{
  "id": 580737,
  "title": "Selection of pseudo-label processing models",
  "url": "/competitions/birdclef-2025/discussion/580737",
  "author_name": "",
  "post_date": "2025-05-26T08:13:44.743784200Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>If I use pseudo-labeling for train_soundscapes, should I use the Google SED pre-trained model or the self-trained model (LB: 0.796),I think my model scores too low and the processing will be negatively optimized.</p>",
  "messages": [
    {
      "id": "3209721",
      "postDate": "05/26/2025 08:13:44",
      "content": "<p>If I use pseudo-labeling for train_soundscapes, should I use the Google SED pre-trained model or the self-trained model (LB: 0.796),I think my model scores too low and the processing will be negatively optimized.</p>",
      "rawMarkdown": "If I use pseudo-labeling for train_soundscapes, should I use the Google SED pre-trained model or the self-trained model (LB: 0.796),I think my model scores too low and the processing will be negatively optimized.",
      "votes": null
    },
    {
      "id": "3210293",
      "postDate": "05/27/2025 02:21:51",
      "content": "<p>I generated pseudo-labels using a strong public notebook (see <a href=\"url\" target=\"_blank\">https://www.kaggle.com/code/myso1987/post-processing-with-power-adjustment-for-low-rank</a> ). <br>\nSimply incorporating those labels boosted my score from 0.818 to 0.836. There are many ways to produce pseudo-labels, and I’m eager to learn which method is truly optimal.</p>",
      "rawMarkdown": "I generated pseudo-labels using a strong public notebook (see [https://www.kaggle.com/code/myso1987/post-processing-with-power-adjustment-for-low-rank](url) ). \nSimply incorporating those labels boosted my score from 0.818 to 0.836. There are many ways to produce pseudo-labels, and I’m eager to learn which method is truly optimal.",
      "votes": null
    },
    {
      "id": "3210590",
      "postDate": "05/27/2025 12:25:47",
      "content": "<p>I have also copied the link you provided, and in the early stages, I replicated the training script according to the network structure in the script, but unfortunately the effect was not satisfactory. Are you using the model input from this inference script to predict pseudo labels?</p>",
      "rawMarkdown": "I have also copied the link you provided, and in the early stages, I replicated the training script according to the network structure in the script, but unfortunately the effect was not satisfactory. Are you using the model input from this inference script to predict pseudo labels?",
      "votes": null
    },
    {
      "id": "3210599",
      "postDate": "05/27/2025 12:38:05",
      "content": "<p>I choose tf_efficientnetv2_s_in21k pre-trained model to make firstly.</p>",
      "rawMarkdown": "I choose tf_efficientnetv2_s_in21k pre-trained model to make firstly.",
      "votes": null
    },
    {
      "id": "3210663",
      "postDate": "05/27/2025 14:32:09",
      "content": "<p>did you use soft or hard labels? and what loss?</p>\n<p>I tried with a lot of combination. lower LB(</p>\n<p>P.S. didnt try CE + logits</p>",
      "rawMarkdown": "did you use soft or hard labels? and what loss?\n\nI tried with a lot of combination. lower LB(\n\nP.S. didnt try CE + logits",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3210293,
      "author_name": "sorawww31",
      "author_url": "",
      "post_date": "05/27/2025 02:21:51",
      "content": "<p>I generated pseudo-labels using a strong public notebook (see <a href=\"url\" target=\"_blank\">https://www.kaggle.com/code/myso1987/post-processing-with-power-adjustment-for-low-rank</a> ). <br>\nSimply incorporating those labels boosted my score from 0.818 to 0.836. There are many ways to produce pseudo-labels, and I’m eager to learn which method is truly optimal.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3210590,
          "author_name": "xiayuxuan",
          "author_url": "",
          "post_date": "05/27/2025 12:25:47",
          "content": "<p>I have also copied the link you provided, and in the early stages, I replicated the training script according to the network structure in the script, but unfortunately the effect was not satisfactory. Are you using the model input from this inference script to predict pseudo labels?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 3210599,
          "author_name": "xiayuxuan",
          "author_url": "",
          "post_date": "05/27/2025 12:38:05",
          "content": "<p>I choose tf_efficientnetv2_s_in21k pre-trained model to make firstly.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 3210663,
          "author_name": "player77",
          "author_url": "",
          "post_date": "05/27/2025 14:32:09",
          "content": "<p>did you use soft or hard labels? and what loss?</p>\n<p>I tried with a lot of combination. lower LB(</p>\n<p>P.S. didnt try CE + logits</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3209721": "If I use pseudo-labeling for train_soundscapes, should I use the Google SED pre-trained model or the self-trained model (LB: 0.796),I think my model scores too low and the processing will be negatively optimized.",
    "3210293": "I generated pseudo-labels using a strong public notebook (see [https://www.kaggle.com/code/myso1987/post-processing-with-power-adjustment-for-low-rank](url) ). \nSimply incorporating those labels boosted my score from 0.818 to 0.836. There are many ways to produce pseudo-labels, and I’m eager to learn which method is truly optimal.",
    "3210590": "I have also copied the link you provided, and in the early stages, I replicated the training script according to the network structure in the script, but unfortunately the effect was not satisfactory. Are you using the model input from this inference script to predict pseudo labels?",
    "3210599": "I choose tf_efficientnetv2_s_in21k pre-trained model to make firstly.",
    "3210663": "did you use soft or hard labels? and what loss?\n\nI tried with a lot of combination. lower LB(\n\nP.S. didnt try CE + logits"
  },
  "source": "meta"
}