{
  "id": 110349,
  "title": "Sharing scores without leak",
  "url": "/competitions/recursion-cellular-image-classification/discussion/110349",
  "author_name": "",
  "post_date": "2019-09-27T02:22:07.315640700Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Congrats to all.</p>\n\n<p>The plate leak is powerful in this competition. In the last week,  I found that improvement of top-1 accuracy contributed little to LB score.</p>\n\n<p>I'm curious about the \"real\" score kagglers could achieve.</p>\n\n<p>Here's mine, simply averaging predictions of two sites, no 1108/277 linear assignment.</p>\n\n<p>| Model | Public LB | Private LB |\n| --- | --- | --- |\n| Single model | 0.95531 | 0.98321 |\n| Ensemble | 0.96366 | 0.98481 |</p>",
  "messages": [
    {
      "id": "634970",
      "postDate": "09/27/2019 02:22:07",
      "content": "<p>Congrats to all.</p>\n\n<p>The plate leak is powerful in this competition. In the last week,  I found that improvement of top-1 accuracy contributed little to LB score.</p>\n\n<p>I'm curious about the \"real\" score kagglers could achieve.</p>\n\n<p>Here's mine, simply averaging predictions of two sites, no 1108/277 linear assignment.</p>\n\n<p>| Model | Public LB | Private LB |\n| --- | --- | --- |\n| Single model | 0.95531 | 0.98321 |\n| Ensemble | 0.96366 | 0.98481 |</p>",
      "rawMarkdown": "Congrats to all.\n\nThe plate leak is powerful in this competition. In the last week,  I found that improvement of top-1 accuracy contributed little to LB score.\n\nI'm curious about the \"real\" score kagglers could achieve.\n\nHere's mine, simply averaging predictions of two sites, no 1108/277 linear assignment.\n\n| Model | Public LB | Private LB |\n| --- | --- | --- |\n| Single model | 0.95531 | 0.98321 |\n| Ensemble | 0.96366 | 0.98481 |",
      "votes": null
    },
    {
      "id": "634975",
      "postDate": "09/27/2019 02:31:58",
      "content": "<p>Wow, is this score without doing linear assignment? It is so impressive. </p>\n\n<p>Settings: \n- 6C5 = 30 combination of channels, \n- Averaging two sites\n- No linear assignment.</p>\n\n<p>|Model| Public | Private|\n|:---|-----|---:|\n|seresnext101| 0.74813 | 0.87979 |\n|seresnext50| 0.74926 | 0.88432 |</p>",
      "rawMarkdown": "Wow, is this score without doing linear assignment? It is so impressive. \n\nSettings: \n- 6C5 = 30 combination of channels, \n- Averaging two sites\n- No linear assignment.\n\n|Model| Public | Private|\n|:---|-----|---:|\n|seresnext101| 0.74813 | 0.87979 |\n|seresnext50| 0.74926 | 0.88432 |",
      "votes": null
    },
    {
      "id": "635155",
      "postDate": "09/27/2019 07:55:17",
      "content": "<p>Models: DenseNet (121, 169); ResNet (50, 101); SE_ResNeXt (50, 101); Efficient (B2,3,4)\nSettings: Average 2 sites\n|  Model | Public  | Private |\n| --- | --- | --- |\n| Ensemble (No linear assignment) | 0.94967 | 0.94212 |\n| Ensemble (No plate leak) | 0.94651 | 0.89233 |</p>",
      "rawMarkdown": "Models: DenseNet (121, 169); ResNet (50, 101); SE_ResNeXt (50, 101); Efficient (B2,3,4)\nSettings: Average 2 sites\n|  Model | Public  | Private |\n| --- | --- | --- |\n| Ensemble (No linear assignment) | 0.94967 | 0.94212 |\n| Ensemble (No plate leak) | 0.94651 | 0.89233 |",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 634975,
      "author_name": "backaggle",
      "author_url": "",
      "post_date": "09/27/2019 02:31:58",
      "content": "<p>Wow, is this score without doing linear assignment? It is so impressive. </p>\n\n<p>Settings: \n- 6C5 = 30 combination of channels, \n- Averaging two sites\n- No linear assignment.</p>\n\n<p>|Model| Public | Private|\n|:---|-----|---:|\n|seresnext101| 0.74813 | 0.87979 |\n|seresnext50| 0.74926 | 0.88432 |</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 635155,
      "author_name": "lego1st",
      "author_url": "",
      "post_date": "09/27/2019 07:55:17",
      "content": "<p>Models: DenseNet (121, 169); ResNet (50, 101); SE_ResNeXt (50, 101); Efficient (B2,3,4)\nSettings: Average 2 sites\n|  Model | Public  | Private |\n| --- | --- | --- |\n| Ensemble (No linear assignment) | 0.94967 | 0.94212 |\n| Ensemble (No plate leak) | 0.94651 | 0.89233 |</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "634970": "Congrats to all.\n\nThe plate leak is powerful in this competition. In the last week,  I found that improvement of top-1 accuracy contributed little to LB score.\n\nI'm curious about the \"real\" score kagglers could achieve.\n\nHere's mine, simply averaging predictions of two sites, no 1108/277 linear assignment.\n\n| Model | Public LB | Private LB |\n| --- | --- | --- |\n| Single model | 0.95531 | 0.98321 |\n| Ensemble | 0.96366 | 0.98481 |",
    "634975": "Wow, is this score without doing linear assignment? It is so impressive. \n\nSettings: \n- 6C5 = 30 combination of channels, \n- Averaging two sites\n- No linear assignment.\n\n|Model| Public | Private|\n|:---|-----|---:|\n|seresnext101| 0.74813 | 0.87979 |\n|seresnext50| 0.74926 | 0.88432 |",
    "635155": "Models: DenseNet (121, 169); ResNet (50, 101); SE_ResNeXt (50, 101); Efficient (B2,3,4)\nSettings: Average 2 sites\n|  Model | Public  | Private |\n| --- | --- | --- |\n| Ensemble (No linear assignment) | 0.94967 | 0.94212 |\n| Ensemble (No plate leak) | 0.94651 | 0.89233 |"
  },
  "source": "meta"
}