{
  "id": 40118,
  "title": "Lessons  learned",
  "url": "/competitions/carvana-image-masking-challenge/discussion/40118",
  "author_name": "",
  "post_date": "2017-09-27T22:56:21.869151500Z",
  "votes": 32,
  "comment_count": 9,
  "views": 0,
  "content": "<p>lessons learned:</p>\n\n<ol>\n<li><p>should have used crops in high resolution instead of increasing  increasing resized input</p></li>\n<li><p>try  different network structure for ensemble, rather than finding the best with the same structure</p></li>\n<li><p>normal practice still work: cross validation, balancing of data, augmentation. These fundamentals are important</p></li>\n</ol>\n\n<p>in short,  \"be brave to try new things\" and \"don't be lazy to  skip n-fold cross validation + ensemble\"</p>",
  "messages": [
    {
      "id": "224913",
      "postDate": "09/27/2017 22:56:21",
      "content": "<p>lessons learned:</p>\n\n<ol>\n<li><p>should have used crops in high resolution instead of increasing  increasing resized input</p></li>\n<li><p>try  different network structure for ensemble, rather than finding the best with the same structure</p></li>\n<li><p>normal practice still work: cross validation, balancing of data, augmentation. These fundamentals are important</p></li>\n</ol>\n\n<p>in short,  \"be brave to try new things\" and \"don't be lazy to  skip n-fold cross validation + ensemble\"</p>",
      "rawMarkdown": "lessons learned:\n\n1. should have used crops in high resolution instead of increasing  increasing resized input\n\n2. try  different network structure for ensemble, rather than finding the best with the same structure\n\n3. normal practice still work: cross validation, balancing of data, augmentation. These fundamentals are important\n\n\nin short,  \"be brave to try new things\" and \"don't be lazy to  skip n-fold cross validation + ensemble\"",
      "votes": null
    },
    {
      "id": "224926",
      "postDate": "09/27/2017 23:28:27",
      "content": "<p>and 4. More GPUs!</p>\n\n<p>Thank you for your input, this competition wouldn't have been that fun without your ideas.</p>",
      "rawMarkdown": "and 4. More GPUs!\n\nThank you for your input, this competition wouldn't have been that fun without your ideas.",
      "votes": null
    },
    {
      "id": "225149",
      "postDate": "09/28/2017 11:27:35",
      "content": "<p>I would love to know some statistics on:\n - What augmentations worked/what didn't\n - Loss functions\n - Did anybody use a smarter train/val split than on a car basis (e.g. by size, color, etc.)</p>",
      "rawMarkdown": "I would love to know some statistics on:\n - What augmentations worked/what didn't\n - Loss functions\n - Did anybody use a smarter train/val split than on a car basis (e.g. by size, color, etc.)",
      "votes": null
    },
    {
      "id": "225153",
      "postDate": "09/28/2017 11:32:39",
      "content": "<p>In my case I decided to use a split using the maker of the car. <br>\nI thought that cars of the same maker have similar shapes.  Also the maker distribution was very different between train and test, but only a little fraction of the test set was used, so maybe this was not true.</p>\n\n<p>I think it was a bad decision and that splitting by car was better.</p>",
      "rawMarkdown": "In my case I decided to use a split using the maker of the car.   \nI thought that cars of the same maker have similar shapes.  Also the maker distribution was very different between train and test, but only a little fraction of the test set was used, so maybe this was not true.\n\nI think it was a bad decision and that splitting by car was better.",
      "votes": null
    },
    {
      "id": "225166",
      "postDate": "09/28/2017 12:07:21",
      "content": "<p>Thanks for all your valuable sharing during! It was my first competition with heavy deeplearning involved. You helped me a lot!</p>",
      "rawMarkdown": "Thanks for all your valuable sharing during! It was my first competition with heavy deeplearning involved. You helped me a lot!",
      "votes": null
    },
    {
      "id": "225172",
      "postDate": "09/28/2017 12:18:35",
      "content": "<p>Thank you, the frog boy!</p>",
      "rawMarkdown": "Thank you, the frog boy!",
      "votes": null
    },
    {
      "id": "225203",
      "postDate": "09/28/2017 13:39:50",
      "content": "<p>You should also add - \"should have used pre-trained encoders\" :-)</p>",
      "rawMarkdown": "You should also add - \"should have used pre-trained encoders\" :-)",
      "votes": null
    },
    {
      "id": "225471",
      "postDate": "09/29/2017 06:33:02",
      "content": "<p>Thank you for sharing!</p>",
      "rawMarkdown": "Thank you for sharing!",
      "votes": null
    },
    {
      "id": "225514",
      "postDate": "09/29/2017 10:28:43",
      "content": "<p>Thank you for your generous contributions to this challenge, Heng! You rock!</p>",
      "rawMarkdown": "Thank you for your generous contributions to this challenge, Heng! You rock!",
      "votes": null
    },
    {
      "id": "225813",
      "postDate": "09/30/2017 02:40:09",
      "content": "<p>Thanks for all your sharing , it helps me a lot!</p>",
      "rawMarkdown": "Thanks for all your sharing , it helps me a lot!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 224926,
      "author_name": "killthekitten",
      "author_url": "",
      "post_date": "09/27/2017 23:28:27",
      "content": "<p>and 4. More GPUs!</p>\n\n<p>Thank you for your input, this competition wouldn't have been that fun without your ideas.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 225149,
      "author_name": "antorsae",
      "author_url": "",
      "post_date": "09/28/2017 11:27:35",
      "content": "<p>I would love to know some statistics on:\n - What augmentations worked/what didn't\n - Loss functions\n - Did anybody use a smarter train/val split than on a car basis (e.g. by size, color, etc.)</p>",
      "votes": null,
      "replies": [
        {
          "id": 225153,
          "author_name": "ironbar",
          "author_url": "",
          "post_date": "09/28/2017 11:32:39",
          "content": "<p>In my case I decided to use a split using the maker of the car. <br>\nI thought that cars of the same maker have similar shapes.  Also the maker distribution was very different between train and test, but only a little fraction of the test set was used, so maybe this was not true.</p>\n\n<p>I think it was a bad decision and that splitting by car was better.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 225166,
      "author_name": "gaborfodor",
      "author_url": "",
      "post_date": "09/28/2017 12:07:21",
      "content": "<p>Thanks for all your valuable sharing during! It was my first competition with heavy deeplearning involved. You helped me a lot!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 225172,
      "author_name": "fengari",
      "author_url": "",
      "post_date": "09/28/2017 12:18:35",
      "content": "<p>Thank you, the frog boy!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 225203,
      "author_name": "alekseit",
      "author_url": "",
      "post_date": "09/28/2017 13:39:50",
      "content": "<p>You should also add - \"should have used pre-trained encoders\" :-)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 225471,
      "author_name": "annapelageya",
      "author_url": "",
      "post_date": "09/29/2017 06:33:02",
      "content": "<p>Thank you for sharing!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 225514,
      "author_name": "slavivanov",
      "author_url": "",
      "post_date": "09/29/2017 10:28:43",
      "content": "<p>Thank you for your generous contributions to this challenge, Heng! You rock!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 225813,
      "author_name": "jingquntang",
      "author_url": "",
      "post_date": "09/30/2017 02:40:09",
      "content": "<p>Thanks for all your sharing , it helps me a lot!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "224913": "lessons learned:\n\n1. should have used crops in high resolution instead of increasing  increasing resized input\n\n2. try  different network structure for ensemble, rather than finding the best with the same structure\n\n3. normal practice still work: cross validation, balancing of data, augmentation. These fundamentals are important\n\n\nin short,  \"be brave to try new things\" and \"don't be lazy to  skip n-fold cross validation + ensemble\"",
    "224926": "and 4. More GPUs!\n\nThank you for your input, this competition wouldn't have been that fun without your ideas.",
    "225149": "I would love to know some statistics on:\n - What augmentations worked/what didn't\n - Loss functions\n - Did anybody use a smarter train/val split than on a car basis (e.g. by size, color, etc.)",
    "225153": "In my case I decided to use a split using the maker of the car.   \nI thought that cars of the same maker have similar shapes.  Also the maker distribution was very different between train and test, but only a little fraction of the test set was used, so maybe this was not true.\n\nI think it was a bad decision and that splitting by car was better.",
    "225166": "Thanks for all your valuable sharing during! It was my first competition with heavy deeplearning involved. You helped me a lot!",
    "225172": "Thank you, the frog boy!",
    "225203": "You should also add - \"should have used pre-trained encoders\" :-)",
    "225471": "Thank you for sharing!",
    "225514": "Thank you for your generous contributions to this challenge, Heng! You rock!",
    "225813": "Thanks for all your sharing , it helps me a lot!"
  },
  "source": "meta"
}