{
  "id": 228776,
  "title": "Is multi class performing better than multi label problem ???",
  "url": "/competitions/plant-pathology-2021-fgvc8/discussion/228776",
  "author_name": "",
  "post_date": "2021-03-26T11:48:21.508861600Z",
  "votes": 5,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hello!</p>\n<p>Even-though this competition is multi-label classification problem, top LB public works considered it as multi-class problem and used softmax+categorical cross entropy.</p>\n<p>This makes me do a quick experimentation using a basic model without any augmentations:</p>\n<ol>\n<li>Multi-label: Sigmoid and Binary Cross entropy - LB: 0.454</li>\n<li>Multi-Class: Softmax and Categorical Cross entropy - LB: 0.582</li>\n</ol>",
  "messages": [
    {
      "id": "1253132",
      "postDate": "03/26/2021 11:48:21",
      "content": "<p>Hello!</p>\n<p>Even-though this competition is multi-label classification problem, top LB public works considered it as multi-class problem and used softmax+categorical cross entropy.</p>\n<p>This makes me do a quick experimentation using a basic model without any augmentations:</p>\n<ol>\n<li>Multi-label: Sigmoid and Binary Cross entropy - LB: 0.454</li>\n<li>Multi-Class: Softmax and Categorical Cross entropy - LB: 0.582</li>\n</ol>",
      "rawMarkdown": "Hello!\n\nEven-though this competition is multi-label classification problem, top LB public works considered it as multi-class problem and used softmax+categorical cross entropy.\n\nThis makes me do a quick experimentation using a basic model without any augmentations:\n\n1. Multi-label: Sigmoid and Binary Cross entropy - LB: 0.454\n2. Multi-Class: Softmax and Categorical Cross entropy - LB: 0.582",
      "votes": null
    },
    {
      "id": "1253388",
      "postDate": "03/26/2021 17:07:02",
      "content": "<p>I've got similar results with both of them.</p>",
      "rawMarkdown": "I've got similar results with both of them.",
      "votes": null
    },
    {
      "id": "1253447",
      "postDate": "03/26/2021 18:20:44",
      "content": "<p>As per the competition description, ideally it should be a multi-label classification. </p>\n<p>The Probable reason for working multi-class could be the hidden test images have similar combination of diseases as training images.</p>\n<p>6 base disease categories ('healthy','scab','complex','rust','frog_eye_leaf_spot','powdery_mildew') and 5 diseases combinations ('scab frog_eye_leaf_spot', 'frog_eye_leaf_spot complex','rust frog_eye_leaf_spot', 'powdery_mildew complex','rust complex'). </p>",
      "rawMarkdown": "As per the competition description, ideally it should be a multi-label classification. \n\nThe Probable reason for working multi-class could be the hidden test images have similar combination of diseases as training images.\n\n6 base disease categories ('healthy','scab','complex','rust','frog_eye_leaf_spot','powdery_mildew') and 5 diseases combinations ('scab frog_eye_leaf_spot', 'frog_eye_leaf_spot complex','rust frog_eye_leaf_spot', 'powdery_mildew complex','rust complex').",
      "votes": null
    },
    {
      "id": "1265718",
      "postDate": "04/07/2021 06:43:11",
      "content": "<p>I think we could use Multi-Class classification for this task if the label distribution of the training and testing dataset is close. But there might not be a way to test this hypothesis. A safer way is to sumbit a multi-label model and a multi-class label. </p>",
      "rawMarkdown": "I think we could use Multi-Class classification for this task if the label distribution of the training and testing dataset is close. But there might not be a way to test this hypothesis. A safer way is to sumbit a multi-label model and a multi-class label.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1253388,
      "author_name": "negreaclaudiu",
      "author_url": "",
      "post_date": "03/26/2021 17:07:02",
      "content": "<p>I've got similar results with both of them.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1253447,
          "author_name": "vigneshirtt",
          "author_url": "",
          "post_date": "03/26/2021 18:20:44",
          "content": "<p>As per the competition description, ideally it should be a multi-label classification. </p>\n<p>The Probable reason for working multi-class could be the hidden test images have similar combination of diseases as training images.</p>\n<p>6 base disease categories ('healthy','scab','complex','rust','frog_eye_leaf_spot','powdery_mildew') and 5 diseases combinations ('scab frog_eye_leaf_spot', 'frog_eye_leaf_spot complex','rust frog_eye_leaf_spot', 'powdery_mildew complex','rust complex'). </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1265718,
      "author_name": "crissallan",
      "author_url": "",
      "post_date": "04/07/2021 06:43:11",
      "content": "<p>I think we could use Multi-Class classification for this task if the label distribution of the training and testing dataset is close. But there might not be a way to test this hypothesis. A safer way is to sumbit a multi-label model and a multi-class label. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1253132": "Hello!\n\nEven-though this competition is multi-label classification problem, top LB public works considered it as multi-class problem and used softmax+categorical cross entropy.\n\nThis makes me do a quick experimentation using a basic model without any augmentations:\n\n1. Multi-label: Sigmoid and Binary Cross entropy - LB: 0.454\n2. Multi-Class: Softmax and Categorical Cross entropy - LB: 0.582",
    "1253388": "I've got similar results with both of them.",
    "1253447": "As per the competition description, ideally it should be a multi-label classification. \n\nThe Probable reason for working multi-class could be the hidden test images have similar combination of diseases as training images.\n\n6 base disease categories ('healthy','scab','complex','rust','frog_eye_leaf_spot','powdery_mildew') and 5 diseases combinations ('scab frog_eye_leaf_spot', 'frog_eye_leaf_spot complex','rust frog_eye_leaf_spot', 'powdery_mildew complex','rust complex').",
    "1265718": "I think we could use Multi-Class classification for this task if the label distribution of the training and testing dataset is close. But there might not be a way to test this hypothesis. A safer way is to sumbit a multi-label model and a multi-class label."
  },
  "source": "meta"
}