{
  "id": 378063,
  "title": "Do high scored notebook reproduce in kaggle hardware?",
  "url": "/competitions/rsna-breast-cancer-detection/discussion/378063",
  "author_name": "Simon Alerdic",
  "post_date": "2023-01-14T08:02:32.867000",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>All the high scored notebook (above .4) are trained from local hardware. Have anyone successfully reproduce them in kaggle system? The possible issue might be the model size, image size and batch size limitation. </p>",
  "messages": [
    {
      "id": 2099160,
      "postDate": "2023-01-14T08:02:32.867Z",
      "content": "<p>All the high scored notebook (above .4) are trained from local hardware. Have anyone successfully reproduce them in kaggle system? The possible issue might be the model size, image size and batch size limitation. </p>",
      "rawMarkdown": "All the high scored notebook (above .4) are trained from local hardware. Have anyone successfully reproduce them in kaggle system? The possible issue might be the model size, image size and batch size limitation. ",
      "votes": 2
    },
    {
      "id": 2099319,
      "postDate": "2023-01-14T10:44:13.767Z",
      "content": "<p>IMO, the answer are: </p>\n<ul>\n<li>Yes: you can train a model with high score in LB with Kaggle System (GPU P100, 2xT4), but you need to apply some tricky solutions like: gradient accumulation, freeze/unfreeze model, … to fit your GPUs.</li>\n<li>No: Unfortunately, this competition requires a lot of experiments, you need to test out your ideas which are enormous. Besides, you also need to finetune your params (a lot) to gain the best score, so in the end, you will need a powerful local hardware (harsh truth).</li>\n</ul>",
      "rawMarkdown": "IMO, the answer are: \n- Yes: you can train a model with high score in LB with Kaggle System (GPU P100, 2xT4), but you need to apply some tricky solutions like: gradient accumulation, freeze/unfreeze model, ... to fit your GPUs.\n- No: Unfortunately, this competition requires a lot of experiments, you need to test out your ideas which are enormous. Besides, you also need to finetune your params (a lot) to gain the best score, so in the end, you will need a powerful local hardware (harsh truth)."
    },
    {
      "id": 2099281,
      "postDate": "2023-01-14T10:07:34.137Z",
      "content": "<p>The very first model that I trained I used the Kaggle notebook (GPU P100) and its LB score is 0.42. However, since then I have used my local GPU (RTX 2080 ti) which has less memory than GPU P100 but I managed to get higher LB scores. I believe that one can use the Kaggle's notebook to train a high performance model. The main challenge in this competition, in my opinion is how to deal with the <strong>data</strong>.</p>",
      "rawMarkdown": "The very first model that I trained I used the Kaggle notebook (GPU P100) and its LB score is 0.42. However, since then I have used my local GPU (RTX 2080 ti) which has less memory than GPU P100 but I managed to get higher LB scores. I believe that one can use the Kaggle's notebook to train a high performance model. The main challenge in this competition, in my opinion is how to deal with the **data**.",
      "replies": [
        {
          "id": 2099698,
          "postDate": "2023-01-14T16:53:08.490Z",
          "content": "<p>kaggle not providing enough disk space for merging external data.</p>",
          "rawMarkdown": "kaggle not providing enough disk space for merging external data."
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2099319,
      "author_name": "i_love_huyen_tran",
      "author_url": "",
      "post_date": "2023-01-14T10:44:13.767000",
      "content": "<p>IMO, the answer are: </p>\n<ul>\n<li>Yes: you can train a model with high score in LB with Kaggle System (GPU P100, 2xT4), but you need to apply some tricky solutions like: gradient accumulation, freeze/unfreeze model, … to fit your GPUs.</li>\n<li>No: Unfortunately, this competition requires a lot of experiments, you need to test out your ideas which are enormous. Besides, you also need to finetune your params (a lot) to gain the best score, so in the end, you will need a powerful local hardware (harsh truth).</li>\n</ul>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2099281,
      "author_name": "Rasoul Mojtahedzadeh",
      "author_url": "",
      "post_date": "2023-01-14T10:07:34.137000",
      "content": "<p>The very first model that I trained I used the Kaggle notebook (GPU P100) and its LB score is 0.42. However, since then I have used my local GPU (RTX 2080 ti) which has less memory than GPU P100 but I managed to get higher LB scores. I believe that one can use the Kaggle's notebook to train a high performance model. The main challenge in this competition, in my opinion is how to deal with the <strong>data</strong>.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 2099698,
          "author_name": "dragon zhang",
          "author_url": "",
          "post_date": "2023-01-14T16:53:08.490000",
          "content": "<p>kaggle not providing enough disk space for merging external data.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2099160": "All the high scored notebook (above .4) are trained from local hardware. Have anyone successfully reproduce them in kaggle system? The possible issue might be the model size, image size and batch size limitation. ",
    "2099319": "IMO, the answer are: \n- Yes: you can train a model with high score in LB with Kaggle System (GPU P100, 2xT4), but you need to apply some tricky solutions like: gradient accumulation, freeze/unfreeze model, ... to fit your GPUs.\n- No: Unfortunately, this competition requires a lot of experiments, you need to test out your ideas which are enormous. Besides, you also need to finetune your params (a lot) to gain the best score, so in the end, you will need a powerful local hardware (harsh truth).",
    "2099281": "The very first model that I trained I used the Kaggle notebook (GPU P100) and its LB score is 0.42. However, since then I have used my local GPU (RTX 2080 ti) which has less memory than GPU P100 but I managed to get higher LB scores. I believe that one can use the Kaggle's notebook to train a high performance model. The main challenge in this competition, in my opinion is how to deal with the **data**."
  }
}