{
  "id": 547291,
  "title": "TTA Question",
  "url": "/competitions/czii-cryo-et-object-identification/discussion/547291",
  "author_name": "",
  "post_date": "2024-11-20T17:00:35.668026700Z",
  "votes": 5,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I encountered some issues while using tta, and when I used methods such as flipping for inference enhancement, my lb was only 0.009. Therefore, I tested the results of tta locally and found that cosine similarity was very low in certain classes of some results without using it. What could be the possible reason for this?🤔</p>",
  "messages": [
    {
      "id": "3050940",
      "postDate": "11/20/2024 17:00:35",
      "content": "<p>I encountered some issues while using tta, and when I used methods such as flipping for inference enhancement, my lb was only 0.009. Therefore, I tested the results of tta locally and found that cosine similarity was very low in certain classes of some results without using it. What could be the possible reason for this?🤔</p>",
      "rawMarkdown": "I encountered some issues while using tta, and when I used methods such as flipping for inference enhancement, my lb was only 0.009. Therefore, I tested the results of tta locally and found that cosine similarity was very low in certain classes of some results without using it. What could be the possible reason for this?🤔",
      "votes": null
    },
    {
      "id": "3050978",
      "postDate": "11/20/2024 17:34:01",
      "content": "<p>I'm not confident into using anything but rot90 since there is too few pixels to transform. Flips should be ok but I had not good results neither, may be stereochemistry.</p>",
      "rawMarkdown": "I'm not confident into using anything but rot90 since there is too few pixels to transform. Flips should be ok but I had not good results neither, may be stereochemistry.",
      "votes": null
    },
    {
      "id": "3050982",
      "postDate": "11/20/2024 17:37:46",
      "content": "<p>bug in your code. results should follow your local validation, what s your local cv for original, tta, original +tta? <br>\ntta may be slightly worse because the train augmentation (i.e. augmented data appears less). the difference is only 0.01~0.02</p>",
      "rawMarkdown": "bug in your code. results should follow your local validation, what s your local cv for original, tta, original +tta? \ntta may be slightly worse because the train augmentation (i.e. augmented data appears less). the difference is only 0.01~0.02",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3050978,
      "author_name": "sacuscreed",
      "author_url": "",
      "post_date": "11/20/2024 17:34:01",
      "content": "<p>I'm not confident into using anything but rot90 since there is too few pixels to transform. Flips should be ok but I had not good results neither, may be stereochemistry.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3050982,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "11/20/2024 17:37:46",
      "content": "<p>bug in your code. results should follow your local validation, what s your local cv for original, tta, original +tta? <br>\ntta may be slightly worse because the train augmentation (i.e. augmented data appears less). the difference is only 0.01~0.02</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3050940": "I encountered some issues while using tta, and when I used methods such as flipping for inference enhancement, my lb was only 0.009. Therefore, I tested the results of tta locally and found that cosine similarity was very low in certain classes of some results without using it. What could be the possible reason for this?🤔",
    "3050978": "I'm not confident into using anything but rot90 since there is too few pixels to transform. Flips should be ok but I had not good results neither, may be stereochemistry.",
    "3050982": "bug in your code. results should follow your local validation, what s your local cv for original, tta, original +tta? \ntta may be slightly worse because the train augmentation (i.e. augmented data appears less). the difference is only 0.01~0.02"
  },
  "source": "meta"
}