{
  "id": 161905,
  "title": "Suggestions to the Kaggle team. ",
  "url": "/competitions/alaska2-image-steganalysis/discussion/161905",
  "author_name": "",
  "post_date": "2020-06-26T15:32:21.898077600Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi Kaggle team,</p>\n\n<p>I would like to thank the Kaggle team in putting wonderful resources for all of us. But, there is something I would like to point to here. </p>\n\n<p>I am not sure if everyone else in this completion is facing the same issues or is it me alone. I can not train models on full training data using Kaggle notebook, no matter what I do. I am just taking subset of whole data but still, the training time is too large. The TPU notebooks have time limit of 3 hours per session. But, there is no code requirements for the competition. So, the question is why not let the user use 30 hours of TPU time in one go? Would not that be good for competitors to have? I am not sure if the GPU notebooks have time limits as well.  </p>\n\n<p>One more thing is that even if it is code requirement completion but still user should be allowed to use all TPU/GPU time in one go. If the notebook run time exceeds allowed time in competition then just do not allow the user to submit the <code>submission.csv</code> file from there. This is simply possible to do I think and it will be more efficient for competitors to train models here on the Kaggle. </p>\n\n<p>Thanks for taking the note. </p>",
  "messages": [
    {
      "id": "903125",
      "postDate": "06/26/2020 15:32:21",
      "content": "<p>Hi Kaggle team,</p>\n\n<p>I would like to thank the Kaggle team in putting wonderful resources for all of us. But, there is something I would like to point to here. </p>\n\n<p>I am not sure if everyone else in this completion is facing the same issues or is it me alone. I can not train models on full training data using Kaggle notebook, no matter what I do. I am just taking subset of whole data but still, the training time is too large. The TPU notebooks have time limit of 3 hours per session. But, there is no code requirements for the competition. So, the question is why not let the user use 30 hours of TPU time in one go? Would not that be good for competitors to have? I am not sure if the GPU notebooks have time limits as well.  </p>\n\n<p>One more thing is that even if it is code requirement completion but still user should be allowed to use all TPU/GPU time in one go. If the notebook run time exceeds allowed time in competition then just do not allow the user to submit the <code>submission.csv</code> file from there. This is simply possible to do I think and it will be more efficient for competitors to train models here on the Kaggle. </p>\n\n<p>Thanks for taking the note. </p>",
      "rawMarkdown": "Hi Kaggle team,\n\nI would like to thank the Kaggle team in putting wonderful resources for all of us. But, there is something I would like to point to here. \n\nI am not sure if everyone else in this completion is facing the same issues or is it me alone. I can not train models on full training data using Kaggle notebook, no matter what I do. I am just taking subset of whole data but still, the training time is too large. The TPU notebooks have time limit of 3 hours per session. But, there is no code requirements for the competition. So, the question is why not let the user use 30 hours of TPU time in one go? Would not that be good for competitors to have? I am not sure if the GPU notebooks have time limits as well.  \n\nOne more thing is that even if it is code requirement completion but still user should be allowed to use all TPU/GPU time in one go. If the notebook run time exceeds allowed time in competition then just do not allow the user to submit the `submission.csv` file from there. This is simply possible to do I think and it will be more efficient for competitors to train models here on the Kaggle. \n\nThanks for taking the note.",
      "votes": null
    },
    {
      "id": "903196",
      "postDate": "06/26/2020 16:25:10",
      "content": "<p>Ah yes, I feel you there.\nThey must have some policy decisions for limiting the one-shot-usage. Suppose you made a mistake while committing?</p>\n\n<p>But there are mechanisms that allow you to continue your training where you left off. That way you can split your training sessions across multiple notebook commits. That's how I am working in this competition since I don't have any serious hardware with me.\n<a href=\"/shonenkov\">@shonenkov</a> has shared an <a href=\"https://www.kaggle.com/shonenkov/train-inference-gpu-baseline\">awesome</a> notebook which describes this mechanism. Look for the <code>save</code> and <code>load</code> functions in the Fitter class.</p>\n\n<p>Hope this helps.</p>",
      "rawMarkdown": "Ah yes, I feel you there.\nThey must have some policy decisions for limiting the one-shot-usage. Suppose you made a mistake while committing?\n\nBut there are mechanisms that allow you to continue your training where you left off. That way you can split your training sessions across multiple notebook commits. That's how I am working in this competition since I don't have any serious hardware with me.\n@shonenkov has shared an [awesome](https://www.kaggle.com/shonenkov/train-inference-gpu-baseline) notebook which describes this mechanism. Look for the `save` and `load` functions in the Fitter class.\n\nHope this helps.",
      "votes": null
    },
    {
      "id": "903201",
      "postDate": "06/26/2020 16:28:21",
      "content": "<p>yeah, that can be done for sure. Like saving the model state and loading them in different notebook and train again. I think we do not have any option than this. </p>",
      "rawMarkdown": "yeah, that can be done for sure. Like saving the model state and loading them in different notebook and train again. I think we do not have any option than this.",
      "votes": null
    },
    {
      "id": "903217",
      "postDate": "06/26/2020 16:40:26",
      "content": "<p>I did not know that on GPU we have 9 hours per session. Really that can work I think and can make a difference. Thanks for pointing out to the kernel as well. </p>",
      "rawMarkdown": "I did not know that on GPU we have 9 hours per session. Really that can work I think and can make a difference. Thanks for pointing out to the kernel as well.",
      "votes": null
    },
    {
      "id": "903255",
      "postDate": "06/26/2020 17:20:10",
      "content": "<p>Glad to be of help :)</p>",
      "rawMarkdown": "Glad to be of help :)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 903196,
      "author_name": "mightyrains",
      "author_url": "",
      "post_date": "06/26/2020 16:25:10",
      "content": "<p>Ah yes, I feel you there.\nThey must have some policy decisions for limiting the one-shot-usage. Suppose you made a mistake while committing?</p>\n\n<p>But there are mechanisms that allow you to continue your training where you left off. That way you can split your training sessions across multiple notebook commits. That's how I am working in this competition since I don't have any serious hardware with me.\n<a href=\"/shonenkov\">@shonenkov</a> has shared an <a href=\"https://www.kaggle.com/shonenkov/train-inference-gpu-baseline\">awesome</a> notebook which describes this mechanism. Look for the <code>save</code> and <code>load</code> functions in the Fitter class.</p>\n\n<p>Hope this helps.</p>",
      "votes": null,
      "replies": [
        {
          "id": 903201,
          "author_name": "urvishp80",
          "author_url": "",
          "post_date": "06/26/2020 16:28:21",
          "content": "<p>yeah, that can be done for sure. Like saving the model state and loading them in different notebook and train again. I think we do not have any option than this. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 903217,
          "author_name": "urvishp80",
          "author_url": "",
          "post_date": "06/26/2020 16:40:26",
          "content": "<p>I did not know that on GPU we have 9 hours per session. Really that can work I think and can make a difference. Thanks for pointing out to the kernel as well. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 903255,
          "author_name": "mightyrains",
          "author_url": "",
          "post_date": "06/26/2020 17:20:10",
          "content": "<p>Glad to be of help :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "903125": "Hi Kaggle team,\n\nI would like to thank the Kaggle team in putting wonderful resources for all of us. But, there is something I would like to point to here. \n\nI am not sure if everyone else in this completion is facing the same issues or is it me alone. I can not train models on full training data using Kaggle notebook, no matter what I do. I am just taking subset of whole data but still, the training time is too large. The TPU notebooks have time limit of 3 hours per session. But, there is no code requirements for the competition. So, the question is why not let the user use 30 hours of TPU time in one go? Would not that be good for competitors to have? I am not sure if the GPU notebooks have time limits as well.  \n\nOne more thing is that even if it is code requirement completion but still user should be allowed to use all TPU/GPU time in one go. If the notebook run time exceeds allowed time in competition then just do not allow the user to submit the `submission.csv` file from there. This is simply possible to do I think and it will be more efficient for competitors to train models here on the Kaggle. \n\nThanks for taking the note.",
    "903196": "Ah yes, I feel you there.\nThey must have some policy decisions for limiting the one-shot-usage. Suppose you made a mistake while committing?\n\nBut there are mechanisms that allow you to continue your training where you left off. That way you can split your training sessions across multiple notebook commits. That's how I am working in this competition since I don't have any serious hardware with me.\n@shonenkov has shared an [awesome](https://www.kaggle.com/shonenkov/train-inference-gpu-baseline) notebook which describes this mechanism. Look for the `save` and `load` functions in the Fitter class.\n\nHope this helps.",
    "903201": "yeah, that can be done for sure. Like saving the model state and loading them in different notebook and train again. I think we do not have any option than this.",
    "903217": "I did not know that on GPU we have 9 hours per session. Really that can work I think and can make a difference. Thanks for pointing out to the kernel as well.",
    "903255": "Glad to be of help :)"
  },
  "source": "meta"
}