{
  "id": 165664,
  "title": "Adding TTA improved my base model, but after Pseudo Labaling LB decreased",
  "url": "/competitions/global-wheat-detection/discussion/165664",
  "author_name": "",
  "post_date": "2020-07-10T14:10:48.838683600Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hey everyone, Im trying to improve my yolov5 model,\nAnd right now my base model gets 0.747~, and after pseudo it gets 0.765+.\nI tryed to add some more TTA functions I made by myself, which did improve my base model from 0.747 to 0.75+, but after Pseudo I get 0.760 which is lower than I expected,\nBecause as it improved my base model, I was sure it gonna improve it after Pseudo also, And I can't find why it decrease it.\nAnyone knows maybe what could be the reason for that? \nThanks 🙂 </p>",
  "messages": [
    {
      "id": "923072",
      "postDate": "07/10/2020 14:10:48",
      "content": "<p>Hey everyone, Im trying to improve my yolov5 model,\nAnd right now my base model gets 0.747~, and after pseudo it gets 0.765+.\nI tryed to add some more TTA functions I made by myself, which did improve my base model from 0.747 to 0.75+, but after Pseudo I get 0.760 which is lower than I expected,\nBecause as it improved my base model, I was sure it gonna improve it after Pseudo also, And I can't find why it decrease it.\nAnyone knows maybe what could be the reason for that? \nThanks 🙂 </p>",
      "rawMarkdown": "Hey everyone, Im trying to improve my yolov5 model,\nAnd right now my base model gets 0.747~, and after pseudo it gets 0.765+.\nI tryed to add some more TTA functions I made by myself, which did improve my base model from 0.747 to 0.75+, but after Pseudo I get 0.760 which is lower than I expected,\nBecause as it improved my base model, I was sure it gonna improve it after Pseudo also, And I can't find why it decrease it.\nAnyone knows maybe what could be the reason for that? \nThanks 🙂",
      "votes": null
    },
    {
      "id": "923242",
      "postDate": "07/10/2020 16:32:55",
      "content": "<p>I'm facing the same too. I am thinking of using the new TTA for final preds and the old TTA to generate pseudo labels. Let's see if that works out..</p>",
      "rawMarkdown": "I'm facing the same too. I am thinking of using the new TTA for final preds and the old TTA to generate pseudo labels. Let's see if that works out..",
      "votes": null
    },
    {
      "id": "923252",
      "postDate": "07/10/2020 16:40:32",
      "content": "<p><a href=\"/mightyrains\">@mightyrains</a>  I tried doing the opposite, I used my new TTA functions at generating pseudo labels, but using the old TTA at final predictions.\nThe results were a bit better but not better than not using my new TTA at all.\nPlease tell me if you get any improvement by doing what you  just said :)</p>",
      "rawMarkdown": "mightyrains  I tried doing the opposite, I used my new TTA functions at generating pseudo labels, but using the old TTA at final predictions.\nThe results were a bit better but not better than not using my new TTA at all.\nPlease tell me if you get any improvement by doing what you  just said :)",
      "votes": null
    },
    {
      "id": "923255",
      "postDate": "07/10/2020 16:42:11",
      "content": "<p>Sure thing :)</p>",
      "rawMarkdown": "Sure thing :)",
      "votes": null
    },
    {
      "id": "926211",
      "postDate": "07/12/2020 15:21:39",
      "content": "<p>Could it be due to target leakage between the pseudo labels and your validation data? </p>",
      "rawMarkdown": "Could it be due to target leakage between the pseudo labels and your validation data?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 923242,
      "author_name": "mightyrains",
      "author_url": "",
      "post_date": "07/10/2020 16:32:55",
      "content": "<p>I'm facing the same too. I am thinking of using the new TTA for final preds and the old TTA to generate pseudo labels. Let's see if that works out..</p>",
      "votes": null,
      "replies": [
        {
          "id": 923252,
          "author_name": "dudkmanron",
          "author_url": "",
          "post_date": "07/10/2020 16:40:32",
          "content": "<p><a href=\"/mightyrains\">@mightyrains</a>  I tried doing the opposite, I used my new TTA functions at generating pseudo labels, but using the old TTA at final predictions.\nThe results were a bit better but not better than not using my new TTA at all.\nPlease tell me if you get any improvement by doing what you  just said :)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 923255,
          "author_name": "mightyrains",
          "author_url": "",
          "post_date": "07/10/2020 16:42:11",
          "content": "<p>Sure thing :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 926211,
      "author_name": "kennethrithvik",
      "author_url": "",
      "post_date": "07/12/2020 15:21:39",
      "content": "<p>Could it be due to target leakage between the pseudo labels and your validation data? </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "923072": "Hey everyone, Im trying to improve my yolov5 model,\nAnd right now my base model gets 0.747~, and after pseudo it gets 0.765+.\nI tryed to add some more TTA functions I made by myself, which did improve my base model from 0.747 to 0.75+, but after Pseudo I get 0.760 which is lower than I expected,\nBecause as it improved my base model, I was sure it gonna improve it after Pseudo also, And I can't find why it decrease it.\nAnyone knows maybe what could be the reason for that? \nThanks 🙂",
    "923242": "I'm facing the same too. I am thinking of using the new TTA for final preds and the old TTA to generate pseudo labels. Let's see if that works out..",
    "923252": "mightyrains  I tried doing the opposite, I used my new TTA functions at generating pseudo labels, but using the old TTA at final predictions.\nThe results were a bit better but not better than not using my new TTA at all.\nPlease tell me if you get any improvement by doing what you  just said :)",
    "923255": "Sure thing :)",
    "926211": "Could it be due to target leakage between the pseudo labels and your validation data?"
  },
  "source": "meta"
}