{
  "id": 97675,
  "title": "You dont need to train in Kernels.",
  "url": "/competitions/aptos2019-blindness-detection/discussion/97675",
  "author_name": "",
  "post_date": "2019-06-28T09:48:12.281845400Z",
  "votes": 29,
  "comment_count": 10,
  "views": 0,
  "content": "<p>In this Synchronous Kernels Only competition, you do not need to train the model in the Kernel. You can do it at home workstation or cloud :D </p>\n\n<p>To demonstrate how to make an inference only kernel, I made a Kernel: <a href=\"https://www.kaggle.com/abhishek/pytorch-inference-kernel\">https://www.kaggle.com/abhishek/pytorch-inference-kernel</a></p>\n\n<p>Hope that helps! </p>",
  "messages": [
    {
      "id": "563468",
      "postDate": "06/28/2019 09:48:12",
      "content": "<p>In this Synchronous Kernels Only competition, you do not need to train the model in the Kernel. You can do it at home workstation or cloud :D </p>\n\n<p>To demonstrate how to make an inference only kernel, I made a Kernel: <a href=\"https://www.kaggle.com/abhishek/pytorch-inference-kernel\">https://www.kaggle.com/abhishek/pytorch-inference-kernel</a></p>\n\n<p>Hope that helps! </p>",
      "rawMarkdown": "In this Synchronous Kernels Only competition, you do not need to train the model in the Kernel. You can do it at home workstation or cloud :D \n\nTo demonstrate how to make an inference only kernel, I made a Kernel: https://www.kaggle.com/abhishek/pytorch-inference-kernel\n\nHope that helps!",
      "votes": null
    },
    {
      "id": "563474",
      "postDate": "06/28/2019 10:00:31",
      "content": "<p>Still very annoying that you have to upload your new models to a new dataset eveytime you make a new submission (and then run it twice) . It should be optional to do inference on kernel and/or commit for private set and see the public result immediately. </p>",
      "rawMarkdown": "Still very annoying that you have to upload your new models to a new dataset eveytime you make a new submission (and then run it twice) . It should be optional to do inference on kernel and/or commit for private set and see the public result immediately.",
      "votes": null
    },
    {
      "id": "563505",
      "postDate": "06/28/2019 10:50:16",
      "content": "<p>Use Local CV :) </p>",
      "rawMarkdown": "Use Local CV :)",
      "votes": null
    },
    {
      "id": "563555",
      "postDate": "06/28/2019 12:19:45",
      "content": "<p>So it's not a kernels only competition</p>",
      "rawMarkdown": "So it's not a kernels only competition",
      "votes": null
    },
    {
      "id": "563569",
      "postDate": "06/28/2019 12:42:41",
      "content": "<p>It is. You can use pretrained models but you can't make submission without kernel.</p>\n\n<p><em>\"In this synchronous Kernels-only competition...\"</em> <a href=\"https://www.kaggle.com/c/aptos2019-blindness-detection/overview\">overview</a></p>\n\n<p><em>\"You can still train a model offline, upload it as a dataset, and use the kernel exclusively to perform inference.\"</em> <a href=\"https://www.kaggle.com/c/aptos2019-blindness-detection/overview/kernels-requirements\">kernels-requirements</a></p>",
      "rawMarkdown": "It is. You can use pretrained models but you can't make submission without kernel.\n\n*\"In this synchronous Kernels-only competition...\"* [overview](https://www.kaggle.com/c/aptos2019-blindness-detection/overview)\n\n*\"You can still train a model offline, upload it as a dataset, and use the kernel exclusively to perform inference.\"* [kernels-requirements](https://www.kaggle.com/c/aptos2019-blindness-detection/overview/kernels-requirements)",
      "votes": null
    },
    {
      "id": "565682",
      "postDate": "07/01/2019 09:23:40",
      "content": "<p>The method you proposed is a good method, but I like to use kaggle kernel to train my model in a kernel and inference in another kernel because of the p100 GPU.</p>",
      "rawMarkdown": "The method you proposed is a good method, but I like to use kaggle kernel to train my model in a kernel and inference in another kernel because of the p100 GPU.",
      "votes": null
    },
    {
      "id": "565751",
      "postDate": "07/01/2019 10:31:12",
      "content": "<p>This often confuses me as well. Why is it a kernels only competition if you can train it with whatever you like outside Kaggle and then just do the prediction by uploading the pre-trained model weights? Well, I guess at least if forces you to show you have a working model that was trained somehow..</p>",
      "rawMarkdown": "This often confuses me as well. Why is it a kernels only competition if you can train it with whatever you like outside Kaggle and then just do the prediction by uploading the pre-trained model weights? Well, I guess at least if forces you to show you have a working model that was trained somehow..",
      "votes": null
    },
    {
      "id": "565908",
      "postDate": "07/01/2019 14:39:52",
      "content": "<p>I am having problens with kernel submission, i have tested two different approaches, 1 - Loaded model weights with .h5 file and  predict the submission inside the kaggle kernel, got the error \"Kernel Out of Resources\"\n2 - upload submission file, but i had another error \"Submission Error\", code below:</p>\n\n<p>\"submission =  pd.read_csv('../input/submission/submission.csv')\nsubmission.to_csv('submission.csv', index=False)\"</p>\n\n<p>Any idea how to solve these problems?</p>",
      "rawMarkdown": "I am having problens with kernel submission, i have tested two different approaches, 1 - Loaded model weights with .h5 file and  predict the submission inside the kaggle kernel, got the error \"Kernel Out of Resources\"\n2 - upload submission file, but i had another error \"Submission Error\", code below:\n\n\"submission =  pd.read_csv('../input/submission/submission.csv')\nsubmission.to_csv('submission.csv', index=False)\"\n\nAny idea how to solve these problems?",
      "votes": null
    },
    {
      "id": "569501",
      "postDate": "07/06/2019 19:12:03",
      "content": "<p><em>Hi! <a href=\"/nandodmelo\">@nandodmelo</a>, try to use small batch size when you run your model training</em></p>",
      "rawMarkdown": "*Hi! @nandodmelo, try to use small batch size when you run your model training*",
      "votes": null
    },
    {
      "id": "570409",
      "postDate": "07/08/2019 08:44:45",
      "content": "<p><a href=\"/beluga\">@beluga</a> I would prefer an additional tag to separate kernel competitions then. </p>\n\n<ol>\n<li>Training and inference need to happen in kernels.</li>\n<li>Only inference need to happen in kernels.</li>\n</ol>",
      "rawMarkdown": "beluga I would prefer an additional tag to separate kernel competitions then. \n\n1. Training and inference need to happen in kernels.\n2. Only inference need to happen in kernels.",
      "votes": null
    },
    {
      "id": "574530",
      "postDate": "07/14/2019 03:26:18",
      "content": "<p>I also tested  the  approach ,upload submission file.  do you solve the problem? can you give me some advices?</p>",
      "rawMarkdown": "I also tested  the  approach ,upload submission file.  do you solve the problem? can you give me some advices?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 563474,
      "author_name": "suicaokhoailang",
      "author_url": "",
      "post_date": "06/28/2019 10:00:31",
      "content": "<p>Still very annoying that you have to upload your new models to a new dataset eveytime you make a new submission (and then run it twice) . It should be optional to do inference on kernel and/or commit for private set and see the public result immediately. </p>",
      "votes": null,
      "replies": [
        {
          "id": 563505,
          "author_name": "gaborfodor",
          "author_url": "",
          "post_date": "06/28/2019 10:50:16",
          "content": "<p>Use Local CV :) </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 563555,
      "author_name": "michaelsnell",
      "author_url": "",
      "post_date": "06/28/2019 12:19:45",
      "content": "<p>So it's not a kernels only competition</p>",
      "votes": null,
      "replies": [
        {
          "id": 563569,
          "author_name": "gaborfodor",
          "author_url": "",
          "post_date": "06/28/2019 12:42:41",
          "content": "<p>It is. You can use pretrained models but you can't make submission without kernel.</p>\n\n<p><em>\"In this synchronous Kernels-only competition...\"</em> <a href=\"https://www.kaggle.com/c/aptos2019-blindness-detection/overview\">overview</a></p>\n\n<p><em>\"You can still train a model offline, upload it as a dataset, and use the kernel exclusively to perform inference.\"</em> <a href=\"https://www.kaggle.com/c/aptos2019-blindness-detection/overview/kernels-requirements\">kernels-requirements</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 565751,
          "author_name": "donkeys",
          "author_url": "",
          "post_date": "07/01/2019 10:31:12",
          "content": "<p>This often confuses me as well. Why is it a kernels only competition if you can train it with whatever you like outside Kaggle and then just do the prediction by uploading the pre-trained model weights? Well, I guess at least if forces you to show you have a working model that was trained somehow..</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 570409,
          "author_name": "michaelsnell",
          "author_url": "",
          "post_date": "07/08/2019 08:44:45",
          "content": "<p><a href=\"/beluga\">@beluga</a> I would prefer an additional tag to separate kernel competitions then. </p>\n\n<ol>\n<li>Training and inference need to happen in kernels.</li>\n<li>Only inference need to happen in kernels.</li>\n</ol>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 565682,
      "author_name": "tenffe",
      "author_url": "",
      "post_date": "07/01/2019 09:23:40",
      "content": "<p>The method you proposed is a good method, but I like to use kaggle kernel to train my model in a kernel and inference in another kernel because of the p100 GPU.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 565908,
      "author_name": "nandodmelo",
      "author_url": "",
      "post_date": "07/01/2019 14:39:52",
      "content": "<p>I am having problens with kernel submission, i have tested two different approaches, 1 - Loaded model weights with .h5 file and  predict the submission inside the kaggle kernel, got the error \"Kernel Out of Resources\"\n2 - upload submission file, but i had another error \"Submission Error\", code below:</p>\n\n<p>\"submission =  pd.read_csv('../input/submission/submission.csv')\nsubmission.to_csv('submission.csv', index=False)\"</p>\n\n<p>Any idea how to solve these problems?</p>",
      "votes": null,
      "replies": [
        {
          "id": 569501,
          "author_name": "cv13j0",
          "author_url": "",
          "post_date": "07/06/2019 19:12:03",
          "content": "<p><em>Hi! <a href=\"/nandodmelo\">@nandodmelo</a>, try to use small batch size when you run your model training</em></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 574530,
          "author_name": "xcz12138",
          "author_url": "",
          "post_date": "07/14/2019 03:26:18",
          "content": "<p>I also tested  the  approach ,upload submission file.  do you solve the problem? can you give me some advices?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "563468": "In this Synchronous Kernels Only competition, you do not need to train the model in the Kernel. You can do it at home workstation or cloud :D \n\nTo demonstrate how to make an inference only kernel, I made a Kernel: https://www.kaggle.com/abhishek/pytorch-inference-kernel\n\nHope that helps!",
    "563474": "Still very annoying that you have to upload your new models to a new dataset eveytime you make a new submission (and then run it twice) . It should be optional to do inference on kernel and/or commit for private set and see the public result immediately.",
    "563505": "Use Local CV :)",
    "563555": "So it's not a kernels only competition",
    "563569": "It is. You can use pretrained models but you can't make submission without kernel.\n\n*\"In this synchronous Kernels-only competition...\"* [overview](https://www.kaggle.com/c/aptos2019-blindness-detection/overview)\n\n*\"You can still train a model offline, upload it as a dataset, and use the kernel exclusively to perform inference.\"* [kernels-requirements](https://www.kaggle.com/c/aptos2019-blindness-detection/overview/kernels-requirements)",
    "565682": "The method you proposed is a good method, but I like to use kaggle kernel to train my model in a kernel and inference in another kernel because of the p100 GPU.",
    "565751": "This often confuses me as well. Why is it a kernels only competition if you can train it with whatever you like outside Kaggle and then just do the prediction by uploading the pre-trained model weights? Well, I guess at least if forces you to show you have a working model that was trained somehow..",
    "565908": "I am having problens with kernel submission, i have tested two different approaches, 1 - Loaded model weights with .h5 file and  predict the submission inside the kaggle kernel, got the error \"Kernel Out of Resources\"\n2 - upload submission file, but i had another error \"Submission Error\", code below:\n\n\"submission =  pd.read_csv('../input/submission/submission.csv')\nsubmission.to_csv('submission.csv', index=False)\"\n\nAny idea how to solve these problems?",
    "569501": "*Hi! @nandodmelo, try to use small batch size when you run your model training*",
    "570409": "beluga I would prefer an additional tag to separate kernel competitions then. \n\n1. Training and inference need to happen in kernels.\n2. Only inference need to happen in kernels.",
    "574530": "I also tested  the  approach ,upload submission file.  do you solve the problem? can you give me some advices?"
  },
  "source": "meta"
}