{
  "id": 228896,
  "title": "shifted but lower score.",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/228896",
  "author_name": "",
  "post_date": "2021-03-27T03:18:12.388016300Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I used shift method like <a href=\"https://www.kaggle.com/wrrosa/hubmap-tf-with-tpu-efficientunet-512x512-tfrecs\" target=\"_blank\">https://www.kaggle.com/wrrosa/hubmap-tf-with-tpu-efficientunet-512x512-tfrecs</a>. but my LB score got lower than even just using pure data with nagative sampling. CV got 0.940 with lb 0.916 compared with LB:0.920 - CV:0.935(around the same size of shited samples) without TTAs in the same fold. overfit? or other reasons.</p>\n<p><strong>Guess:</strong> One of the reasons could be using too large learning rate causing my model converging too fast in the early steps.</p>",
  "messages": [
    {
      "id": "1253789",
      "postDate": "03/27/2021 03:18:12",
      "content": "<p>I used shift method like <a href=\"https://www.kaggle.com/wrrosa/hubmap-tf-with-tpu-efficientunet-512x512-tfrecs\" target=\"_blank\">https://www.kaggle.com/wrrosa/hubmap-tf-with-tpu-efficientunet-512x512-tfrecs</a>. but my LB score got lower than even just using pure data with nagative sampling. CV got 0.940 with lb 0.916 compared with LB:0.920 - CV:0.935(around the same size of shited samples) without TTAs in the same fold. overfit? or other reasons.</p>\n<p><strong>Guess:</strong> One of the reasons could be using too large learning rate causing my model converging too fast in the early steps.</p>",
      "rawMarkdown": "I used shift method like https://www.kaggle.com/wrrosa/hubmap-tf-with-tpu-efficientunet-512x512-tfrecs. but my LB score got lower than even just using pure data with nagative sampling. CV got 0.940 with lb 0.916 compared with LB:0.920 - CV:0.935(around the same size of shited samples) without TTAs in the same fold. overfit? or other reasons.\n\n**Guess:** One of the reasons could be using too large learning rate causing my model converging too fast in the early steps.",
      "votes": null
    },
    {
      "id": "1254060",
      "postDate": "03/27/2021 10:01:53",
      "content": "<p>Old saying goes like <code>Trust your CV</code>. What do you think of that? </p>",
      "rawMarkdown": "Old saying goes like `Trust your CV`. What do you think of that?",
      "votes": null
    },
    {
      "id": "1254376",
      "postDate": "03/27/2021 15:56:54",
      "content": "<p>I cant agree in my case :(. since my test dataset is chosen to be the simplest images. CV score is unreliable I thought.</p>",
      "rawMarkdown": "I cant agree in my case :(. since my test dataset is chosen to be the simplest images. CV score is unreliable I thought.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1254060,
      "author_name": "louieshao",
      "author_url": "",
      "post_date": "03/27/2021 10:01:53",
      "content": "<p>Old saying goes like <code>Trust your CV</code>. What do you think of that? </p>",
      "votes": null,
      "replies": [
        {
          "id": 1254376,
          "author_name": "southsakura",
          "author_url": "",
          "post_date": "03/27/2021 15:56:54",
          "content": "<p>I cant agree in my case :(. since my test dataset is chosen to be the simplest images. CV score is unreliable I thought.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1253789": "I used shift method like https://www.kaggle.com/wrrosa/hubmap-tf-with-tpu-efficientunet-512x512-tfrecs. but my LB score got lower than even just using pure data with nagative sampling. CV got 0.940 with lb 0.916 compared with LB:0.920 - CV:0.935(around the same size of shited samples) without TTAs in the same fold. overfit? or other reasons.\n\n**Guess:** One of the reasons could be using too large learning rate causing my model converging too fast in the early steps.",
    "1254060": "Old saying goes like `Trust your CV`. What do you think of that?",
    "1254376": "I cant agree in my case :(. since my test dataset is chosen to be the simplest images. CV score is unreliable I thought."
  },
  "source": "meta"
}