{
  "id": 198989,
  "title": "Internet should be disabled",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/198989",
  "author_name": "lukachkhetiani",
  "post_date": "2020-11-23T23:29:19.486000",
  "votes": 2,
  "comment_count": 11,
  "views": 0,
  "content": "<p>Love the organizers' idea: Okay, we know that even google/ibm/fair researchers use pretrained models for almost any regular task, but let's ask competitors to disable the internet for fun:)</p>",
  "messages": [
    {
      "id": 1088826,
      "postDate": "2020-11-24T01:00:11.280Z",
      "content": "<p>As explained elsewhere, the committed notebooks have to turn off Internet access to prevent mining the hidden test dataset.</p>\n<p>That is why you need to load packages as datasets.</p>\n<p>It is fully expected that kagglers will use prebuilt and pretrained models.</p>",
      "rawMarkdown": "As explained elsewhere, the committed notebooks have to turn off Internet access to prevent mining the hidden test dataset.\n\nThat is why you need to load packages as datasets.\n\nIt is fully expected that kagglers will use prebuilt and pretrained models.",
      "votes": 2,
      "replies": [
        {
          "id": 1089071,
          "postDate": "2020-11-24T07:27:22.447Z",
          "content": "<p>I didn't think of mining the test set. Still what are the possible reasons for not using TPU when submitting? Is it mine-able too?</p>",
          "rawMarkdown": "I didn't think of mining the test set. Still what are the possible reasons for not using TPU when submitting? Is it mine-able too?",
          "replies": [
            {
              "id": 1089273,
              "postDate": "2020-11-24T11:07:31.387Z",
              "content": "<p>I believe TPU requires Internet to be on.</p>",
              "rawMarkdown": "I believe TPU requires Internet to be on.",
              "votes": 1
            },
            {
              "id": 1089281,
              "postDate": "2020-11-24T11:12:54.003Z",
              "content": "<p>Oh right, they read data from GCS.</p>",
              "rawMarkdown": "Oh right, they read data from GCS."
            }
          ]
        }
      ]
    },
    {
      "id": 1088764,
      "postDate": "2020-11-23T23:29:19.487Z",
      "content": "<p>Love the organizers' idea: Okay, we know that even google/ibm/fair researchers use pretrained models for almost any regular task, but let's ask competitors to disable the internet for fun:)</p>",
      "rawMarkdown": "Love the organizers' idea: Okay, we know that even google/ibm/fair researchers use pretrained models for almost any regular task, but let's ask competitors to disable the internet for fun:)\n\n",
      "votes": 2
    },
    {
      "id": 1090478,
      "postDate": "2020-11-25T11:12:16.813Z",
      "content": "<p>Hey guys. My initial submission needed some packages that required torch 1.7; <br>\nAlso, the model architectures I've used needed complex file structure that you can't rewrite in a day.  </p>\n<p>I was able to build new versions of frameworks &amp; upload network files. In case you're having a hard time, I can write a simple notebook.</p>",
      "rawMarkdown": "Hey guys. My initial submission needed some packages that required torch 1.7; \nAlso, the model architectures I've used needed complex file structure that you can't rewrite in a day.  \n\nI was able to build new versions of frameworks & upload network files. In case you're having a hard time, I can write a simple notebook."
    },
    {
      "id": 1090161,
      "postDate": "2020-11-25T06:04:06.563Z",
      "content": "<p>how to upload the pretrained models or models trained by local GPU?</p>",
      "rawMarkdown": "how to upload the pretrained models or models trained by local GPU?",
      "replies": [
        {
          "id": 1090355,
          "postDate": "2020-11-25T09:27:28.793Z",
          "rawMarkdown": "",
          "votes": 1,
          "isDeleted": true
        }
      ]
    },
    {
      "id": 1088766,
      "postDate": "2020-11-23T23:30:25.527Z",
      "content": "<p>P.S. uploading packages and whole repositories as datasets is humiliating.  </p>",
      "rawMarkdown": "P.S. uploading packages and whole repositories as datasets is humiliating.  ",
      "replies": [
        {
          "id": 1089072,
          "postDate": "2020-11-24T07:28:07.383Z",
          "content": "<p>yeah it's kinda annoying. For every code competitions I think I upload about 3-4 datasets for libraries.</p>",
          "rawMarkdown": "yeah it's kinda annoying. For every code competitions I think I upload about 3-4 datasets for libraries."
        }
      ]
    },
    {
      "id": 1089872,
      "postDate": "2020-11-24T21:15:00.443Z",
      "rawMarkdown": "",
      "isDeleted": true,
      "replies": [
        {
          "id": 1089901,
          "postDate": "2020-11-24T21:51:43.487Z",
          "content": "<p>You are correct</p>\n<ol>\n<li>Train model using Internet and TPU if you wish</li>\n<li>download model/weights</li>\n<li>create dataset and load model/weights into it</li>\n<li>create inference notebook and add dataset with \"add data\"</li>\n<li>run inference notebook and commit. Turn off Internet. You can use GPU.</li>\n</ol>",
          "rawMarkdown": "You are correct\n\n1. Train model using Internet and TPU if you wish\n2. download model/weights\n3. create dataset and load model/weights into it\n4. create inference notebook and add dataset with \"add data\"\n5. run inference notebook and commit. Turn off Internet. You can use GPU.",
          "votes": 1
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1088826,
      "author_name": "quadcore/Richard Epstein",
      "author_url": "",
      "post_date": "2020-11-24T01:00:11.280000",
      "content": "<p>As explained elsewhere, the committed notebooks have to turn off Internet access to prevent mining the hidden test dataset.</p>\n<p>That is why you need to load packages as datasets.</p>\n<p>It is fully expected that kagglers will use prebuilt and pretrained models.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 1089071,
          "author_name": "Long Luu",
          "author_url": "",
          "post_date": "2020-11-24T07:27:22.447000",
          "content": "<p>I didn't think of mining the test set. Still what are the possible reasons for not using TPU when submitting? Is it mine-able too?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 1089273,
              "author_name": "quadcore/Richard Epstein",
              "author_url": "",
              "post_date": "2020-11-24T11:07:31.387000",
              "content": "<p>I believe TPU requires Internet to be on.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 1089281,
              "author_name": "Long Luu",
              "author_url": "",
              "post_date": "2020-11-24T11:12:54.003000",
              "content": "<p>Oh right, they read data from GCS.</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 1090478,
      "author_name": "lukachkhetiani",
      "author_url": "",
      "post_date": "2020-11-25T11:12:16.813000",
      "content": "<p>Hey guys. My initial submission needed some packages that required torch 1.7; <br>\nAlso, the model architectures I've used needed complex file structure that you can't rewrite in a day.  </p>\n<p>I was able to build new versions of frameworks &amp; upload network files. In case you're having a hard time, I can write a simple notebook.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1090161,
      "author_name": "Yi Wu",
      "author_url": "",
      "post_date": "2020-11-25T06:04:06.563000",
      "content": "<p>how to upload the pretrained models or models trained by local GPU?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1090355,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-11-25T09:27:28.793000",
          "content": "",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1088766,
      "author_name": "lukachkhetiani",
      "author_url": "",
      "post_date": "2020-11-23T23:30:25.527000",
      "content": "<p>P.S. uploading packages and whole repositories as datasets is humiliating.  </p>",
      "votes": 0,
      "replies": [
        {
          "id": 1089072,
          "author_name": "Long Luu",
          "author_url": "",
          "post_date": "2020-11-24T07:28:07.383000",
          "content": "<p>yeah it's kinda annoying. For every code competitions I think I upload about 3-4 datasets for libraries.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1089872,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-11-24T21:15:00.443000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 1089901,
          "author_name": "quadcore/Richard Epstein",
          "author_url": "",
          "post_date": "2020-11-24T21:51:43.487000",
          "content": "<p>You are correct</p>\n<ol>\n<li>Train model using Internet and TPU if you wish</li>\n<li>download model/weights</li>\n<li>create dataset and load model/weights into it</li>\n<li>create inference notebook and add dataset with \"add data\"</li>\n<li>run inference notebook and commit. Turn off Internet. You can use GPU.</li>\n</ol>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1088826": "As explained elsewhere, the committed notebooks have to turn off Internet access to prevent mining the hidden test dataset.\n\nThat is why you need to load packages as datasets.\n\nIt is fully expected that kagglers will use prebuilt and pretrained models.",
    "1088764": "Love the organizers' idea: Okay, we know that even google/ibm/fair researchers use pretrained models for almost any regular task, but let's ask competitors to disable the internet for fun:)\n\n",
    "1090478": "Hey guys. My initial submission needed some packages that required torch 1.7; \nAlso, the model architectures I've used needed complex file structure that you can't rewrite in a day.  \n\nI was able to build new versions of frameworks & upload network files. In case you're having a hard time, I can write a simple notebook.",
    "1090161": "how to upload the pretrained models or models trained by local GPU?",
    "1088766": "P.S. uploading packages and whole repositories as datasets is humiliating.  ",
    "1089872": ""
  }
}