{
  "id": 542010,
  "title": "Generating additional training data?",
  "url": "/competitions/ariel-data-challenge-2024/discussion/542010",
  "author_name": "Araik Tamazian",
  "post_date": "2024-10-22T14:58:29.719000",
  "votes": 2,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Since we have synthetic dataset here, I'm just wondering if anyone managed to generate additional training data which can be used for models training/validation.</p>",
  "messages": [
    {
      "id": 3025236,
      "postDate": "2024-10-22T15:07:30.270Z",
      "content": "<p>I'm wondering the same, do we have example code for the generation?</p>",
      "rawMarkdown": "I'm wondering the same, do we have example code for the generation?",
      "votes": 1,
      "replies": [
        {
          "id": 3025246,
          "postDate": "2024-10-22T15:14:39.503Z",
          "content": "<p>I don't have such code, only ideas. But I'm not sure about generation parameters that should be used.</p>",
          "rawMarkdown": "I don't have such code, only ideas. But I'm not sure about generation parameters that should be used.",
          "replies": [
            {
              "id": 3025264,
              "postDate": "2024-10-22T15:27:16.530Z",
              "content": "<p>Oh, and another problem - how one can get ground truth for the generated samples?</p>",
              "rawMarkdown": "Oh, and another problem - how one can get ground truth for the generated samples?"
            },
            {
              "id": 3025361,
              "postDate": "2024-10-22T17:18:33.323Z",
              "content": "<p>IDK, but it will be good to have more data</p>",
              "rawMarkdown": "IDK, but it will be good to have more data"
            },
            {
              "id": 3025478,
              "postDate": "2024-10-22T20:05:02.997Z",
              "content": "<p>I assume one can use TransitModel from the Batman package to generate curves vs time. It should be possible to vary the transit depth to mimic the data dependence vs wavelength. So you can choose any ground truth as you wish. There is a freedom to fold those distributions to mimic detector effects and any flux dependence versus time (linear, quadratic, etc). And even more, one can vary TransitModel parameters to generate more distortions and test how stable your model is against them :)</p>",
              "rawMarkdown": "I assume one can use TransitModel from the Batman package to generate curves vs time. It should be possible to vary the transit depth to mimic the data dependence vs wavelength. So you can choose any ground truth as you wish. There is a freedom to fold those distributions to mimic detector effects and any flux dependence versus time (linear, quadratic, etc). And even more, one can vary TransitModel parameters to generate more distortions and test how stable your model is against them :)",
              "votes": 6
            },
            {
              "id": 3026237,
              "postDate": "2024-10-23T14:57:12.220Z",
              "content": "<p>I've run some experiments with <code>batman</code>, and I was able to generate light curves quite similar to ones which can be obtained from the train set (by adding some noise as well).</p>",
              "rawMarkdown": "I've run some experiments with `batman`, and I was able to generate light curves quite similar to ones which can be obtained from the train set (by adding some noise as well)."
            }
          ]
        }
      ]
    },
    {
      "id": 3025227,
      "postDate": "2024-10-22T14:58:29.720Z",
      "content": "<p>Since we have synthetic dataset here, I'm just wondering if anyone managed to generate additional training data which can be used for models training/validation.</p>",
      "rawMarkdown": "Since we have synthetic dataset here, I'm just wondering if anyone managed to generate additional training data which can be used for models training/validation.",
      "votes": 2
    }
  ],
  "comments": [
    {
      "id": 3025236,
      "author_name": "yuanzhe zhou",
      "author_url": "",
      "post_date": "2024-10-22T15:07:30.270000",
      "content": "<p>I'm wondering the same, do we have example code for the generation?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3025246,
          "author_name": "Araik Tamazian",
          "author_url": "",
          "post_date": "2024-10-22T15:14:39.503000",
          "content": "<p>I don't have such code, only ideas. But I'm not sure about generation parameters that should be used.</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3025264,
              "author_name": "Araik Tamazian",
              "author_url": "",
              "post_date": "2024-10-22T15:27:16.530000",
              "content": "<p>Oh, and another problem - how one can get ground truth for the generated samples?</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3025361,
              "author_name": "yuanzhe zhou",
              "author_url": "",
              "post_date": "2024-10-22T17:18:33.323000",
              "content": "<p>IDK, but it will be good to have more data</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3025478,
              "author_name": "Oleh Kivernyk",
              "author_url": "",
              "post_date": "2024-10-22T20:05:02.997000",
              "content": "<p>I assume one can use TransitModel from the Batman package to generate curves vs time. It should be possible to vary the transit depth to mimic the data dependence vs wavelength. So you can choose any ground truth as you wish. There is a freedom to fold those distributions to mimic detector effects and any flux dependence versus time (linear, quadratic, etc). And even more, one can vary TransitModel parameters to generate more distortions and test how stable your model is against them :)</p>",
              "votes": 6,
              "replies": []
            },
            {
              "id": 3026237,
              "author_name": "Araik Tamazian",
              "author_url": "",
              "post_date": "2024-10-23T14:57:12.220000",
              "content": "<p>I've run some experiments with <code>batman</code>, and I was able to generate light curves quite similar to ones which can be obtained from the train set (by adding some noise as well).</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3025236": "I'm wondering the same, do we have example code for the generation?",
    "3025227": "Since we have synthetic dataset here, I'm just wondering if anyone managed to generate additional training data which can be used for models training/validation."
  }
}