{
  "id": 559649,
  "title": "Training of 3DUNET using synthetic data",
  "url": "/competitions/czii-cryo-et-object-identification/discussion/559649",
  "author_name": "",
  "post_date": "2025-01-27T02:54:01.770341700Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I have successfully trained YOLO using synthetic data. However, using synthetic data on 3DUNET resulted in lower scores. Does anyone else have a similar situation?</p>",
  "messages": [
    {
      "id": "3107851",
      "postDate": "01/27/2025 02:54:01",
      "content": "<p>I have successfully trained YOLO using synthetic data. However, using synthetic data on 3DUNET resulted in lower scores. Does anyone else have a similar situation?</p>",
      "rawMarkdown": "I have successfully trained YOLO using synthetic data. However, using synthetic data on 3DUNET resulted in lower scores. Does anyone else have a similar situation?",
      "votes": null
    },
    {
      "id": "3107874",
      "postDate": "01/27/2025 04:01:16",
      "content": "<p>The synthetic data is different than the denoised data we have in the competition.  I recommend pre training on the synthetic data and then train on the competition data with the starting point from the synthetic</p>",
      "rawMarkdown": "The synthetic data is different than the denoised data we have in the competition.  I recommend pre training on the synthetic data and then train on the competition data with the starting point from the synthetic",
      "votes": null
    },
    {
      "id": "3107880",
      "postDate": "01/27/2025 04:13:43",
      "content": "<p>Essentially train on only synthetic data -&gt; save the weights -&gt; load the weights and train on competition data if that makes sense.  Let me know if you’re already doing that</p>",
      "rawMarkdown": "Essentially train on only synthetic data -> save the weights -> load the weights and train on competition data if that makes sense.  Let me know if you’re already doing that",
      "votes": null
    },
    {
      "id": "3107884",
      "postDate": "01/27/2025 04:18:28",
      "content": "<p>Thanks for the great idea. I was using the synthetic data with noise removal and the data provided in the competition together for training. I would like to train with the pre-training method.</p>",
      "rawMarkdown": "Thanks for the great idea. I was using the synthetic data with noise removal and the data provided in the competition together for training. I would like to train with the pre-training method.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3107874,
      "author_name": "connorjd",
      "author_url": "",
      "post_date": "01/27/2025 04:01:16",
      "content": "<p>The synthetic data is different than the denoised data we have in the competition.  I recommend pre training on the synthetic data and then train on the competition data with the starting point from the synthetic</p>",
      "votes": null,
      "replies": [
        {
          "id": 3107880,
          "author_name": "connorjd",
          "author_url": "",
          "post_date": "01/27/2025 04:13:43",
          "content": "<p>Essentially train on only synthetic data -&gt; save the weights -&gt; load the weights and train on competition data if that makes sense.  Let me know if you’re already doing that</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 3107884,
          "author_name": "yoshinarikawashima",
          "author_url": "",
          "post_date": "01/27/2025 04:18:28",
          "content": "<p>Thanks for the great idea. I was using the synthetic data with noise removal and the data provided in the competition together for training. I would like to train with the pre-training method.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3107851": "I have successfully trained YOLO using synthetic data. However, using synthetic data on 3DUNET resulted in lower scores. Does anyone else have a similar situation?",
    "3107874": "The synthetic data is different than the denoised data we have in the competition.  I recommend pre training on the synthetic data and then train on the competition data with the starting point from the synthetic",
    "3107880": "Essentially train on only synthetic data -> save the weights -> load the weights and train on competition data if that makes sense.  Let me know if you’re already doing that",
    "3107884": "Thanks for the great idea. I was using the synthetic data with noise removal and the data provided in the competition together for training. I would like to train with the pre-training method."
  },
  "source": "meta"
}