{
  "id": 133748,
  "title": "What about clipping?",
  "url": "/competitions/deepfake-detection-challenge/discussion/133748",
  "author_name": "Ankit Saini",
  "post_date": "2020-03-04T04:38:51.737000",
  "votes": 1,
  "comment_count": 14,
  "views": 0,
  "content": "<p>Recently I made a submission and it scored 0.68 on lb but just after clipping of 0.1 I got 0.48 at lb. </p>\n\n<p>Maybe some of you're missing your best models due to clipping. \nPlease share any insights on how you're dealing with it.</p>",
  "messages": [
    {
      "id": 763528,
      "postDate": "2020-03-04T15:15:19.893Z",
      "content": "<p>0.99 and 0.01 clip is the best one I could suggest if your score is below 0.36, else 0.05 and 0.95 is a good one too.</p>",
      "rawMarkdown": "0.99 and 0.01 clip is the best one I could suggest if your score is below 0.36, else 0.05 and 0.95 is a good one too.",
      "votes": 3,
      "replies": [
        {
          "id": 763572,
          "postDate": "2020-03-04T16:03:43.947Z",
          "content": "<p><a href=\"/harshitsheoran\">@harshitsheoran</a> , Whats the val score you are getting on your model? Not the LB score.</p>",
          "rawMarkdown": "@harshitsheoran , Whats the val score you are getting on your model? Not the LB score."
        },
        {
          "id": 763614,
          "postDate": "2020-03-04T17:00:10.703Z",
          "content": "<p>best benchmark score was 0.16, which was 0.31 on leaderboard.</p>",
          "rawMarkdown": "best benchmark score was 0.16, which was 0.31 on leaderboard.",
          "votes": 1
        },
        {
          "id": 769876,
          "postDate": "2020-03-12T10:54:22.923Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 769892,
          "postDate": "2020-03-12T11:18:11.083Z",
          "content": "<p>make sure pred is in form of <code>(num_preds, 1)</code> or <code>(num_preds, )</code>\n<code>preds = np.clip(preds, a_max = 0.95, a_min = 0.05)</code></p>",
          "rawMarkdown": "make sure pred is in form of `(num_preds, 1)` or `(num_preds, )`\n`preds = np.clip(preds, a_max = 0.95, a_min = 0.05)`",
          "votes": 1
        }
      ]
    },
    {
      "id": 763498,
      "postDate": "2020-03-04T14:53:20.157Z",
      "content": "<p>when my LB score is lower,such as 0.33, clipping will not implement your score.</p>",
      "rawMarkdown": "when my LB score is lower,such as 0.33, clipping will not implement your score.",
      "votes": 1,
      "replies": [
        {
          "id": 769872,
          "postDate": "2020-03-12T10:53:03.200Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 769891,
          "postDate": "2020-03-12T11:16:54.253Z",
          "content": "<p>make sure pred is in form of <code>(num_preds, 1)</code> or <code>(num_preds, )</code>\n<code>preds = np.clip(preds, a_max = 0.95, a_min = 0.05)</code></p>",
          "rawMarkdown": "make sure pred is in form of `(num_preds, 1)` or `(num_preds, )`\n`preds = np.clip(preds, a_max = 0.95, a_min = 0.05)`"
        },
        {
          "id": 769894,
          "postDate": "2020-03-12T11:20:07.373Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 770177,
          "postDate": "2020-03-12T16:38:15.350Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 770216,
          "postDate": "2020-03-12T17:32:04.493Z",
          "content": "<p>are you clipping on 1 and 0, also sometimes (in very very) rare occasion for this competition, that is theoratlically possible, try out different clipps, like too low or too high, recommended is from 0.85/0.15 to 0.99/0.01</p>",
          "rawMarkdown": "are you clipping on 1 and 0, also sometimes (in very very) rare occasion for this competition, that is theoratlically possible, try out different clipps, like too low or too high, recommended is from 0.85/0.15 to 0.99/0.01"
        },
        {
          "id": 770226,
          "postDate": "2020-03-12T17:49:59.650Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 763053,
      "postDate": "2020-03-04T04:38:51.737Z",
      "content": "<p>Recently I made a submission and it scored 0.68 on lb but just after clipping of 0.1 I got 0.48 at lb. </p>\n\n<p>Maybe some of you're missing your best models due to clipping. \nPlease share any insights on how you're dealing with it.</p>",
      "rawMarkdown": "Recently I made a submission and it scored 0.68 on lb but just after clipping of 0.1 I got 0.48 at lb. \n\nMaybe some of you're missing your best models due to clipping. \nPlease share any insights on how you're dealing with it.",
      "votes": 1
    },
    {
      "id": 769880,
      "postDate": "2020-03-12T11:00:18.463Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 769869,
      "postDate": "2020-03-12T10:51:57.070Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 763528,
      "author_name": "Harshit Sheoran",
      "author_url": "",
      "post_date": "2020-03-04T15:15:19.893000",
      "content": "<p>0.99 and 0.01 clip is the best one I could suggest if your score is below 0.36, else 0.05 and 0.95 is a good one too.</p>",
      "votes": 3,
      "replies": [
        {
          "id": 763572,
          "author_name": "Akash",
          "author_url": "",
          "post_date": "2020-03-04T16:03:43.947000",
          "content": "<p><a href=\"/harshitsheoran\">@harshitsheoran</a> , Whats the val score you are getting on your model? Not the LB score.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 763614,
          "author_name": "Harshit Sheoran",
          "author_url": "",
          "post_date": "2020-03-04T17:00:10.703000",
          "content": "<p>best benchmark score was 0.16, which was 0.31 on leaderboard.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 769876,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-03-12T10:54:22.923000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 769892,
          "author_name": "Harshit Sheoran",
          "author_url": "",
          "post_date": "2020-03-12T11:18:11.083000",
          "content": "<p>make sure pred is in form of <code>(num_preds, 1)</code> or <code>(num_preds, )</code>\n<code>preds = np.clip(preds, a_max = 0.95, a_min = 0.05)</code></p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 763498,
      "author_name": "BokingChen",
      "author_url": "",
      "post_date": "2020-03-04T14:53:20.157000",
      "content": "<p>when my LB score is lower,such as 0.33, clipping will not implement your score.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 769872,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-03-12T10:53:03.200000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 769891,
          "author_name": "Harshit Sheoran",
          "author_url": "",
          "post_date": "2020-03-12T11:16:54.253000",
          "content": "<p>make sure pred is in form of <code>(num_preds, 1)</code> or <code>(num_preds, )</code>\n<code>preds = np.clip(preds, a_max = 0.95, a_min = 0.05)</code></p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 769894,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-03-12T11:20:07.373000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 770177,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-03-12T16:38:15.350000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 770216,
          "author_name": "Harshit Sheoran",
          "author_url": "",
          "post_date": "2020-03-12T17:32:04.493000",
          "content": "<p>are you clipping on 1 and 0, also sometimes (in very very) rare occasion for this competition, that is theoratlically possible, try out different clipps, like too low or too high, recommended is from 0.85/0.15 to 0.99/0.01</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 770226,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-03-12T17:49:59.650000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 769880,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-03-12T11:00:18.463000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 769869,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-03-12T10:51:57.070000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "763528": "0.99 and 0.01 clip is the best one I could suggest if your score is below 0.36, else 0.05 and 0.95 is a good one too.",
    "763498": "when my LB score is lower,such as 0.33, clipping will not implement your score.",
    "763053": "Recently I made a submission and it scored 0.68 on lb but just after clipping of 0.1 I got 0.48 at lb. \n\nMaybe some of you're missing your best models due to clipping. \nPlease share any insights on how you're dealing with it.",
    "769880": "",
    "769869": ""
  }
}