you give me a resource?</p>",
      "rawMarkdown": "thanks, @zhangeng  I will try ensembling, can you give me a resource?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1159190,
      "author_name": "zhangeng",
      "author_url": "",
      "post_date": "01/19/2021 05:09:55",
      "content": "<p>I am doing multiple model ensemble, it may not improve the LB score, but the robustness of the model may be useful for private data sets…</p>",
      "votes": null,
      "replies": [
        {
          "id": 1159225,
          "author_name": "kutaykutlu",
          "author_url": "",
          "post_date": "01/19/2021 05:38:41",
          "content": "<p>thanks <a href=\"https://www.kaggle.com/zhangeng\" target=\"_blank\">@zhangeng</a>, can you cite some resources?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1159235,
          "author_name": "thakurudit",
          "author_url": "",
          "post_date": "01/19/2021 05:45:34",
          "content": "<p><a href=\"https://www.kaggle.com/zhangeng\" target=\"_blank\">@zhangeng</a> I've also tried ensembling my models but didn't see any improvements on the lb, only on cv.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1159256,
          "author_name": "zhangeng",
          "author_url": "",
          "post_date": "01/19/2021 06:17:58",
          "content": "<p>For example, I am using vit, B4, and resnet networks for ensemble, and the model is being trained…</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1168122,
          "author_name": "pratt3000",
          "author_url": "",
          "post_date": "01/24/2021 18:22:50",
          "content": "<p>use tampered loss functions</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1168528,
          "author_name": "zhangeng",
          "author_url": "",
          "post_date": "01/25/2021 03:00:51",
          "content": "<p>It's true that ensemble sometimes increases lb, sometimes decreases lb, so it's necessary to know that many ensemble methods, such as weighted average or simple average, etc.. Ensemble makes me higher than 90.0. Now I'm going to use all the data for training. Although it's not easy to master the training times and find the extreme value, I think it's worth trying</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1168615,
          "author_name": "kutaykutlu",
          "author_url": "",
          "post_date": "01/25/2021 04:41:36",
          "content": "<p>thanks <a href=\"https://www.kaggle.com/pratt3000\" target=\"_blank\">@pratt3000</a> it was said :) </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1168632,
          "author_name": "kutaykutlu",
          "author_url": "",
          "post_date": "01/25/2021 04:46:08",
          "content": "<p>thanks, <a href=\"https://www.kaggle.com/zhangeng\" target=\"_blank\">@zhangeng</a>  I will try ensembling, can you give me a resource?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1159290,
      "author_name": "saurabhshahane",
      "author_url": "",
      "post_date": "01/19/2021 07:12:17",
      "content": "<ol>\n<li>Feeding good quality of data</li>\n<li>Reducing the noise</li>\n</ol>",
      "votes": null,
      "replies": [
        {
          "id": 1159293,
          "author_name": "kutaykutlu",
          "author_url": "",
          "post_date": "01/19/2021 07:14:00",
          "content": "<p>thanks <a href=\"https://www.kaggle.com/saurabhshahane\" target=\"_blank\">@saurabhshahane</a> 👍</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1159304,
          "author_name": "zhangeng",
          "author_url": "",
          "post_date": "01/19/2021 07:20:30",
          "content": "<p>The problem is that the test set is also very noisy…</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1159342,
          "author_name": "saurabhshahane",
          "author_url": "",
          "post_date": "01/19/2021 07:45:07",
          "content": "<p>I agree with you <a href=\"https://www.kaggle.com/zhangeng\" target=\"_blank\">@zhangeng</a> . I'm saying in general.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1159718,
      "author_name": "tamilselvanmoorthy",
      "author_url": "",
      "post_date": "01/19/2021 12:27:20",
      "content": "<ol>\n<li>try different augmentation techniques. </li>\n<li>efficient net gave good accuracy. </li>\n<li>use tampered loss functions as there are many mislabelling. </li>\n</ol>",
      "votes": null,
      "replies": [
        {
          "id": 1159732,
          "author_name": "kutaykutlu",
          "author_url": "",
          "post_date": "01/19/2021 12:37:59",
          "content": "<p>thanks, <a href=\"https://www.kaggle.com/tamilselvanmoorthy\" target=\"_blank\">@tamilselvanmoorthy</a> I didn't know about tampered function, I will look and try it</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1160365,
      "author_name": "capiru",
      "author_url": "",
      "post_date": "01/19/2021 21:22:02",
      "content": "<p>There's lots of methods to improve accuracy: search \"training in noisy labels\" there's lots of different ideas</p>",
      "votes": null,
      "replies": [
        {
          "id": 1160367,
          "author_name": "kutaykutlu",
          "author_url": "",
          "post_date": "01/19/2021 21:23:58",
          "content": "<p>thanks <a href=\"https://www.kaggle.com/capiru\" target=\"_blank\">@capiru</a> I am trying💪 </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1160480,
          "author_name": "capiru",
          "author_url": "",
          "post_date": "01/20/2021 00:32:36",
          "content": "<p>If you search the discussion for: cleanlab is what i'm trying right now for cleaning labels</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1160851,
          "author_name": "kutaykutlu",
          "author_url": "",
          "post_date": "01/20/2021 07:10:59",
          "content": "<p>ok thanks, I will try </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1159183": "Collect more data.\n- Collect a more diverse training set.\n- Train algorithm longer with gradient descent.\n- Try different optimization algorithm \n- Try a bigger or a smaller network.\n- Add L2 regularization or dropout.\n- Change network architecture (activation functions, hidden units, etc.)\n\nfrom @saurabhshahane \n- Feeding good quality of data\n- Reducing the noise\n\nfrom @tamilselvanmoorthy \n- use tampered loss functions as there are many mislabelling.",
    "1159190": "I am doing multiple model ensemble, it may not improve the LB score, but the robustness of the model may be useful for private data sets...",
    "1159225": "thanks @zhangeng, can you cite some resources?",
    "1159235": "zhangeng I've also tried ensembling my models but didn't see any improvements on the lb, only on cv.",
    "1159256": "For example, I am using vit, B4, and resnet networks for ensemble, and the model is being trained...",
    "1159290": "1. Feeding good quality of data\n2. Reducing the noise",
    "1159293": "thanks @saurabhshahane 👍",
    "1159304": "The problem is that the test set is also very noisy...",
    "1159342": "I agree with you @zhangeng . I'm saying in general.",
    "1159718": "1. try different augmentation techniques. \n2. efficient net gave good accuracy. \n3. use tampered loss functions as there are many mislabelling.",
    "1159732": "thanks, @tamilselvanmoorthy I didn't know about tampered function, I will look and try it",
    "1160365": "There's lots of methods to improve accuracy: search \"training in noisy labels\" there's lots of different ideas",
    "1160367": "thanks @capiru I am trying💪",
    "1160480": "If you search the discussion for: cleanlab is what i'm trying right now for cleaning labels",
    "1160851": "ok thanks, I will try",
    "1168122": "use tampered loss functions",
    "1168528": "It's true that ensemble sometimes increases lb, sometimes decreases lb, so it's necessary to know that many ensemble methods, such as weighted average or simple average, etc.. Ensemble makes me higher than 90.0. Now I'm going to use all the data for training. Although it's not easy to master the training times and find the extreme value, I think it's worth trying",
    "1168615": "thanks @pratt3000 it was said :)",
    "1168632": "thanks, @zhangeng  I will try ensembling, can you give me a resource?"
  },
  "source": "meta"
}