{
  "id": 130666,
  "title": "Any success with Test Time Augmentation?",
  "url": "/competitions/deepfake-detection-challenge/discussion/130666",
  "author_name": "",
  "post_date": "2020-02-15T16:04:38.282164100Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi, I have tried TTA with my existing models, but results got worse ☹️ Anyone who has successfully used it to improve LB? Or any general suggestions would be highly appreciated!</p>",
  "messages": [
    {
      "id": "746837",
      "postDate": "02/15/2020 16:04:38",
      "content": "<p>Hi, I have tried TTA with my existing models, but results got worse ☹️ Anyone who has successfully used it to improve LB? Or any general suggestions would be highly appreciated!</p>",
      "rawMarkdown": "Hi, I have tried TTA with my existing models, but results got worse ☹️ Anyone who has successfully used it to improve LB? Or any general suggestions would be highly appreciated!",
      "votes": null
    },
    {
      "id": "746844",
      "postDate": "02/15/2020 16:11:55",
      "content": "<p>Improved ~0.001 with hflip tta. I guess tta in this competition does not matter. But yeah, using hflip tta is a good one for balancing the predictions.</p>",
      "rawMarkdown": "Improved ~0.001 with hflip tta. I guess tta in this competition does not matter. But yeah, using hflip tta is a good one for balancing the predictions.",
      "votes": null
    },
    {
      "id": "746849",
      "postDate": "02/15/2020 16:14:20",
      "content": "<p>Thanks <a href=\"/harshitsheoran\">@harshitsheoran</a>!</p>",
      "rawMarkdown": "Thanks @harshitsheoran!",
      "votes": null
    },
    {
      "id": "747046",
      "postDate": "02/15/2020 22:03:58",
      "content": "<p>Running inference on multiple frames in the video and averaging the resulting predictions is a form of TTA that, I think, most use. It is included in the kernel of Human Analogue.</p>",
      "rawMarkdown": "Running inference on multiple frames in the video and averaging the resulting predictions is a form of TTA that, I think, most use. It is included in the kernel of Human Analogue.",
      "votes": null
    },
    {
      "id": "747056",
      "postDate": "02/15/2020 22:33:37",
      "content": "<p>Yes, I tried to combine it with some other augmentations, got worse. But, in the case of <a href=\"/harshitsheoran\">@harshitsheoran</a> flipping gave a small boost.</p>",
      "rawMarkdown": "Yes, I tried to combine it with some other augmentations, got worse. But, in the case of @harshitsheoran flipping gave a small boost.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 746844,
      "author_name": "harshitsheoran",
      "author_url": "",
      "post_date": "02/15/2020 16:11:55",
      "content": "<p>Improved ~0.001 with hflip tta. I guess tta in this competition does not matter. But yeah, using hflip tta is a good one for balancing the predictions.</p>",
      "votes": null,
      "replies": [
        {
          "id": 746849,
          "author_name": "debanga",
          "author_url": "",
          "post_date": "02/15/2020 16:14:20",
          "content": "<p>Thanks <a href=\"/harshitsheoran\">@harshitsheoran</a>!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 747046,
      "author_name": "petewills",
      "author_url": "",
      "post_date": "02/15/2020 22:03:58",
      "content": "<p>Running inference on multiple frames in the video and averaging the resulting predictions is a form of TTA that, I think, most use. It is included in the kernel of Human Analogue.</p>",
      "votes": null,
      "replies": [
        {
          "id": 747056,
          "author_name": "debanga",
          "author_url": "",
          "post_date": "02/15/2020 22:33:37",
          "content": "<p>Yes, I tried to combine it with some other augmentations, got worse. But, in the case of <a href=\"/harshitsheoran\">@harshitsheoran</a> flipping gave a small boost.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "746837": "Hi, I have tried TTA with my existing models, but results got worse ☹️ Anyone who has successfully used it to improve LB? Or any general suggestions would be highly appreciated!",
    "746844": "Improved ~0.001 with hflip tta. I guess tta in this competition does not matter. But yeah, using hflip tta is a good one for balancing the predictions.",
    "746849": "Thanks @harshitsheoran!",
    "747046": "Running inference on multiple frames in the video and averaging the resulting predictions is a form of TTA that, I think, most use. It is included in the kernel of Human Analogue.",
    "747056": "Yes, I tried to combine it with some other augmentations, got worse. But, in the case of @harshitsheoran flipping gave a small boost."
  },
  "source": "meta"
}