{
  "id": 105391,
  "title": "Hint",
  "url": "/competitions/aptos2019-blindness-detection/discussion/105391",
  "author_name": "OmerS",
  "post_date": "2019-08-22T19:37:02.500000",
  "votes": 4,
  "comment_count": 10,
  "views": 0,
  "content": "<p>Despite multiple training, I am still stuck at LB 0.8 for the last couple of week.\nI tried multiple  approachs  as suggest h here : \nTrain with  old data \nTrain with old data and freeze some layers \nMultiple image resolutions\n Different models (Efficient Net B0 - B6 )\nResnet50 \nStop training based on Validation \nKfolds \nTTA </p>\n\n<p>Any hint for a breakthrough ?\nMy goal is to be in the medal zone  I am not looking for the top </p>",
  "messages": [
    {
      "id": 605826,
      "postDate": "2019-08-22T19:37:02.500Z",
      "content": "<p>Despite multiple training, I am still stuck at LB 0.8 for the last couple of week.\nI tried multiple  approachs  as suggest h here : \nTrain with  old data \nTrain with old data and freeze some layers \nMultiple image resolutions\n Different models (Efficient Net B0 - B6 )\nResnet50 \nStop training based on Validation \nKfolds \nTTA </p>\n\n<p>Any hint for a breakthrough ?\nMy goal is to be in the medal zone  I am not looking for the top </p>",
      "rawMarkdown": "Despite multiple training, I am still stuck at LB 0.8 for the last couple of week.\nI tried multiple  approachs  as suggest h here : \nTrain with  old data \nTrain with old data and freeze some layers \nMultiple image resolutions\n Different models (Efficient Net B0 - B6 )\nResnet50 \nStop training based on Validation \nKfolds \nTTA \n\nAny hint for a breakthrough ?\nMy goal is to be in the medal zone  I am not looking for the top \n\n",
      "votes": 4
    },
    {
      "id": 605893,
      "postDate": "2019-08-22T23:07:33.993Z",
      "content": "<p>Me too, I am in the 0.8 zone. \nI am now trying without pre-processing (grey scale and cropping)</p>",
      "rawMarkdown": "Me too, I am in the 0.8 zone. \nI am now trying without pre-processing (grey scale and cropping)",
      "votes": 1,
      "replies": [
        {
          "id": 607115,
          "postDate": "2019-08-24T16:11:43.517Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 607345,
          "postDate": "2019-08-25T04:11:19.693Z",
          "content": "<p>no, I think they are the same, at least for the B4 Im running. Or I should say the pre-processing does slightly better than the un-preprocessing one, but if you do more argumentation it's pretty much the same</p>",
          "rawMarkdown": "no, I think they are the same, at least for the B4 Im running. Or I should say the pre-processing does slightly better than the un-preprocessing one, but if you do more argumentation it's pretty much the same\n"
        }
      ]
    },
    {
      "id": 606085,
      "postDate": "2019-08-23T07:07:11.110Z",
      "content": "<p>Try Ensembling using the approaches mentioned <a href=\"https://www.kaggle.com/c/aptos2019-blindness-detection/discussion/104981#latest-605844\">here</a>. Mostly likely it improve ur model performance</p>",
      "rawMarkdown": "Try Ensembling using the approaches mentioned [here](https://www.kaggle.com/c/aptos2019-blindness-detection/discussion/104981#latest-605844). Mostly likely it improve ur model performance",
      "votes": 2
    },
    {
      "id": 607057,
      "postDate": "2019-08-24T14:30:40.970Z",
      "content": "<p>Have you tried removing the TTA?</p>",
      "rawMarkdown": "Have you tried removing the TTA?\n"
    },
    {
      "id": 606568,
      "postDate": "2019-08-23T18:38:34.987Z",
      "content": "<p>Just got into the 0.8 zone! Have you used any pre-processing? In my case careful ensembling different models gave a  score boost.</p>",
      "rawMarkdown": "Just got into the 0.8 zone! Have you used any pre-processing? In my case careful ensembling different models gave a  score boost.",
      "replies": [
        {
          "id": 607114,
          "postDate": "2019-08-24T16:11:34.387Z",
          "content": "<p>how many models and how much boost</p>",
          "rawMarkdown": "how many models and how much boost\n",
          "votes": 1
        },
        {
          "id": 607166,
          "postDate": "2019-08-24T17:38:06.167Z",
          "content": "<p>3 models ~0.79 to 0.801,  so around .01 boost</p>",
          "rawMarkdown": "3 models ~0.79 to 0.801,  so around .01 boost",
          "votes": 2
        }
      ]
    },
    {
      "id": 607006,
      "postDate": "2019-08-24T12:38:25.527Z",
      "rawMarkdown": "",
      "isDeleted": true,
      "replies": [
        {
          "id": 607110,
          "postDate": "2019-08-24T16:07:30.460Z",
          "content": "<p>\"We can't find that page.\" </p>",
          "rawMarkdown": "\"We can't find that page.\" "
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 605893,
      "author_name": "Hao He",
      "author_url": "",
      "post_date": "2019-08-22T23:07:33.993000",
      "content": "<p>Me too, I am in the 0.8 zone. \nI am now trying without pre-processing (grey scale and cropping)</p>",
      "votes": 1,
      "replies": [
        {
          "id": 607115,
          "author_name": "",
          "author_url": "",
          "post_date": "2019-08-24T16:11:43.517000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 607345,
          "author_name": "Hao He",
          "author_url": "",
          "post_date": "2019-08-25T04:11:19.693000",
          "content": "<p>no, I think they are the same, at least for the B4 Im running. Or I should say the pre-processing does slightly better than the un-preprocessing one, but if you do more argumentation it's pretty much the same</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 606085,
      "author_name": "Bibek",
      "author_url": "",
      "post_date": "2019-08-23T07:07:11.110000",
      "content": "<p>Try Ensembling using the approaches mentioned <a href=\"https://www.kaggle.com/c/aptos2019-blindness-detection/discussion/104981#latest-605844\">here</a>. Mostly likely it improve ur model performance</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 607057,
      "author_name": "AmardeepGanguly",
      "author_url": "",
      "post_date": "2019-08-24T14:30:40.970000",
      "content": "<p>Have you tried removing the TTA?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 606568,
      "author_name": "Brian Lee",
      "author_url": "",
      "post_date": "2019-08-23T18:38:34.987000",
      "content": "<p>Just got into the 0.8 zone! Have you used any pre-processing? In my case careful ensembling different models gave a  score boost.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 607114,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-08-24T16:11:34.387000",
          "content": "<p>how many models and how much boost</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 607166,
          "author_name": "Brian Lee",
          "author_url": "",
          "post_date": "2019-08-24T17:38:06.167000",
          "content": "<p>3 models ~0.79 to 0.801,  so around .01 boost</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 607006,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-08-24T12:38:25.527000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 607110,
          "author_name": "hahaha",
          "author_url": "",
          "post_date": "2019-08-24T16:07:30.460000",
          "content": "<p>\"We can't find that page.\" </p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "605826": "Despite multiple training, I am still stuck at LB 0.8 for the last couple of week.\nI tried multiple  approachs  as suggest h here : \nTrain with  old data \nTrain with old data and freeze some layers \nMultiple image resolutions\n Different models (Efficient Net B0 - B6 )\nResnet50 \nStop training based on Validation \nKfolds \nTTA \n\nAny hint for a breakthrough ?\nMy goal is to be in the medal zone  I am not looking for the top \n\n",
    "605893": "Me too, I am in the 0.8 zone. \nI am now trying without pre-processing (grey scale and cropping)",
    "606085": "Try Ensembling using the approaches mentioned [here](https://www.kaggle.com/c/aptos2019-blindness-detection/discussion/104981#latest-605844). Mostly likely it improve ur model performance",
    "607057": "Have you tried removing the TTA?\n",
    "606568": "Just got into the 0.8 zone! Have you used any pre-processing? In my case careful ensembling different models gave a  score boost.",
    "607006": ""
  }
}