{
  "id": 101192,
  "title": "Inconsistent LB score ",
  "url": "/competitions/aptos2019-blindness-detection/discussion/101192",
  "author_name": "",
  "post_date": "2019-07-23T22:00:56.432590200Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I just trained my ResNet and once i submitted it, i got 0.741. Later, I decided to run and submit my kernel again and got 0.751(didn't change my code) . Is it due to the small sample size of the validation set or some kaggle problems? </p>",
  "messages": [
    {
      "id": "583006",
      "postDate": "07/23/2019 22:00:56",
      "content": "<p>I just trained my ResNet and once i submitted it, i got 0.741. Later, I decided to run and submit my kernel again and got 0.751(didn't change my code) . Is it due to the small sample size of the validation set or some kaggle problems? </p>",
      "rawMarkdown": "I just trained my ResNet and once i submitted it, i got 0.741. Later, I decided to run and submit my kernel again and got 0.751(didn't change my code) . Is it due to the small sample size of the validation set or some kaggle problems?",
      "votes": null
    },
    {
      "id": "583050",
      "postDate": "07/24/2019 00:24:22",
      "content": "<p>1/ Maybe you are not using the same seed in your experiments\n2/ Transformations like flips and rotations are performed randomly\n3/ You can use TTA or kfold validation. It acts like ensembling, it will reduce variance and probably boost your score.</p>",
      "rawMarkdown": "1/ Maybe you are not using the same seed in your experiments\n2/ Transformations like flips and rotations are performed randomly\n3/ You can use TTA or kfold validation. It acts like ensembling, it will reduce variance and probably boost your score.",
      "votes": null
    },
    {
      "id": "583080",
      "postDate": "07/24/2019 02:21:19",
      "content": "<p>I think TTA can solve your problem</p>",
      "rawMarkdown": "I think TTA can solve your problem",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 583050,
      "author_name": "seif95",
      "author_url": "",
      "post_date": "07/24/2019 00:24:22",
      "content": "<p>1/ Maybe you are not using the same seed in your experiments\n2/ Transformations like flips and rotations are performed randomly\n3/ You can use TTA or kfold validation. It acts like ensembling, it will reduce variance and probably boost your score.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 583080,
      "author_name": "suzhengpeng",
      "author_url": "",
      "post_date": "07/24/2019 02:21:19",
      "content": "<p>I think TTA can solve your problem</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "583006": "I just trained my ResNet and once i submitted it, i got 0.741. Later, I decided to run and submit my kernel again and got 0.751(didn't change my code) . Is it due to the small sample size of the validation set or some kaggle problems?",
    "583050": "1/ Maybe you are not using the same seed in your experiments\n2/ Transformations like flips and rotations are performed randomly\n3/ You can use TTA or kfold validation. It acts like ensembling, it will reduce variance and probably boost your score.",
    "583080": "I think TTA can solve your problem"
  },
  "source": "meta"
}