{
  "id": 139502,
  "title": "single image models - Average score worst?",
  "url": "/competitions/deepfake-detection-challenge/discussion/139502",
  "author_name": "",
  "post_date": "2020-03-29T04:13:30.578045400Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>My single image model giving good cross validation score. but at interface average of 10 frames giving LB closer to 0.6, is average is bad approach because few real and few fake frames?</p>",
  "messages": [
    {
      "id": "789868",
      "postDate": "03/29/2020 04:13:30",
      "content": "<p>My single image model giving good cross validation score. but at interface average of 10 frames giving LB closer to 0.6, is average is bad approach because few real and few fake frames?</p>",
      "rawMarkdown": "My single image model giving good cross validation score. but at interface average of 10 frames giving LB closer to 0.6, is average is bad approach because few real and few fake frames?",
      "votes": null
    },
    {
      "id": "789890",
      "postDate": "03/29/2020 04:42:46",
      "content": "<p>No, if you compare the real to fake videos frame by frame you will see that all frames (or almost all) were modified. You are probably over fitting. </p>",
      "rawMarkdown": "No, if you compare the real to fake videos frame by frame you will see that all frames (or almost all) were modified. You are probably over fitting.",
      "votes": null
    },
    {
      "id": "789892",
      "postDate": "03/29/2020 04:46:49",
      "content": "<p><a href=\"/seshurajup\">@seshurajup</a>, I also think that average is a bad approach. As of now I haven't tried <strong>weighted mean</strong> but I think it would be better than average when making a prediction.</p>\n\n<p>My idea will be that if we have <strong>10 frames</strong> where <em>8 are fake</em> and <strong>2 are real</strong> so we should give weight-age to fake and thus calculate a weighted mean which would give us a prediction nearer to 1 thus making a prediction for the video to be fake. </p>\n\n<p>But when we use average and let's say due to extremities it turn out that many a times prediction is nearer to 0.5 and therefore you are having a LB==0.6 as logloss(0.5)==0.69314.</p>\n\n<p>Please correct me if I am wrong and if anyone has any other suggestions, please let us know.</p>\n\n<p>Regards,\n<strong>Ashutosh Pande</strong>\n<a href=\"/lepusarcticus\">@lepusarcticus</a> </p>",
      "rawMarkdown": "seshurajup, I also think that average is a bad approach. As of now I haven't tried **weighted mean** but I think it would be better than average when making a prediction.\n\nMy idea will be that if we have **10 frames** where *8 are fake* and **2 are real** so we should give weight-age to fake and thus calculate a weighted mean which would give us a prediction nearer to 1 thus making a prediction for the video to be fake. \n\nBut when we use average and let's say due to extremities it turn out that many a times prediction is nearer to 0.5 and therefore you are having a LB==0.6 as logloss(0.5)==0.69314.\n\nPlease correct me if I am wrong and if anyone has any other suggestions, please let us know.\n\nRegards,\n**Ashutosh Pande**\n@lepusarcticus",
      "votes": null
    },
    {
      "id": "789923",
      "postDate": "03/29/2020 06:02:23",
      "content": "<p><a href=\"/moshel\">@moshel</a> Thanks</p>",
      "rawMarkdown": "moshel Thanks",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 789890,
      "author_name": "moshel",
      "author_url": "",
      "post_date": "03/29/2020 04:42:46",
      "content": "<p>No, if you compare the real to fake videos frame by frame you will see that all frames (or almost all) were modified. You are probably over fitting. </p>",
      "votes": null,
      "replies": [
        {
          "id": 789923,
          "author_name": "seshurajup",
          "author_url": "",
          "post_date": "03/29/2020 06:02:23",
          "content": "<p><a href=\"/moshel\">@moshel</a> Thanks</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 789892,
      "author_name": "lepusarcticus",
      "author_url": "",
      "post_date": "03/29/2020 04:46:49",
      "content": "<p><a href=\"/seshurajup\">@seshurajup</a>, I also think that average is a bad approach. As of now I haven't tried <strong>weighted mean</strong> but I think it would be better than average when making a prediction.</p>\n\n<p>My idea will be that if we have <strong>10 frames</strong> where <em>8 are fake</em> and <strong>2 are real</strong> so we should give weight-age to fake and thus calculate a weighted mean which would give us a prediction nearer to 1 thus making a prediction for the video to be fake. </p>\n\n<p>But when we use average and let's say due to extremities it turn out that many a times prediction is nearer to 0.5 and therefore you are having a LB==0.6 as logloss(0.5)==0.69314.</p>\n\n<p>Please correct me if I am wrong and if anyone has any other suggestions, please let us know.</p>\n\n<p>Regards,\n<strong>Ashutosh Pande</strong>\n<a href=\"/lepusarcticus\">@lepusarcticus</a> </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "789868": "My single image model giving good cross validation score. but at interface average of 10 frames giving LB closer to 0.6, is average is bad approach because few real and few fake frames?",
    "789890": "No, if you compare the real to fake videos frame by frame you will see that all frames (or almost all) were modified. You are probably over fitting.",
    "789892": "seshurajup, I also think that average is a bad approach. As of now I haven't tried **weighted mean** but I think it would be better than average when making a prediction.\n\nMy idea will be that if we have **10 frames** where *8 are fake* and **2 are real** so we should give weight-age to fake and thus calculate a weighted mean which would give us a prediction nearer to 1 thus making a prediction for the video to be fake. \n\nBut when we use average and let's say due to extremities it turn out that many a times prediction is nearer to 0.5 and therefore you are having a LB==0.6 as logloss(0.5)==0.69314.\n\nPlease correct me if I am wrong and if anyone has any other suggestions, please let us know.\n\nRegards,\n**Ashutosh Pande**\n@lepusarcticus",
    "789923": "moshel Thanks"
  },
  "source": "meta"
}