{
  "id": 212974,
  "title": "Random predict good for public LB? or Big shake may happen?",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/212974",
  "author_name": "",
  "post_date": "2021-01-21T00:31:33.237227900Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I tried various training setting like model, optimizer, loss, learning rate etc.</p>\n<p>From my tihs competition try, I think randomness of prediction contributes to public LB.</p>\n<p>Bitempered loss and large epoch looks good at local CV(accurayc = 0.89~0.9), but public LB score is 0.891. On the other hand softmax cross entropy is looks worse at local CV(0.88~0.89), but public LB is luckly 0.898~0.899.<br>\nAnd cutmix is better for LB than without it.</p>\n<p>I think public LB image contains noisy, so uncertainly prediction near random is good for LB.</p>\n<p>Or, big shake may happen?</p>\n<p>If possible, please tell me how do you think.</p>\n<p>Thanks</p>",
  "messages": [
    {
      "id": "1162074",
      "postDate": "01/21/2021 00:31:33",
      "content": "<p>I tried various training setting like model, optimizer, loss, learning rate etc.</p>\n<p>From my tihs competition try, I think randomness of prediction contributes to public LB.</p>\n<p>Bitempered loss and large epoch looks good at local CV(accurayc = 0.89~0.9), but public LB score is 0.891. On the other hand softmax cross entropy is looks worse at local CV(0.88~0.89), but public LB is luckly 0.898~0.899.<br>\nAnd cutmix is better for LB than without it.</p>\n<p>I think public LB image contains noisy, so uncertainly prediction near random is good for LB.</p>\n<p>Or, big shake may happen?</p>\n<p>If possible, please tell me how do you think.</p>\n<p>Thanks</p>",
      "rawMarkdown": "I tried various training setting like model, optimizer, loss, learning rate etc.\n\nFrom my tihs competition try, I think randomness of prediction contributes to public LB.\n\nBitempered loss and large epoch looks good at local CV(accurayc = 0.89~0.9), but public LB score is 0.891. On the other hand softmax cross entropy is looks worse at local CV(0.88~0.89), but public LB is luckly 0.898~0.899.\nAnd cutmix is better for LB than without it.\n\nI think public LB image contains noisy, so uncertainly prediction near random is good for LB.\n\nOr, big shake may happen?\n\nIf possible, please tell me how do you think.\n\nThanks",
      "votes": null
    },
    {
      "id": "1163581",
      "postDate": "01/21/2021 19:18:58",
      "content": "<p>Trust your CV, i also think public LB is noisy. </p>",
      "rawMarkdown": "Trust your CV, i also think public LB is noisy.",
      "votes": null
    },
    {
      "id": "1163710",
      "postDate": "01/21/2021 21:48:37",
      "content": "<p>There is a discussion which explains that the test dataset is also noisy. Here is the <a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017\" target=\"_blank\">link</a> </p>",
      "rawMarkdown": "There is a discussion which explains that the test dataset is also noisy. Here is the [link](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1163581,
      "author_name": "anku5hk",
      "author_url": "",
      "post_date": "01/21/2021 19:18:58",
      "content": "<p>Trust your CV, i also think public LB is noisy. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1163710,
      "author_name": "snnclsr",
      "author_url": "",
      "post_date": "01/21/2021 21:48:37",
      "content": "<p>There is a discussion which explains that the test dataset is also noisy. Here is the <a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017\" target=\"_blank\">link</a> </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1162074": "I tried various training setting like model, optimizer, loss, learning rate etc.\n\nFrom my tihs competition try, I think randomness of prediction contributes to public LB.\n\nBitempered loss and large epoch looks good at local CV(accurayc = 0.89~0.9), but public LB score is 0.891. On the other hand softmax cross entropy is looks worse at local CV(0.88~0.89), but public LB is luckly 0.898~0.899.\nAnd cutmix is better for LB than without it.\n\nI think public LB image contains noisy, so uncertainly prediction near random is good for LB.\n\nOr, big shake may happen?\n\nIf possible, please tell me how do you think.\n\nThanks",
    "1163581": "Trust your CV, i also think public LB is noisy.",
    "1163710": "There is a discussion which explains that the test dataset is also noisy. Here is the [link](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017)"
  },
  "source": "meta"
}