{
  "id": 155424,
  "title": "Some ideas that work for me",
  "url": "/competitions/prostate-cancer-grade-assessment/discussion/155424",
  "author_name": "Qishen Ha",
  "post_date": "2020-06-01T16:15:12.900000",
  "votes": 87,
  "comment_count": 28,
  "views": 0,
  "content": "<ul>\n<li>Using tiling method based on <a href=\"https://www.kaggle.com/iafoss/panda-16x128x128-tiles\">https://www.kaggle.com/iafoss/panda-16x128x128-tiles</a></li>\n<li>Simply setting the N = 36 and sz=256 then extract from median resolution</li>\n<li>Create 6x6 big image from 36 tiles</li>\n</ul>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F448347%2Fab6decd7abbd6f38be4b2204f1eac8e4%2F__results___9_1.png?generation=1591032799588488&amp;alt=media\" alt=\"\"></p>\n\n<ul>\n<li>Binning label\n<ul><li>E.g.\n<ul><li><code>label = [0,0,0,0,0]</code> means <code>isup_grade = 0</code></li>\n<li><code>label = [1,1,1,0,0]</code> means <code>isup_grade = 3</code></li>\n<li><code>label = [1,1,1,1,1]</code> means <code>isup_grade = 5</code></li></ul></li></ul></li>\n<li>BCE loss</li>\n<li>Augmentation on both tile level and big image level</li>\n<li>CosineAnnealingLR for one round</li>\n</ul>\n\n<p>I've publish my training &amp; inference code. For more details please refer to them:\n* Train: <a href=\"https://www.kaggle.com/haqishen/train-efficientnet-b0-w-36-tiles-256-lb0-87\">https://www.kaggle.com/haqishen/train-efficientnet-b0-w-36-tiles-256-lb0-87</a>\n* Inference: <a href=\"https://www.kaggle.com/haqishen/panda-inference-w-36-tiles-256\">https://www.kaggle.com/haqishen/panda-inference-w-36-tiles-256</a></p>",
  "messages": [
    {
      "id": 870322,
      "postDate": "2020-06-01T16:15:12.900Z",
      "content": "<ul>\n<li>Using tiling method based on <a href=\"https://www.kaggle.com/iafoss/panda-16x128x128-tiles\">https://www.kaggle.com/iafoss/panda-16x128x128-tiles</a></li>\n<li>Simply setting the N = 36 and sz=256 then extract from median resolution</li>\n<li>Create 6x6 big image from 36 tiles</li>\n</ul>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F448347%2Fab6decd7abbd6f38be4b2204f1eac8e4%2F__results___9_1.png?generation=1591032799588488&amp;alt=media\" alt=\"\"></p>\n\n<ul>\n<li>Binning label\n<ul><li>E.g.\n<ul><li><code>label = [0,0,0,0,0]</code> means <code>isup_grade = 0</code></li>\n<li><code>label = [1,1,1,0,0]</code> means <code>isup_grade = 3</code></li>\n<li><code>label = [1,1,1,1,1]</code> means <code>isup_grade = 5</code></li></ul></li></ul></li>\n<li>BCE loss</li>\n<li>Augmentation on both tile level and big image level</li>\n<li>CosineAnnealingLR for one round</li>\n</ul>\n\n<p>I've publish my training &amp; inference code. For more details please refer to them:\n* Train: <a href=\"https://www.kaggle.com/haqishen/train-efficientnet-b0-w-36-tiles-256-lb0-87\">https://www.kaggle.com/haqishen/train-efficientnet-b0-w-36-tiles-256-lb0-87</a>\n* Inference: <a href=\"https://www.kaggle.com/haqishen/panda-inference-w-36-tiles-256\">https://www.kaggle.com/haqishen/panda-inference-w-36-tiles-256</a></p>",
      "rawMarkdown": "* Using tiling method based on https://www.kaggle.com/iafoss/panda-16x128x128-tiles\n* Simply setting the N = 36 and sz=256 then extract from median resolution\n* Create 6x6 big image from 36 tiles\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F448347%2Fab6decd7abbd6f38be4b2204f1eac8e4%2F__results___9_1.png?generation=1591032799588488&amp;alt=media)\n\n\n* Binning label\n  * E.g.\n      * `label = [0,0,0,0,0]` means `isup_grade = 0`\n      * `label = [1,1,1,0,0]` means `isup_grade = 3`\n      * `label = [1,1,1,1,1]` means `isup_grade = 5`\n* BCE loss\n* Augmentation on both tile level and big image level\n* CosineAnnealingLR for one round\n\n\n\nI've publish my training &amp; inference code. For more details please refer to them:\n* Train: https://www.kaggle.com/haqishen/train-efficientnet-b0-w-36-tiles-256-lb0-87\n* Inference: https://www.kaggle.com/haqishen/panda-inference-w-36-tiles-256",
      "votes": 86
    },
    {
      "id": 873518,
      "postDate": "2020-06-04T08:03:57.093Z",
      "content": "<p>Just wondering why \"label binning\" would increase performance, as opposed to regression? Is it easier to optimize?</p>",
      "rawMarkdown": "Just wondering why \"label binning\" would increase performance, as opposed to regression? Is it easier to optimize?",
      "votes": 3
    },
    {
      "id": 870771,
      "postDate": "2020-06-01T22:39:53.593Z",
      "content": "<p>Label binning is quite an interesting idea. Did you compare it to just using regression? </p>",
      "rawMarkdown": "Label binning is quite an interesting idea. Did you compare it to just using regression? ",
      "votes": 3,
      "replies": [
        {
          "id": 870877,
          "postDate": "2020-06-02T02:18:32.260Z",
          "content": "<p>I blend it with a MSE model and got a tiny boost ;)</p>",
          "rawMarkdown": "I blend it with a MSE model and got a tiny boost ;)",
          "votes": 1
        }
      ]
    },
    {
      "id": 870834,
      "postDate": "2020-06-02T01:03:47.147Z",
      "content": "<p>Now the game is raising to next level, 0.87 would be the baseline score😂 \nLet's make this better!</p>",
      "rawMarkdown": "Now the game is raising to next level, 0.87 would be the baseline score😂 \nLet's make this better!",
      "votes": 4,
      "replies": [
        {
          "id": 873603,
          "postDate": "2020-06-04T09:52:42.443Z",
          "content": "<p>yeah, but I have to say, this guy is genius.</p>",
          "rawMarkdown": "yeah, but I have to say, this guy is genius.",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 875965,
      "postDate": "2020-06-06T10:21:41.310Z",
      "content": "<p>How do you manage to apply simple methods in all kind of comps and get a good score thats beyond me . Your codes are never complex to understand and your sharing is always very useful . Best of luck.</p>\n\n<p>P.S. For me in all NLP comps it has become a norm to use \"Qishen Ha's Multi Sample Dropout\" method :P </p>",
      "rawMarkdown": "How do you manage to apply simple methods in all kind of comps and get a good score thats beyond me . Your codes are never complex to understand and your sharing is always very useful . Best of luck.\n\nP.S. For me in all NLP comps it has become a norm to use \"Qishen Ha's Multi Sample Dropout\" method :P ",
      "votes": 3,
      "replies": [
        {
          "id": 876268,
          "postDate": "2020-06-06T15:32:59.573Z",
          "content": "<p>Thanks for your kind words ;)\nThe consequence may looks simple, but the process is not.\nWorking hard to explore the data, explore papers, then do experiments is the only way to get better score.\nIn my experience, almost no one can win a gold only by reading public discussions and kernels. (\nAs you don’t expect to win a gold by the information that everyone can easily get)</p>",
          "rawMarkdown": "Thanks for your kind words ;)\nThe consequence may looks simple, but the process is not.\nWorking hard to explore the data, explore papers, then do experiments is the only way to get better score.\nIn my experience, almost no one can win a gold only by reading public discussions and kernels. (\nAs you don’t expect to win a gold by the information that everyone can easily get)",
          "votes": 15
        }
      ]
    },
    {
      "id": 917743,
      "postDate": "2020-07-06T18:19:23.120Z",
      "content": "<p>Amazing kernel!</p>",
      "rawMarkdown": "Amazing kernel!",
      "votes": 1
    },
    {
      "id": 871947,
      "postDate": "2020-06-02T18:45:18.693Z",
      "content": "<p>Amazing ! Thanks a lot. Very authentic and interesting ideas.</p>",
      "rawMarkdown": "Amazing ! Thanks a lot. Very authentic and interesting ideas.",
      "votes": 1
    },
    {
      "id": 870915,
      "postDate": "2020-06-02T03:06:05.087Z",
      "content": "<p>Some interesting ideas there.\nDo you remember the improvement you got over tiles + pooling ?</p>",
      "rawMarkdown": "Some interesting ideas there.\nDo you remember the improvement you got over tiles + pooling ?",
      "votes": 2,
      "replies": [
        {
          "id": 871005,
          "postDate": "2020-06-02T04:56:58.933Z",
          "content": "<p>I got a significant improvement on local score but not that much on LB by replacing tiles + pooling with big image.</p>",
          "rawMarkdown": "I got a significant improvement on local score but not that much on LB by replacing tiles + pooling with big image.\n",
          "votes": 2
        }
      ]
    },
    {
      "id": 871130,
      "postDate": "2020-06-02T06:56:42.230Z",
      "content": "<p>Thanks for sharing! Ordinal regression is useful in grading problem, but do you compare it with simple classification or simple regression?</p>",
      "rawMarkdown": "Thanks for sharing! Ordinal regression is useful in grading problem, but do you compare it with simple classification or simple regression?",
      "replies": [
        {
          "id": 871374,
          "postDate": "2020-06-02T09:57:54.667Z",
          "content": "<p>It's not necessarily better than regression but it works when I did ensemble with regression model.</p>",
          "rawMarkdown": "It's not necessarily better than regression but it works when I did ensemble with regression model."
        },
        {
          "id": 871710,
          "postDate": "2020-06-02T15:31:00.137Z",
          "content": "<p>In your ordinal regression method, it is also possible to use thresholds optimized in oof to quantify predictions and get higher score (also higher risk of over-fitting);\nIn simple regression, we can also use the ‘round’ function to apply thresholds easily. \nSo I don't think there are much different in regression model ensemble. <br>\nDue to the instability of cv and lb, I can't decide which regression method is better. Combining multiple regression methods is a good idea, it may reduce the risk of shaking? 233</p>",
          "rawMarkdown": "In your ordinal regression method, it is also possible to use thresholds optimized in oof to quantify predictions and get higher score (also higher risk of over-fitting);\nIn simple regression, we can also use the ‘round’ function to apply thresholds easily. \nSo I don't think there are much different in regression model ensemble.  \nDue to the instability of cv and lb, I can't decide which regression method is better. Combining multiple regression methods is a good idea, it may reduce the risk of shaking? 233"
        }
      ]
    },
    {
      "id": 1759761,
      "postDate": "2022-04-18T21:41:59.113Z",
      "content": "<p>Good work how can we customize it for other fields sir …</p>",
      "rawMarkdown": "Good work how can we customize it for other fields sir ..."
    },
    {
      "id": 902237,
      "postDate": "2020-06-26T02:45:47.697Z",
      "content": "<p>I fail to get images like yours...\nMine are so sparse with lots of white regions.</p>",
      "rawMarkdown": "I fail to get images like yours...\nMine are so sparse with lots of white regions.",
      "replies": [
        {
          "id": 917981,
          "postDate": "2020-07-06T21:20:25.740Z",
          "content": "<p>this happen depending on the resolution\nimage = skimage.io.MultiImage(tiff_file)[-1] will give you that</p>\n\n<p>I have used like in the code\nimage = skimage.io.MultiImage(tiff_file)[1]</p>",
          "rawMarkdown": "this happen depending on the resolution\nimage = skimage.io.MultiImage(tiff_file)[-1] will give you that\n\nI have used like in the code\nimage = skimage.io.MultiImage(tiff_file)[1]"
        }
      ]
    },
    {
      "id": 885378,
      "postDate": "2020-06-14T06:35:58.113Z",
      "content": "<p>A rookie question, How will you compute the Kappa Score for binned labels?</p>",
      "rawMarkdown": "A rookie question, How will you compute the Kappa Score for binned labels?",
      "replies": [
        {
          "id": 885471,
          "postDate": "2020-06-14T08:46:25.103Z",
          "content": "<p>See my code out there ;)</p>",
          "rawMarkdown": "See my code out there ;)",
          "votes": 1
        }
      ]
    },
    {
      "id": 881114,
      "postDate": "2020-06-10T18:54:12.827Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true
    },
    {
      "id": 874594,
      "postDate": "2020-06-05T05:50:29.003Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true
    },
    {
      "id": 875737,
      "postDate": "2020-06-06T06:04:31.207Z",
      "content": "<p>Thanks for sharing</p>",
      "rawMarkdown": "Thanks for sharing",
      "votes": 4
    },
    {
      "id": 874617,
      "postDate": "2020-06-05T06:22:22.533Z",
      "content": "<p>Thanks for sharing, great!</p>",
      "rawMarkdown": "Thanks for sharing, great!",
      "votes": 1
    },
    {
      "id": 874395,
      "postDate": "2020-06-05T00:01:32.083Z",
      "content": "<p>Thanks!</p>",
      "rawMarkdown": "Thanks!",
      "votes": 1
    },
    {
      "id": 873570,
      "postDate": "2020-06-04T09:19:03.173Z",
      "content": "<p>Thanks for sharing cool stuff!!</p>",
      "rawMarkdown": "Thanks for sharing cool stuff!!",
      "votes": 1
    },
    {
      "id": 870967,
      "postDate": "2020-06-02T04:23:40.860Z",
      "content": "<p>Nice Idea. Thanks for sharing.</p>",
      "rawMarkdown": "Nice Idea. Thanks for sharing.",
      "votes": 1
    },
    {
      "id": 870339,
      "postDate": "2020-06-01T16:29:13.320Z",
      "content": "<p>Thanks for sharing!</p>",
      "rawMarkdown": "Thanks for sharing!",
      "votes": 1
    },
    {
      "id": 924102,
      "postDate": "2020-07-11T09:01:14.760Z",
      "content": "<p>Thanks for sharing</p>",
      "rawMarkdown": "Thanks for sharing"
    }
  ],
  "comments": [
    {
      "id": 873518,
      "author_name": "Green Tea",
      "author_url": "",
      "post_date": "2020-06-04T08:03:57.093000",
      "content": "<p>Just wondering why \"label binning\" would increase performance, as opposed to regression? Is it easier to optimize?</p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 870771,
      "author_name": "Shujun",
      "author_url": "",
      "post_date": "2020-06-01T22:39:53.593000",
      "content": "<p>Label binning is quite an interesting idea. Did you compare it to just using regression? </p>",
      "votes": 3,
      "replies": [
        {
          "id": 870877,
          "author_name": "Qishen Ha",
          "author_url": "",
          "post_date": "2020-06-02T02:18:32.260000",
          "content": "<p>I blend it with a MSE model and got a tiny boost ;)</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 870834,
      "author_name": "Tsai29",
      "author_url": "",
      "post_date": "2020-06-02T01:03:47.147000",
      "content": "<p>Now the game is raising to next level, 0.87 would be the baseline score😂 \nLet's make this better!</p>",
      "votes": 4,
      "replies": [
        {
          "id": 873603,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-06-04T09:52:42.443000",
          "content": "<p>yeah, but I have to say, this guy is genius.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 875965,
      "author_name": "Nirjhar Roy",
      "author_url": "",
      "post_date": "2020-06-06T10:21:41.310000",
      "content": "<p>How do you manage to apply simple methods in all kind of comps and get a good score thats beyond me . Your codes are never complex to understand and your sharing is always very useful . Best of luck.</p>\n\n<p>P.S. For me in all NLP comps it has become a norm to use \"Qishen Ha's Multi Sample Dropout\" method :P </p>",
      "votes": 3,
      "replies": [
        {
          "id": 876268,
          "author_name": "Qishen Ha",
          "author_url": "",
          "post_date": "2020-06-06T15:32:59.573000",
          "content": "<p>Thanks for your kind words ;)\nThe consequence may looks simple, but the process is not.\nWorking hard to explore the data, explore papers, then do experiments is the only way to get better score.\nIn my experience, almost no one can win a gold only by reading public discussions and kernels. (\nAs you don’t expect to win a gold by the information that everyone can easily get)</p>",
          "votes": 15,
          "replies": []
        }
      ]
    },
    {
      "id": 917743,
      "author_name": "Salman Ibne Eunus",
      "author_url": "",
      "post_date": "2020-07-06T18:19:23.120000",
      "content": "<p>Amazing kernel!</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 871947,
      "author_name": "F20180097",
      "author_url": "",
      "post_date": "2020-06-02T18:45:18.693000",
      "content": "<p>Amazing ! Thanks a lot. Very authentic and interesting ideas.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 870915,
      "author_name": "Arnaud Roussel",
      "author_url": "",
      "post_date": "2020-06-02T03:06:05.087000",
      "content": "<p>Some interesting ideas there.\nDo you remember the improvement you got over tiles + pooling ?</p>",
      "votes": 2,
      "replies": [
        {
          "id": 871005,
          "author_name": "Qishen Ha",
          "author_url": "",
          "post_date": "2020-06-02T04:56:58.933000",
          "content": "<p>I got a significant improvement on local score but not that much on LB by replacing tiles + pooling with big image.</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 871130,
      "author_name": "seefun",
      "author_url": "",
      "post_date": "2020-06-02T06:56:42.230000",
      "content": "<p>Thanks for sharing! Ordinal regression is useful in grading problem, but do you compare it with simple classification or simple regression?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 871374,
          "author_name": "Qishen Ha",
          "author_url": "",
          "post_date": "2020-06-02T09:57:54.667000",
          "content": "<p>It's not necessarily better than regression but it works when I did ensemble with regression model.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 871710,
          "author_name": "seefun",
          "author_url": "",
          "post_date": "2020-06-02T15:31:00.137000",
          "content": "<p>In your ordinal regression method, it is also possible to use thresholds optimized in oof to quantify predictions and get higher score (also higher risk of over-fitting);\nIn simple regression, we can also use the ‘round’ function to apply thresholds easily. \nSo I don't think there are much different in regression model ensemble. <br>\nDue to the instability of cv and lb, I can't decide which regression method is better. Combining multiple regression methods is a good idea, it may reduce the risk of shaking? 233</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1759761,
      "author_name": "Iqbal Ali",
      "author_url": "",
      "post_date": "2022-04-18T21:41:59.113000",
      "content": "<p>Good work how can we customize it for other fields sir …</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 902237,
      "author_name": "ChrisXie",
      "author_url": "",
      "post_date": "2020-06-26T02:45:47.697000",
      "content": "<p>I fail to get images like yours...\nMine are so sparse with lots of white regions.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 917981,
          "author_name": "TheStoneMX",
          "author_url": "",
          "post_date": "2020-07-06T21:20:25.740000",
          "content": "<p>this happen depending on the resolution\nimage = skimage.io.MultiImage(tiff_file)[-1] will give you that</p>\n\n<p>I have used like in the code\nimage = skimage.io.MultiImage(tiff_file)[1]</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 885378,
      "author_name": "Son of Anton v3.0",
      "author_url": "",
      "post_date": "2020-06-14T06:35:58.113000",
      "content": "<p>A rookie question, How will you compute the Kappa Score for binned labels?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 885471,
          "author_name": "Qishen Ha",
          "author_url": "",
          "post_date": "2020-06-14T08:46:25.103000",
          "content": "<p>See my code out there ;)</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 881114,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-06-10T18:54:12.827000",
      "content": "",
      "votes": 1,
      "replies": []
    },
    {
      "id": 874594,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-06-05T05:50:29.003000",
      "content": "",
      "votes": 1,
      "replies": []
    },
    {
      "id": 875737,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-06-06T06:04:31.207000",
      "content": "<p>Thanks for sharing</p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 874617,
      "author_name": "Burak  Batıbay",
      "author_url": "",
      "post_date": "2020-06-05T06:22:22.533000",
      "content": "<p>Thanks for sharing, great!</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 874395,
      "author_name": "Michał Ołtarzewski",
      "author_url": "",
      "post_date": "2020-06-05T00:01:32.083000",
      "content": "<p>Thanks!</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 873570,
      "author_name": "Niraj",
      "author_url": "",
      "post_date": "2020-06-04T09:19:03.173000",
      "content": "<p>Thanks for sharing cool stuff!!</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 870967,
      "author_name": "Karan",
      "author_url": "",
      "post_date": "2020-06-02T04:23:40.860000",
      "content": "<p>Nice Idea. Thanks for sharing.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 870339,
      "author_name": "Kurian Benoy",
      "author_url": "",
      "post_date": "2020-06-01T16:29:13.320000",
      "content": "<p>Thanks for sharing!</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 924102,
      "author_name": "Prachi",
      "author_url": "",
      "post_date": "2020-07-11T09:01:14.760000",
      "content": "<p>Thanks for sharing</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "870322": "* Using tiling method based on https://www.kaggle.com/iafoss/panda-16x128x128-tiles\n* Simply setting the N = 36 and sz=256 then extract from median resolution\n* Create 6x6 big image from 36 tiles\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F448347%2Fab6decd7abbd6f38be4b2204f1eac8e4%2F__results___9_1.png?generation=1591032799588488&amp;alt=media)\n\n\n* Binning label\n  * E.g.\n      * `label = [0,0,0,0,0]` means `isup_grade = 0`\n      * `label = [1,1,1,0,0]` means `isup_grade = 3`\n      * `label = [1,1,1,1,1]` means `isup_grade = 5`\n* BCE loss\n* Augmentation on both tile level and big image level\n* CosineAnnealingLR for one round\n\n\n\nI've publish my training &amp; inference code. For more details please refer to them:\n* Train: https://www.kaggle.com/haqishen/train-efficientnet-b0-w-36-tiles-256-lb0-87\n* Inference: https://www.kaggle.com/haqishen/panda-inference-w-36-tiles-256",
    "873518": "Just wondering why \"label binning\" would increase performance, as opposed to regression? Is it easier to optimize?",
    "870771": "Label binning is quite an interesting idea. Did you compare it to just using regression? ",
    "870834": "Now the game is raising to next level, 0.87 would be the baseline score😂 \nLet's make this better!",
    "875965": "How do you manage to apply simple methods in all kind of comps and get a good score thats beyond me . Your codes are never complex to understand and your sharing is always very useful . Best of luck.\n\nP.S. For me in all NLP comps it has become a norm to use \"Qishen Ha's Multi Sample Dropout\" method :P ",
    "917743": "Amazing kernel!",
    "871947": "Amazing ! Thanks a lot. Very authentic and interesting ideas.",
    "870915": "Some interesting ideas there.\nDo you remember the improvement you got over tiles + pooling ?",
    "871130": "Thanks for sharing! Ordinal regression is useful in grading problem, but do you compare it with simple classification or simple regression?",
    "1759761": "Good work how can we customize it for other fields sir ...",
    "902237": "I fail to get images like yours...\nMine are so sparse with lots of white regions.",
    "885378": "A rookie question, How will you compute the Kappa Score for binned labels?",
    "881114": "",
    "874594": "",
    "875737": "Thanks for sharing",
    "874617": "Thanks for sharing, great!",
    "874395": "Thanks!",
    "873570": "Thanks for sharing cool stuff!!",
    "870967": "Nice Idea. Thanks for sharing.",
    "870339": "Thanks for sharing!",
    "924102": "Thanks for sharing"
  }
}