{
  "id": 209408,
  "title": "How can I get more kaggle GPU hours? 40hours have used off",
  "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/209408",
  "author_name": "",
  "post_date": "2021-01-07T13:56:13.360817400Z",
  "votes": 2,
  "comment_count": 6,
  "views": 0,
  "content": "<p>so I can not inference with GPU，then I can not submit csv</p>\n<p>How can I get more kaggle GPU hours for inferencing?</p>",
  "messages": [
    {
      "id": "1142616",
      "postDate": "01/07/2021 13:56:13",
      "content": "<p>so I can not inference with GPU，then I can not submit csv</p>\n<p>How can I get more kaggle GPU hours for inferencing?</p>",
      "rawMarkdown": "so I can not inference with GPU，then I can not submit csv\n\nHow can I get more kaggle GPU hours for inferencing?",
      "votes": null
    },
    {
      "id": "1142787",
      "postDate": "01/07/2021 15:34:40",
      "content": "<p>I have heard some mention getting a type of google account. Kaggle might also have a program similar to its TPU star program (<a href=\"https://www.kaggle.com/tpu-prize\" target=\"_blank\">https://www.kaggle.com/tpu-prize</a> ), but I'm unsure. Hope this helps.</p>",
      "rawMarkdown": "I have heard some mention getting a type of google account. Kaggle might also have a program similar to its TPU star program (https://www.kaggle.com/tpu-prize ), but I'm unsure. Hope this helps.",
      "votes": null
    },
    {
      "id": "1143440",
      "postDate": "01/07/2021 22:47:39",
      "content": "<p>If you are using PyTorch, then switching to PyTorch Lightning might be a (relatively…) easy option for also being able to use the TPU hours without changing your code. Another idea to avoid using GPU hours (assuming you can use compute somewhere else to train your models) is to write your script so that when you save it (you can basically test whether it's a real submission run or just the one saving the notebook via the length of the test set like <a href=\"https://www.kaggle.com/underwearfitting/make-final-submission-the-efficient-way\" target=\"_blank\">in this notebook</a>), you do not actually run the time consuming part.</p>",
      "rawMarkdown": "If you are using PyTorch, then switching to PyTorch Lightning might be a (relatively...) easy option for also being able to use the TPU hours without changing your code. Another idea to avoid using GPU hours (assuming you can use compute somewhere else to train your models) is to write your script so that when you save it (you can basically test whether it's a real submission run or just the one saving the notebook via the length of the test set like [in this notebook](https://www.kaggle.com/underwearfitting/make-final-submission-the-efficient-way)), you do not actually run the time consuming part.",
      "votes": null
    },
    {
      "id": "1143800",
      "postDate": "01/08/2021 04:28:38",
      "content": "<p>Thank you！</p>",
      "rawMarkdown": "Thank you！",
      "votes": null
    },
    {
      "id": "1145056",
      "postDate": "01/08/2021 21:02:39",
      "content": "<p>I have just made a post about using Google Colab for extra TPU/GPU. Giving you an upwards of 20hrs+ of accelerator quota. <a href=\"https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/209846\" target=\"_blank\">https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/209846</a> Meaning that you can do the training on Colab then inference on Kaggle.</p>",
      "rawMarkdown": "I have just made a post about using Google Colab for extra TPU/GPU. Giving you an upwards of 20hrs+ of accelerator quota. https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/209846 Meaning that you can do the training on Colab then inference on Kaggle.",
      "votes": null
    },
    {
      "id": "1145062",
      "postDate": "01/08/2021 21:07:19",
      "content": "<p>Link fixed</p>",
      "rawMarkdown": "Link fixed",
      "votes": null
    },
    {
      "id": "1146208",
      "postDate": "01/09/2021 15:50:41",
      "content": "<p>You can essentially use zero GPU time while saving using the trick in this template: <a href=\"https://www.kaggle.com/marcelbischoff/ranzcr-quick-submission-template\" target=\"_blank\">https://www.kaggle.com/marcelbischoff/ranzcr-quick-submission-template</a>. Make sure to only active GPU for saving and deactivate it immediately - even while it is still saving, which now will only take a minute or so. This way it only takes around a minute of GPU time. Running the submission against the private set apparently doesn't count towards the quota.</p>",
      "rawMarkdown": "You can essentially use zero GPU time while saving using the trick in this template: https://www.kaggle.com/marcelbischoff/ranzcr-quick-submission-template. Make sure to only active GPU for saving and deactivate it immediately - even while it is still saving, which now will only take a minute or so. This way it only takes around a minute of GPU time. Running the submission against the private set apparently doesn't count towards the quota.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1142787,
      "author_name": "kyleberdy",
      "author_url": "",
      "post_date": "01/07/2021 15:34:40",
      "content": "<p>I have heard some mention getting a type of google account. Kaggle might also have a program similar to its TPU star program (<a href=\"https://www.kaggle.com/tpu-prize\" target=\"_blank\">https://www.kaggle.com/tpu-prize</a> ), but I'm unsure. Hope this helps.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1145062,
          "author_name": "kyleberdy",
          "author_url": "",
          "post_date": "01/08/2021 21:07:19",
          "content": "<p>Link fixed</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1143440,
      "author_name": "bjoernholzhauer",
      "author_url": "",
      "post_date": "01/07/2021 22:47:39",
      "content": "<p>If you are using PyTorch, then switching to PyTorch Lightning might be a (relatively…) easy option for also being able to use the TPU hours without changing your code. Another idea to avoid using GPU hours (assuming you can use compute somewhere else to train your models) is to write your script so that when you save it (you can basically test whether it's a real submission run or just the one saving the notebook via the length of the test set like <a href=\"https://www.kaggle.com/underwearfitting/make-final-submission-the-efficient-way\" target=\"_blank\">in this notebook</a>), you do not actually run the time consuming part.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1143800,
          "author_name": "guochangzhang",
          "author_url": "",
          "post_date": "01/08/2021 04:28:38",
          "content": "<p>Thank you！</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1145056,
      "author_name": "andy1010",
      "author_url": "",
      "post_date": "01/08/2021 21:02:39",
      "content": "<p>I have just made a post about using Google Colab for extra TPU/GPU. Giving you an upwards of 20hrs+ of accelerator quota. <a href=\"https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/209846\" target=\"_blank\">https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/209846</a> Meaning that you can do the training on Colab then inference on Kaggle.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1146208,
      "author_name": "marcelbischoff",
      "author_url": "",
      "post_date": "01/09/2021 15:50:41",
      "content": "<p>You can essentially use zero GPU time while saving using the trick in this template: <a href=\"https://www.kaggle.com/marcelbischoff/ranzcr-quick-submission-template\" target=\"_blank\">https://www.kaggle.com/marcelbischoff/ranzcr-quick-submission-template</a>. Make sure to only active GPU for saving and deactivate it immediately - even while it is still saving, which now will only take a minute or so. This way it only takes around a minute of GPU time. Running the submission against the private set apparently doesn't count towards the quota.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1142616": "so I can not inference with GPU，then I can not submit csv\n\nHow can I get more kaggle GPU hours for inferencing?",
    "1142787": "I have heard some mention getting a type of google account. Kaggle might also have a program similar to its TPU star program (https://www.kaggle.com/tpu-prize ), but I'm unsure. Hope this helps.",
    "1143440": "If you are using PyTorch, then switching to PyTorch Lightning might be a (relatively...) easy option for also being able to use the TPU hours without changing your code. Another idea to avoid using GPU hours (assuming you can use compute somewhere else to train your models) is to write your script so that when you save it (you can basically test whether it's a real submission run or just the one saving the notebook via the length of the test set like [in this notebook](https://www.kaggle.com/underwearfitting/make-final-submission-the-efficient-way)), you do not actually run the time consuming part.",
    "1143800": "Thank you！",
    "1145056": "I have just made a post about using Google Colab for extra TPU/GPU. Giving you an upwards of 20hrs+ of accelerator quota. https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/209846 Meaning that you can do the training on Colab then inference on Kaggle.",
    "1145062": "Link fixed",
    "1146208": "You can essentially use zero GPU time while saving using the trick in this template: https://www.kaggle.com/marcelbischoff/ranzcr-quick-submission-template. Make sure to only active GPU for saving and deactivate it immediately - even while it is still saving, which now will only take a minute or so. This way it only takes around a minute of GPU time. Running the submission against the private set apparently doesn't count towards the quota."
  },
  "source": "meta"
}