{
  "id": 130734,
  "title": "Pseudo-Label",
  "url": "/competitions/flower-classification-with-tpus/discussion/130734",
  "author_name": "",
  "post_date": "2020-02-16T02:35:44.851304900Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Have someone used the Pseudo-Label. How to use it?</p>",
  "messages": [
    {
      "id": "747140",
      "postDate": "02/16/2020 02:35:44",
      "content": "<p>Have someone used the Pseudo-Label. How to use it?</p>",
      "rawMarkdown": "Have someone used the Pseudo-Label. How to use it?",
      "votes": null
    },
    {
      "id": "747549",
      "postDate": "02/16/2020 15:03:46",
      "content": "<p>In general, we have two ways to use pseudo-label. \nA. Use the test set as pseudo-label .  For this purpose, we should use our best model to predict the test set and get the target prediction.  We use these target prediction as the pseudo labels of test set .  In the end ,we add the test set with the pseudo labels  into our train data.</p>\n\n<p>For your reference, you can see <a href=\"/cdeotte\">@cdeotte</a> 's great  <a href=\"https://www.kaggle.com/cdeotte/pseudo-labeling-qda-0-969\">kernel.</a> </p>\n\n<p>B. Find other external datas as  pseudo-label .  Then do the same works as A (Use the test set as pseudo-label).  </p>",
      "rawMarkdown": "In general, we have two ways to use pseudo-label. \nA. Use the test set as pseudo-label .  For this purpose, we should use our best model to predict the test set and get the target prediction.  We use these target prediction as the pseudo labels of test set .  In the end ,we add the test set with the pseudo labels  into our train data.\n\nFor your reference, you can see @cdeotte 's great  [kernel.](https://www.kaggle.com/cdeotte/pseudo-labeling-qda-0-969) \n\nB. Find other external datas as  pseudo-label .  Then do the same works as A (Use the test set as pseudo-label).",
      "votes": null
    },
    {
      "id": "749618",
      "postDate": "02/18/2020 20:15:16",
      "content": "<p>Great idea to try semi-supervised training. The test dataset is large here so this might actually work. On the other hand, the problematic classes are the ones with very few entries. In any case, it is worth trying. I hope you will share your Notebook if you get it to work.</p>",
      "rawMarkdown": "Great idea to try semi-supervised training. The test dataset is large here so this might actually work. On the other hand, the problematic classes are the ones with very few entries. In any case, it is worth trying. I hope you will share your Notebook if you get it to work.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 747549,
      "author_name": "qinhui1999",
      "author_url": "",
      "post_date": "02/16/2020 15:03:46",
      "content": "<p>In general, we have two ways to use pseudo-label. \nA. Use the test set as pseudo-label .  For this purpose, we should use our best model to predict the test set and get the target prediction.  We use these target prediction as the pseudo labels of test set .  In the end ,we add the test set with the pseudo labels  into our train data.</p>\n\n<p>For your reference, you can see <a href=\"/cdeotte\">@cdeotte</a> 's great  <a href=\"https://www.kaggle.com/cdeotte/pseudo-labeling-qda-0-969\">kernel.</a> </p>\n\n<p>B. Find other external datas as  pseudo-label .  Then do the same works as A (Use the test set as pseudo-label).  </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 749618,
      "author_name": "mgorner",
      "author_url": "",
      "post_date": "02/18/2020 20:15:16",
      "content": "<p>Great idea to try semi-supervised training. The test dataset is large here so this might actually work. On the other hand, the problematic classes are the ones with very few entries. In any case, it is worth trying. I hope you will share your Notebook if you get it to work.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "747140": "Have someone used the Pseudo-Label. How to use it?",
    "747549": "In general, we have two ways to use pseudo-label. \nA. Use the test set as pseudo-label .  For this purpose, we should use our best model to predict the test set and get the target prediction.  We use these target prediction as the pseudo labels of test set .  In the end ,we add the test set with the pseudo labels  into our train data.\n\nFor your reference, you can see @cdeotte 's great  [kernel.](https://www.kaggle.com/cdeotte/pseudo-labeling-qda-0-969) \n\nB. Find other external datas as  pseudo-label .  Then do the same works as A (Use the test set as pseudo-label).",
    "749618": "Great idea to try semi-supervised training. The test dataset is large here so this might actually work. On the other hand, the problematic classes are the ones with very few entries. In any case, it is worth trying. I hope you will share your Notebook if you get it to work."
  },
  "source": "meta"
}