{
  "id": 40241,
  "title": "Comparing ranking and framework",
  "url": "/competitions/carvana-image-masking-challenge/discussion/40241",
  "author_name": "Vladimir Iglovikov",
  "post_date": "2017-09-29T22:59:44.074000",
  "votes": 5,
  "comment_count": 14,
  "views": 0,
  "content": "<p>The framework is much less important than GPUs in this particular task, but still. I really liked what people shared in <a href=\"https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40121\">https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40121</a></p>\n\n<p>It would be nice to know what software participants used and if there is any correlation between framework and result.</p>\n\n<p>To start: </p>\n\n<p>1st place. PyTorch</p>",
  "messages": [
    {
      "id": 226316,
      "postDate": "2017-10-01T20:22:42.593Z",
      "content": "<p>4th place using a TF-backed Keras. </p>\n\n<p>Once I've bought some extra GPUs, I've tried to switch the whole pipeline to Pytorch. It took me a night or two to get it running, but I've faced some memory-related issues (most likely related to how I've used DataLoader) and turned back to Keras. </p>\n\n<p>Still, Pytorch feels much more transparent and intuitive, as well it had some influence on our later pipeline. I'm definitely will prefer it over Keras in the future.</p>",
      "rawMarkdown": "4th place using a TF-backed Keras. \n\nOnce I've bought some extra GPUs, I've tried to switch the whole pipeline to Pytorch. It took me a night or two to get it running, but I've faced some memory-related issues (most likely related to how I've used DataLoader) and turned back to Keras. \n\nStill, Pytorch feels much more transparent and intuitive, as well it had some influence on our later pipeline. I'm definitely will prefer it over Keras in the future.",
      "votes": 5
    },
    {
      "id": 225720,
      "postDate": "2017-09-29T22:59:44.073Z",
      "content": "<p>The framework is much less important than GPUs in this particular task, but still. I really liked what people shared in <a href=\"https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40121\">https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40121</a></p>\n\n<p>It would be nice to know what software participants used and if there is any correlation between framework and result.</p>\n\n<p>To start: </p>\n\n<p>1st place. PyTorch</p>",
      "rawMarkdown": "The framework is much less important than GPUs in this particular task, but still. I really liked what people shared in https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40121\n\nIt would be nice to know what software participants used and if there is any correlation between framework and result.\n\nTo start: \n\n1st place. PyTorch",
      "votes": 5
    },
    {
      "id": 225848,
      "postDate": "2017-09-30T05:04:41.580Z",
      "content": "<p>5th: Keras/TF</p>",
      "rawMarkdown": "5th: Keras/TF",
      "votes": 3
    },
    {
      "id": 225821,
      "postDate": "2017-09-30T03:16:27.680Z",
      "content": "<p>The same team (1st place), and Pytorch as well.</p>",
      "rawMarkdown": "The same team (1st place), and Pytorch as well.",
      "votes": 3
    },
    {
      "id": 225774,
      "postDate": "2017-09-30T01:48:03.383Z",
      "content": "<p>13: pytorch and keras ( different team members uses different framework )</p>",
      "rawMarkdown": "13: pytorch and keras ( different team members uses different framework )",
      "votes": 4
    },
    {
      "id": 225725,
      "postDate": "2017-09-29T23:48:14.710Z",
      "content": "<p>7th mxnet</p>",
      "rawMarkdown": "7th mxnet",
      "votes": 4
    },
    {
      "id": 226178,
      "postDate": "2017-10-01T09:40:10.907Z",
      "content": "<p>6th place: pytorch, mxnet, keras(TF) </p>",
      "rawMarkdown": "6th place: pytorch, mxnet, keras(TF) ",
      "votes": 2
    },
    {
      "id": 231310,
      "postDate": "2017-10-14T11:11:53.253Z",
      "content": "<p>36th - Keras/TF</p>",
      "rawMarkdown": "36th - Keras/TF",
      "votes": 1
    },
    {
      "id": 226526,
      "postDate": "2017-10-02T15:49:17.383Z",
      "content": "<p>70th - PyTorch</p>",
      "rawMarkdown": "70th - PyTorch",
      "votes": 2
    },
    {
      "id": 226423,
      "postDate": "2017-10-02T08:22:20.220Z",
      "content": "<p>20th - Keras/TF</p>",
      "rawMarkdown": "20th - Keras/TF",
      "votes": 2
    },
    {
      "id": 226020,
      "postDate": "2017-09-30T18:08:52.073Z",
      "content": "<p>39th - Keras+TF</p>",
      "rawMarkdown": "39th - Keras+TF",
      "votes": 2
    },
    {
      "id": 225930,
      "postDate": "2017-09-30T12:45:57.513Z",
      "content": "<p>Keras/TF. Gave up on Keras/TH as it was way slower than TF to train but it has given me better results in prior competition.</p>",
      "rawMarkdown": "Keras/TF. Gave up on Keras/TH as it was way slower than TF to train but it has given me better results in prior competition.",
      "votes": 2
    },
    {
      "id": 225857,
      "postDate": "2017-09-30T06:33:36.247Z",
      "content": "<p>Keras/TF</p>",
      "rawMarkdown": "Keras/TF",
      "votes": 2
    },
    {
      "id": 225811,
      "postDate": "2017-09-30T02:32:56.027Z",
      "content": "<p>10th: Keras (TF backend)</p>",
      "rawMarkdown": "10th: Keras (TF backend)",
      "votes": 2
    },
    {
      "id": 227419,
      "postDate": "2017-10-04T09:25:06.290Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 226316,
      "author_name": "Nikolay Shebanov",
      "author_url": "",
      "post_date": "2017-10-01T20:22:42.593000",
      "content": "<p>4th place using a TF-backed Keras. </p>\n\n<p>Once I've bought some extra GPUs, I've tried to switch the whole pipeline to Pytorch. It took me a night or two to get it running, but I've faced some memory-related issues (most likely related to how I've used DataLoader) and turned back to Keras. </p>\n\n<p>Still, Pytorch feels much more transparent and intuitive, as well it had some influence on our later pipeline. I'm definitely will prefer it over Keras in the future.</p>",
      "votes": 5,
      "replies": []
    },
    {
      "id": 225848,
      "author_name": "Kyle Lee",
      "author_url": "",
      "post_date": "2017-09-30T05:04:41.580000",
      "content": "<p>5th: Keras/TF</p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 225821,
      "author_name": "Artem.Sanakoev",
      "author_url": "",
      "post_date": "2017-09-30T03:16:27.680000",
      "content": "<p>The same team (1st place), and Pytorch as well.</p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 225774,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2017-09-30T01:48:03.383000",
      "content": "<p>13: pytorch and keras ( different team members uses different framework )</p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 225725,
      "author_name": "Renat Bashirov",
      "author_url": "",
      "post_date": "2017-09-29T23:48:14.710000",
      "content": "<p>7th mxnet</p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 226178,
      "author_name": "n01z3",
      "author_url": "",
      "post_date": "2017-10-01T09:40:10.907000",
      "content": "<p>6th place: pytorch, mxnet, keras(TF) </p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 231310,
      "author_name": "YuhaoTang",
      "author_url": "",
      "post_date": "2017-10-14T11:11:53.253000",
      "content": "<p>36th - Keras/TF</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 226526,
      "author_name": "Matt Boyd",
      "author_url": "",
      "post_date": "2017-10-02T15:49:17.383000",
      "content": "<p>70th - PyTorch</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 226423,
      "author_name": "Igor Praznik",
      "author_url": "",
      "post_date": "2017-10-02T08:22:20.220000",
      "content": "<p>20th - Keras/TF</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 226020,
      "author_name": "Planckton",
      "author_url": "",
      "post_date": "2017-09-30T18:08:52.073000",
      "content": "<p>39th - Keras+TF</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 225930,
      "author_name": "eagle4",
      "author_url": "",
      "post_date": "2017-09-30T12:45:57.513000",
      "content": "<p>Keras/TF. Gave up on Keras/TH as it was way slower than TF to train but it has given me better results in prior competition.</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 225857,
      "author_name": "DmitryKustikov",
      "author_url": "",
      "post_date": "2017-09-30T06:33:36.247000",
      "content": "<p>Keras/TF</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 225811,
      "author_name": "David Austin",
      "author_url": "",
      "post_date": "2017-09-30T02:32:56.027000",
      "content": "<p>10th: Keras (TF backend)</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 227419,
      "author_name": "",
      "author_url": "",
      "post_date": "2017-10-04T09:25:06.290000",
      "content": "",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "226316": "4th place using a TF-backed Keras. \n\nOnce I've bought some extra GPUs, I've tried to switch the whole pipeline to Pytorch. It took me a night or two to get it running, but I've faced some memory-related issues (most likely related to how I've used DataLoader) and turned back to Keras. \n\nStill, Pytorch feels much more transparent and intuitive, as well it had some influence on our later pipeline. I'm definitely will prefer it over Keras in the future.",
    "225720": "The framework is much less important than GPUs in this particular task, but still. I really liked what people shared in https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40121\n\nIt would be nice to know what software participants used and if there is any correlation between framework and result.\n\nTo start: \n\n1st place. PyTorch",
    "225848": "5th: Keras/TF",
    "225821": "The same team (1st place), and Pytorch as well.",
    "225774": "13: pytorch and keras ( different team members uses different framework )",
    "225725": "7th mxnet",
    "226178": "6th place: pytorch, mxnet, keras(TF) ",
    "231310": "36th - Keras/TF",
    "226526": "70th - PyTorch",
    "226423": "20th - Keras/TF",
    "226020": "39th - Keras+TF",
    "225930": "Keras/TF. Gave up on Keras/TH as it was way slower than TF to train but it has given me better results in prior competition.",
    "225857": "Keras/TF",
    "225811": "10th: Keras (TF backend)",
    "227419": ""
  }
}