{
  "id": 326957,
  "title": "Looks like traditional classifer still dominate this competition? How many people use few-shot model?",
  "url": "/competitions/birdclef-2022/discussion/326957",
  "author_name": "",
  "post_date": "2022-05-25T03:29:44.896678300Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>As most of the 21 socred birds has very few training samples, this year's competition looks obvious few-shot problem. <br>\nI`m wondering how your classifier overcome the serious overfitting issue? In my case I used few-shot model and RelationNet as learnable non-linear comparator,  my single model can reach 0.77 public LB, but in the end it still overfit the public LB so much.</p>",
  "messages": [
    {
      "id": "1800566",
      "postDate": "05/25/2022 03:29:44",
      "content": "<p>As most of the 21 socred birds has very few training samples, this year's competition looks obvious few-shot problem. <br>\nI`m wondering how your classifier overcome the serious overfitting issue? In my case I used few-shot model and RelationNet as learnable non-linear comparator,  my single model can reach 0.77 public LB, but in the end it still overfit the public LB so much.</p>",
      "rawMarkdown": "As most of the 21 socred birds has very few training samples, this year's competition looks obvious few-shot problem. \nI`m wondering how your classifier overcome the serious overfitting issue? In my case I used few-shot model and RelationNet as learnable non-linear comparator,  my single model can reach 0.77 public LB, but in the end it still overfit the public LB so much.",
      "votes": null
    },
    {
      "id": "1800582",
      "postDate": "05/25/2022 04:10:17",
      "content": "<p>I tried to extract features from regular classifier and then create a separate FF model/SVM model to predict on these features for only scored birds (with weighted loss), but somehow couldn't get the model to work properly due to paucity of time. I will be curious to go back and explore what went wrong. </p>",
      "rawMarkdown": "I tried to extract features from regular classifier and then create a separate FF model/SVM model to predict on these features for only scored birds (with weighted loss), but somehow couldn't get the model to work properly due to paucity of time. I will be curious to go back and explore what went wrong.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1800582,
      "author_name": "allohvk",
      "author_url": "",
      "post_date": "05/25/2022 04:10:17",
      "content": "<p>I tried to extract features from regular classifier and then create a separate FF model/SVM model to predict on these features for only scored birds (with weighted loss), but somehow couldn't get the model to work properly due to paucity of time. I will be curious to go back and explore what went wrong. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1800566": "As most of the 21 socred birds has very few training samples, this year's competition looks obvious few-shot problem. \nI`m wondering how your classifier overcome the serious overfitting issue? In my case I used few-shot model and RelationNet as learnable non-linear comparator,  my single model can reach 0.77 public LB, but in the end it still overfit the public LB so much.",
    "1800582": "I tried to extract features from regular classifier and then create a separate FF model/SVM model to predict on these features for only scored birds (with weighted loss), but somehow couldn't get the model to work properly due to paucity of time. I will be curious to go back and explore what went wrong."
  },
  "source": "meta"
}