{
  "id": 172407,
  "title": "pseudo labelling with soft labels leads to bad results?",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/172407",
  "author_name": "",
  "post_date": "2020-08-05T00:23:40.503762200Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Has anybody tried pseudo labelling with soft labels? </p>\n\n<p>I continue to see worse results if I use soft pseudo labelled test set to my training data. </p>\n\n<p>Wondering if this is the case for everyone?</p>\n\n<p>Here's the process:\n- Train a model\n- Predict on test using best model\n- Add soft preditions to train \n- Train new model with soft predictions added to train set \n- Finetune without pseudo labels (optional)\n- Predict on test</p>\n\n<p>I see pretty bad scores when I try pseudo labelling. </p>",
  "messages": [
    {
      "id": "958359",
      "postDate": "08/05/2020 00:23:40",
      "content": "<p>Has anybody tried pseudo labelling with soft labels? </p>\n\n<p>I continue to see worse results if I use soft pseudo labelled test set to my training data. </p>\n\n<p>Wondering if this is the case for everyone?</p>\n\n<p>Here's the process:\n- Train a model\n- Predict on test using best model\n- Add soft preditions to train \n- Train new model with soft predictions added to train set \n- Finetune without pseudo labels (optional)\n- Predict on test</p>\n\n<p>I see pretty bad scores when I try pseudo labelling. </p>",
      "rawMarkdown": "Has anybody tried pseudo labelling with soft labels? \n\nI continue to see worse results if I use soft pseudo labelled test set to my training data. \n\nWondering if this is the case for everyone?\n\nHere's the process:\n- Train a model\n- Predict on test using best model\n- Add soft preditions to train \n- Train new model with soft predictions added to train set \n- Finetune without pseudo labels (optional)\n- Predict on test\n\nI see pretty bad scores when I try pseudo labelling.",
      "votes": null
    },
    {
      "id": "958432",
      "postDate": "08/05/2020 01:39:09",
      "content": "<p>I have tried using pseudo labels (hard labeled 0/1). Nothing bad happened. I don't recommend it unless you know what you're doing.</p>",
      "rawMarkdown": "I have tried using pseudo labels (hard labeled 0/1). Nothing bad happened. I don't recommend it unless you know what you're doing.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 958432,
      "author_name": "teeyee314",
      "author_url": "",
      "post_date": "08/05/2020 01:39:09",
      "content": "<p>I have tried using pseudo labels (hard labeled 0/1). Nothing bad happened. I don't recommend it unless you know what you're doing.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "958359": "Has anybody tried pseudo labelling with soft labels? \n\nI continue to see worse results if I use soft pseudo labelled test set to my training data. \n\nWondering if this is the case for everyone?\n\nHere's the process:\n- Train a model\n- Predict on test using best model\n- Add soft preditions to train \n- Train new model with soft predictions added to train set \n- Finetune without pseudo labels (optional)\n- Predict on test\n\nI see pretty bad scores when I try pseudo labelling.",
    "958432": "I have tried using pseudo labels (hard labeled 0/1). Nothing bad happened. I don't recommend it unless you know what you're doing."
  },
  "source": "meta"
}