{
  "id": 214826,
  "title": "Can we down sample the dominant class named Cassava Mosaic Disease ?",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/214826",
  "author_name": "",
  "post_date": "2021-01-27T18:32:01.796629900Z",
  "votes": 2,
  "comment_count": 8,
  "views": 0,
  "content": "<p>In this Cassava disease classification out of 5 classes <strong>4th class(Cassava Mosaic Disease (CMD))</strong> is more dominant. My question is can we down sample this class by taking 40%  of 4th class to get better results?</p>\n<p>Thanks in advance!</p>",
  "messages": [
    {
      "id": "1173227",
      "postDate": "01/27/2021 18:32:01",
      "content": "<p>In this Cassava disease classification out of 5 classes <strong>4th class(Cassava Mosaic Disease (CMD))</strong> is more dominant. My question is can we down sample this class by taking 40%  of 4th class to get better results?</p>\n<p>Thanks in advance!</p>",
      "rawMarkdown": "In this Cassava disease classification out of 5 classes **4th class(Cassava Mosaic Disease (CMD))** is more dominant. My question is can we down sample this class by taking 40%  of 4th class to get better results?\n\nThanks in advance!",
      "votes": null
    },
    {
      "id": "1173287",
      "postDate": "01/27/2021 19:12:38",
      "content": "<p>Try it.I had try to increase 0 1 2 4 classes with a simple way,single model PB 0.898change to 0.897😭</p>",
      "rawMarkdown": "Try it.I had try to increase 0 1 2 4 classes with a simple way,single model PB 0.898change to 0.897😭",
      "votes": null
    },
    {
      "id": "1173705",
      "postDate": "01/28/2021 04:29:16",
      "content": "<p>ok, sure I will try and update to you. Thank you <a href=\"https://www.kaggle.com/darknesszx\" target=\"_blank\">@darknesszx</a> </p>",
      "rawMarkdown": "ok, sure I will try and update to you. Thank you @darknesszx",
      "votes": null
    },
    {
      "id": "1174833",
      "postDate": "01/28/2021 18:36:13",
      "content": "<p>Yes definitely<br>\nBut only downsampling wont do all of it<br>\nyou also need to increase the sample of other classes as well. <br>\nif you want to achieve a better accuracy.<br>\nGood Luck</p>",
      "rawMarkdown": "Yes definitely\nBut only downsampling wont do all of it\nyou also need to increase the sample of other classes as well. \nif you want to achieve a better accuracy.\nGood Luck",
      "votes": null
    },
    {
      "id": "1175263",
      "postDate": "01/29/2021 03:45:37",
      "content": "<p>One of the ideas that we have is to oversample the non-CMD classes by producing more images through augmentations, meaning using up to ~13,000 images per class. The training time increased a lot yet the performance of the model did not improve. </p>\n<p>If time permits I may revisit our methodology if there were any errors in it, but that's another idea you can explore.</p>",
      "rawMarkdown": "One of the ideas that we have is to oversample the non-CMD classes by producing more images through augmentations, meaning using up to ~13,000 images per class. The training time increased a lot yet the performance of the model did not improve. \n\nIf time permits I may revisit our methodology if there were any errors in it, but that's another idea you can explore.",
      "votes": null
    },
    {
      "id": "1175293",
      "postDate": "01/29/2021 04:27:12",
      "content": "<p>hi <a href=\"https://www.kaggle.com/ayannareda\" target=\"_blank\">@ayannareda</a> ,Yeah i did like that. At first I downsample Cassava Mosaic Disease (CMD) and apply real time augmenatation using albumentation for all the classes. </p>\n<p>is this ok??</p>",
      "rawMarkdown": "hi @ayannareda ,Yeah i did like that. At first I downsample Cassava Mosaic Disease (CMD) and apply real time augmenatation using albumentation for all the classes. \n\nis this ok??",
      "votes": null
    },
    {
      "id": "1175442",
      "postDate": "01/29/2021 07:08:17",
      "content": "<p>To solve the problem of category balance has been discussed before. My simple idea is to classify the CMD and non CMD categories into two categories, and then to classify the remaining non CMD categories into several categories. Under the same conditions, the two categories may be more accurate than the multi categories, for reference only</p>",
      "rawMarkdown": "To solve the problem of category balance has been discussed before. My simple idea is to classify the CMD and non CMD categories into two categories, and then to classify the remaining non CMD categories into several categories. Under the same conditions, the two categories may be more accurate than the multi categories, for reference only",
      "votes": null
    },
    {
      "id": "1175572",
      "postDate": "01/29/2021 08:26:11",
      "content": "<p>sure man, I will try👍👍.</p>",
      "rawMarkdown": "sure man, I will try👍👍.",
      "votes": null
    },
    {
      "id": "1196175",
      "postDate": "02/11/2021 09:56:26",
      "content": "<p>hi <a href=\"https://www.kaggle.com/zhangeng\" target=\"_blank\">@zhangeng</a> , As per your suggestion I did  for CMD and non-CMD  classification it gives 93% accuracy for efficientnetB05 model. After that I classify non-CMD class into several categories it gives 73% accuracy using the same model. But overall accuracy is very low which is 59.9%.</p>\n<p>I think Non-CMD images are not well taken images. It's better to use as a single model and I will get 84.9% accuracy by using same efficientnetb05 model.</p>\n<p>Have you tried anything like your approach with increasing accuracy help me with that.</p>\n<p>Thank you</p>",
      "rawMarkdown": "hi @zhangeng , As per your suggestion I did  for CMD and non-CMD  classification it gives 93% accuracy for efficientnetB05 model. After that I classify non-CMD class into several categories it gives 73% accuracy using the same model. But overall accuracy is very low which is 59.9%.\n\nI think Non-CMD images are not well taken images. It's better to use as a single model and I will get 84.9% accuracy by using same efficientnetb05 model.\n\nHave you tried anything like your approach with increasing accuracy help me with that.\n\nThank you",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1173287,
      "author_name": "darknesszx",
      "author_url": "",
      "post_date": "01/27/2021 19:12:38",
      "content": "<p>Try it.I had try to increase 0 1 2 4 classes with a simple way,single model PB 0.898change to 0.897😭</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1173705,
      "author_name": "dhurky",
      "author_url": "",
      "post_date": "01/28/2021 04:29:16",
      "content": "<p>ok, sure I will try and update to you. Thank you <a href=\"https://www.kaggle.com/darknesszx\" target=\"_blank\">@darknesszx</a> </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1174833,
      "author_name": "ayannareda",
      "author_url": "",
      "post_date": "01/28/2021 18:36:13",
      "content": "<p>Yes definitely<br>\nBut only downsampling wont do all of it<br>\nyou also need to increase the sample of other classes as well. <br>\nif you want to achieve a better accuracy.<br>\nGood Luck</p>",
      "votes": null,
      "replies": [
        {
          "id": 1175293,
          "author_name": "dhurky",
          "author_url": "",
          "post_date": "01/29/2021 04:27:12",
          "content": "<p>hi <a href=\"https://www.kaggle.com/ayannareda\" target=\"_blank\">@ayannareda</a> ,Yeah i did like that. At first I downsample Cassava Mosaic Disease (CMD) and apply real time augmenatation using albumentation for all the classes. </p>\n<p>is this ok??</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1175263,
      "author_name": "jasondolorso",
      "author_url": "",
      "post_date": "01/29/2021 03:45:37",
      "content": "<p>One of the ideas that we have is to oversample the non-CMD classes by producing more images through augmentations, meaning using up to ~13,000 images per class. The training time increased a lot yet the performance of the model did not improve. </p>\n<p>If time permits I may revisit our methodology if there were any errors in it, but that's another idea you can explore.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1175442,
      "author_name": "zhangeng",
      "author_url": "",
      "post_date": "01/29/2021 07:08:17",
      "content": "<p>To solve the problem of category balance has been discussed before. My simple idea is to classify the CMD and non CMD categories into two categories, and then to classify the remaining non CMD categories into several categories. Under the same conditions, the two categories may be more accurate than the multi categories, for reference only</p>",
      "votes": null,
      "replies": [
        {
          "id": 1175572,
          "author_name": "dhurky",
          "author_url": "",
          "post_date": "01/29/2021 08:26:11",
          "content": "<p>sure man, I will try👍👍.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1196175,
          "author_name": "dhurky",
          "author_url": "",
          "post_date": "02/11/2021 09:56:26",
          "content": "<p>hi <a href=\"https://www.kaggle.com/zhangeng\" target=\"_blank\">@zhangeng</a> , As per your suggestion I did  for CMD and non-CMD  classification it gives 93% accuracy for efficientnetB05 model. After that I classify non-CMD class into several categories it gives 73% accuracy using the same model. But overall accuracy is very low which is 59.9%.</p>\n<p>I think Non-CMD images are not well taken images. It's better to use as a single model and I will get 84.9% accuracy by using same efficientnetb05 model.</p>\n<p>Have you tried anything like your approach with increasing accuracy help me with that.</p>\n<p>Thank you</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1173227": "In this Cassava disease classification out of 5 classes **4th class(Cassava Mosaic Disease (CMD))** is more dominant. My question is can we down sample this class by taking 40%  of 4th class to get better results?\n\nThanks in advance!",
    "1173287": "Try it.I had try to increase 0 1 2 4 classes with a simple way,single model PB 0.898change to 0.897😭",
    "1173705": "ok, sure I will try and update to you. Thank you @darknesszx",
    "1174833": "Yes definitely\nBut only downsampling wont do all of it\nyou also need to increase the sample of other classes as well. \nif you want to achieve a better accuracy.\nGood Luck",
    "1175263": "One of the ideas that we have is to oversample the non-CMD classes by producing more images through augmentations, meaning using up to ~13,000 images per class. The training time increased a lot yet the performance of the model did not improve. \n\nIf time permits I may revisit our methodology if there were any errors in it, but that's another idea you can explore.",
    "1175293": "hi @ayannareda ,Yeah i did like that. At first I downsample Cassava Mosaic Disease (CMD) and apply real time augmenatation using albumentation for all the classes. \n\nis this ok??",
    "1175442": "To solve the problem of category balance has been discussed before. My simple idea is to classify the CMD and non CMD categories into two categories, and then to classify the remaining non CMD categories into several categories. Under the same conditions, the two categories may be more accurate than the multi categories, for reference only",
    "1175572": "sure man, I will try👍👍.",
    "1196175": "hi @zhangeng , As per your suggestion I did  for CMD and non-CMD  classification it gives 93% accuracy for efficientnetB05 model. After that I classify non-CMD class into several categories it gives 73% accuracy using the same model. But overall accuracy is very low which is 59.9%.\n\nI think Non-CMD images are not well taken images. It's better to use as a single model and I will get 84.9% accuracy by using same efficientnetb05 model.\n\nHave you tried anything like your approach with increasing accuracy help me with that.\n\nThank you"
  },
  "source": "meta"
}