{
  "id": 110352,
  "title": "How did you use the control experiments?",
  "url": "/competitions/recursion-cellular-image-classification/discussion/110352",
  "author_name": "",
  "post_date": "2019-09-27T02:42:14.777493500Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>How did you use the control images? Were they useful to your results?</p>\n\n<p>We trained with all control images to let the model \"see\" all experiments. We also generate control image features using the same embedding model and compare them to the actual images. </p>",
  "messages": [
    {
      "id": "634980",
      "postDate": "09/27/2019 02:42:14",
      "content": "<p>How did you use the control images? Were they useful to your results?</p>\n\n<p>We trained with all control images to let the model \"see\" all experiments. We also generate control image features using the same embedding model and compare them to the actual images. </p>",
      "rawMarkdown": "How did you use the control images? Were they useful to your results?\n\nWe trained with all control images to let the model \"see\" all experiments. We also generate control image features using the same embedding model and compare them to the actual images.",
      "votes": null
    },
    {
      "id": "634983",
      "postDate": "09/27/2019 02:47:38",
      "content": "<p>Yes. I reported the way to use control images in <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110337#634969\">here</a> </p>\n\n<ul>\n<li>Pretraining model with the control images gives us 2% increment. </li>\n<li>Training model with 1139 (1108 + 31) gives us a little diversity. </li>\n</ul>",
      "rawMarkdown": "Yes. I reported the way to use control images in [here](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110337#634969) \n\n- Pretraining model with the control images gives us 2% increment. \n- Training model with 1139 (1108 + 31) gives us a little diversity.",
      "votes": null
    },
    {
      "id": "634993",
      "postDate": "09/27/2019 03:12:17",
      "content": "<p>amazing, that's is the magic of control images. I used it the wrong way and hesitant to retrain again the other way due to time constraint! :-(</p>",
      "rawMarkdown": "amazing, that's is the magic of control images. I used it the wrong way and hesitant to retrain again the other way due to time constraint! :-(",
      "votes": null
    },
    {
      "id": "635777",
      "postDate": "09/28/2019 07:26:08",
      "content": "<p>I also tried pretraining on control set, but I trained separate model first with 31 classes and than finetuned it replacing the output layer. But this approach scored slightly less on LB </p>",
      "rawMarkdown": "I also tried pretraining on control set, but I trained separate model first with 31 classes and than finetuned it replacing the output layer. But this approach scored slightly less on LB",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 634983,
      "author_name": "backaggle",
      "author_url": "",
      "post_date": "09/27/2019 02:47:38",
      "content": "<p>Yes. I reported the way to use control images in <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110337#634969\">here</a> </p>\n\n<ul>\n<li>Pretraining model with the control images gives us 2% increment. </li>\n<li>Training model with 1139 (1108 + 31) gives us a little diversity. </li>\n</ul>",
      "votes": null,
      "replies": [
        {
          "id": 634993,
          "author_name": "projdev",
          "author_url": "",
          "post_date": "09/27/2019 03:12:17",
          "content": "<p>amazing, that's is the magic of control images. I used it the wrong way and hesitant to retrain again the other way due to time constraint! :-(</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 635777,
      "author_name": "cateek",
      "author_url": "",
      "post_date": "09/28/2019 07:26:08",
      "content": "<p>I also tried pretraining on control set, but I trained separate model first with 31 classes and than finetuned it replacing the output layer. But this approach scored slightly less on LB </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "634980": "How did you use the control images? Were they useful to your results?\n\nWe trained with all control images to let the model \"see\" all experiments. We also generate control image features using the same embedding model and compare them to the actual images.",
    "634983": "Yes. I reported the way to use control images in [here](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110337#634969) \n\n- Pretraining model with the control images gives us 2% increment. \n- Training model with 1139 (1108 + 31) gives us a little diversity.",
    "634993": "amazing, that's is the magic of control images. I used it the wrong way and hesitant to retrain again the other way due to time constraint! :-(",
    "635777": "I also tried pretraining on control set, but I trained separate model first with 31 classes and than finetuned it replacing the output layer. But this approach scored slightly less on LB"
  },
  "source": "meta"
}