{
  "id": 650742,
  "title": "how batch size weight in model training?",
  "url": "/competitions/vesuvius-challenge-surface-detection/discussion/650742",
  "author_name": "",
  "post_date": "2025-12-02T06:49:35.610853700Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>changing imagsize larger leads to OOM on kaggle. </p>\n<p>what  image size and batch size   for better model if enough memory available?</p>",
  "messages": [
    {
      "id": "3360422",
      "postDate": "12/02/2025 06:49:35",
      "content": "<p>changing imagsize larger leads to OOM on kaggle. </p>\n<p>what  image size and batch size   for better model if enough memory available?</p>",
      "rawMarkdown": "changing imagsize larger leads to OOM on kaggle. \n\nwhat  image size and batch size   for better model if enough memory available?",
      "votes": null
    },
    {
      "id": "3360731",
      "postDate": "12/02/2025 14:09:29",
      "content": "<p>If you train with PyTorch Lightning, you can use accumulated gradient which scale up the virtual batch, check it <a href=\"https://www.kaggle.com/code/jirkaborovec/surface-train-inference-slice-2d-segmentation\" target=\"_blank\">here</a>… </p>",
      "rawMarkdown": "If you train with PyTorch Lightning, you can use accumulated gradient which scale up the virtual batch, check it [here](https://www.kaggle.com/code/jirkaborovec/surface-train-inference-slice-2d-segmentation)...",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3360731,
      "author_name": "jirkaborovec",
      "author_url": "",
      "post_date": "12/02/2025 14:09:29",
      "content": "<p>If you train with PyTorch Lightning, you can use accumulated gradient which scale up the virtual batch, check it <a href=\"https://www.kaggle.com/code/jirkaborovec/surface-train-inference-slice-2d-segmentation\" target=\"_blank\">here</a>… </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3360422": "changing imagsize larger leads to OOM on kaggle. \n\nwhat  image size and batch size   for better model if enough memory available?",
    "3360731": "If you train with PyTorch Lightning, you can use accumulated gradient which scale up the virtual batch, check it [here](https://www.kaggle.com/code/jirkaborovec/surface-train-inference-slice-2d-segmentation)..."
  },
  "source": "meta"
}