{
  "id": 117364,
  "title": "Best Private LB models without vs with neighbor slice(s) info",
  "url": "/competitions/rsna-intracranial-hemorrhage-detection/discussion/117364",
  "author_name": "Kerem Turgutlu",
  "post_date": "2019-11-14T23:36:51.154000",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>First, congratulations to everyone involved in stage-2 and medal winning top teams. After a brief overview of the top solutions in discussions it seems like using neighboring slices as features either with RNN models or with feature engineering is the key to boost LB scores. I thought it would be nice to capture the impact without and with neighbor slice features.</p>\n\n<p>My best model scored 0.061 in Private LB - stack of 5 fold TTA preds of 2 models (eff-b0 and resnet50) - no neighbor slice info or no metadata info.</p>\n\n<p>Thanks!</p>",
  "messages": [
    {
      "id": 674221,
      "postDate": "2019-11-16T03:00:29.397Z",
      "content": "<p>I got 0.058, without using neighbouring slides ResNext50 with 5 fold, 3 window setting.</p>",
      "rawMarkdown": "I got 0.058, without using neighbouring slides ResNext50 with 5 fold, 3 window setting.",
      "votes": 1
    },
    {
      "id": 673657,
      "postDate": "2019-11-15T09:42:35.530Z",
      "content": "<p>I got .057 without using neighbouring slides Efficientnet b5 5 fold</p>",
      "rawMarkdown": "I got .057 without using neighbouring slides Efficientnet b5 5 fold",
      "votes": 1,
      "replies": [
        {
          "id": 674105,
          "postDate": "2019-11-15T22:29:27.193Z",
          "content": "<p>Can u explain more ? Thanks :)</p>",
          "rawMarkdown": "Can u explain more ? Thanks :)"
        },
        {
          "id": 674192,
          "postDate": "2019-11-16T01:44:22.587Z",
          "content": "<p>Sure, like my private score is just based on treating all slides as different. I trained Efficientnet B5 with 2 different preprocessings brain subdural bone and brain subdural tissue 5 fold then simply average ensemble 10 models. Actually to be honest I didn't know that we could use metadata for post processing. What was your approch? Thanks</p>",
          "rawMarkdown": "Sure, like my private score is just based on treating all slides as different. I trained Efficientnet B5 with 2 different preprocessings brain subdural bone and brain subdural tissue 5 fold then simply average ensemble 10 models. Actually to be honest I didn't know that we could use metadata for post processing. What was your approch? Thanks"
        }
      ]
    },
    {
      "id": 673404,
      "postDate": "2019-11-14T23:36:51.153Z",
      "content": "<p>First, congratulations to everyone involved in stage-2 and medal winning top teams. After a brief overview of the top solutions in discussions it seems like using neighboring slices as features either with RNN models or with feature engineering is the key to boost LB scores. I thought it would be nice to capture the impact without and with neighbor slice features.</p>\n\n<p>My best model scored 0.061 in Private LB - stack of 5 fold TTA preds of 2 models (eff-b0 and resnet50) - no neighbor slice info or no metadata info.</p>\n\n<p>Thanks!</p>",
      "rawMarkdown": "First, congratulations to everyone involved in stage-2 and medal winning top teams. After a brief overview of the top solutions in discussions it seems like using neighboring slices as features either with RNN models or with feature engineering is the key to boost LB scores. I thought it would be nice to capture the impact without and with neighbor slice features.\n\nMy best model scored 0.061 in Private LB - stack of 5 fold TTA preds of 2 models (eff-b0 and resnet50) - no neighbor slice info or no metadata info.\n\nThanks!\n\n",
      "votes": 2
    }
  ],
  "comments": [
    {
      "id": 674221,
      "author_name": "hqhz1817",
      "author_url": "",
      "post_date": "2019-11-16T03:00:29.397000",
      "content": "<p>I got 0.058, without using neighbouring slides ResNext50 with 5 fold, 3 window setting.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 673657,
      "author_name": "Udbhav Bamba",
      "author_url": "",
      "post_date": "2019-11-15T09:42:35.530000",
      "content": "<p>I got .057 without using neighbouring slides Efficientnet b5 5 fold</p>",
      "votes": 1,
      "replies": [
        {
          "id": 674105,
          "author_name": "Kartik Nighania",
          "author_url": "",
          "post_date": "2019-11-15T22:29:27.193000",
          "content": "<p>Can u explain more ? Thanks :)</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 674192,
          "author_name": "Udbhav Bamba",
          "author_url": "",
          "post_date": "2019-11-16T01:44:22.587000",
          "content": "<p>Sure, like my private score is just based on treating all slides as different. I trained Efficientnet B5 with 2 different preprocessings brain subdural bone and brain subdural tissue 5 fold then simply average ensemble 10 models. Actually to be honest I didn't know that we could use metadata for post processing. What was your approch? Thanks</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "674221": "I got 0.058, without using neighbouring slides ResNext50 with 5 fold, 3 window setting.",
    "673657": "I got .057 without using neighbouring slides Efficientnet b5 5 fold",
    "673404": "First, congratulations to everyone involved in stage-2 and medal winning top teams. After a brief overview of the top solutions in discussions it seems like using neighboring slices as features either with RNN models or with feature engineering is the key to boost LB scores. I thought it would be nice to capture the impact without and with neighbor slice features.\n\nMy best model scored 0.061 in Private LB - stack of 5 fold TTA preds of 2 models (eff-b0 and resnet50) - no neighbor slice info or no metadata info.\n\nThanks!\n\n"
  }
}