{
  "id": 117301,
  "title": "6th place solution End to End Sequence to Sequence with sliding window.",
  "url": "/competitions/rsna-intracranial-hemorrhage-detection/discussion/117301",
  "author_name": "DavidGbodiOdaibo",
  "post_date": "2019-11-14T14:18:34.892000",
  "votes": 51,
  "comment_count": 25,
  "views": 0,
  "content": "<p>Congrats to all participants and the winners, and myself, I must say, for becoming a Kaggle GRANDMASTER!  Catching that elusive fifth Gold medal that I have been chasing for some time now.\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F369212%2Ff3b3bd56f8e374c317bc1fadc08a0a72%2FRSNA.png?generation=1573741035979705&amp;alt=media\" alt=\"\"></p>\n\n<p>Down to business. We utilized an end to end sequence to sequence model. A sliding window approach was used to select a fixed window of “n” slices from the ct volume. The FIG above shows the architecture. The architecture made things nice and simple, training was end to end, no data shuffling and gymnastics. Prediction on the slices was done by the LSTM at each time step. This conveniently also enabled some nice test time augmentation (TTA) with the sliding window approach.</p>",
  "messages": [
    {
      "id": 673104,
      "postDate": "2019-11-14T14:18:34.893Z",
      "content": "<p>Congrats to all participants and the winners, and myself, I must say, for becoming a Kaggle GRANDMASTER!  Catching that elusive fifth Gold medal that I have been chasing for some time now.\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F369212%2Ff3b3bd56f8e374c317bc1fadc08a0a72%2FRSNA.png?generation=1573741035979705&amp;alt=media\" alt=\"\"></p>\n\n<p>Down to business. We utilized an end to end sequence to sequence model. A sliding window approach was used to select a fixed window of “n” slices from the ct volume. The FIG above shows the architecture. The architecture made things nice and simple, training was end to end, no data shuffling and gymnastics. Prediction on the slices was done by the LSTM at each time step. This conveniently also enabled some nice test time augmentation (TTA) with the sliding window approach.</p>",
      "rawMarkdown": "Congrats to all participants and the winners, and myself, I must say, for becoming a Kaggle GRANDMASTER!  Catching that elusive fifth Gold medal that I have been chasing for some time now.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F369212%2Ff3b3bd56f8e374c317bc1fadc08a0a72%2FRSNA.png?generation=1573741035979705&amp;alt=media)\n\nDown to business. We utilized an end to end sequence to sequence model. A sliding window approach was used to select a fixed window of “n” slices from the ct volume. The FIG above shows the architecture. The architecture made things nice and simple, training was end to end, no data shuffling and gymnastics. Prediction on the slices was done by the LSTM at each time step. This conveniently also enabled some nice test time augmentation (TTA) with the sliding window approach.",
      "votes": 51
    },
    {
      "id": 673848,
      "postDate": "2019-11-15T14:54:07.887Z",
      "content": "<p>Congratulations on becoming a Kaggle GRANDMASTER! (It should always be fully capitalised ;)</p>\n\n<p>I wasn't following this competition, but am following you so this came up in my feed... I followed you because of <a href=\"https://www.kaggle.com/c/state-farm-distracted-driver-detection/discussion/22627#129848\">this great mixup idea</a> which at first I couldn't understand - working through the consequences of that idea lead me to understand ConvNets a lot better, so thank you for that!</p>\n\n<p>(It's my top example of a massively under-voted forum post :/)</p>\n\n<p>In fact, I was reminded of it recently when Jeremy Howard <a href=\"https://twitter.com/jeremyphoward/status/1143364106893684737\">tweeted a link to it</a>... Congrats on that too - not sure if you knew!</p>",
      "rawMarkdown": "Congratulations on becoming a Kaggle GRANDMASTER! (It should always be fully capitalised ;)\n\nI wasn't following this competition, but am following you so this came up in my feed... I followed you because of [this great mixup idea][1] which at first I couldn't understand - working through the consequences of that idea lead me to understand ConvNets a lot better, so thank you for that!\n\n(It's my top example of a massively under-voted forum post :/)\n\nIn fact, I was reminded of it recently when Jeremy Howard [tweeted a link to it][2]... Congrats on that too - not sure if you knew!\n\n [1]: https://www.kaggle.com/c/state-farm-distracted-driver-detection/discussion/22627#129848\n [2]: https://twitter.com/jeremyphoward/status/1143364106893684737\n\n",
      "votes": 4,
      "replies": [
        {
          "id": 673890,
          "postDate": "2019-11-15T15:51:21.127Z",
          "content": "<p>Wow! Thanks James for letting me know.  How do we fix this? need the credit!!! just kidding. Honored Jeremy is fighting for the little guy.</p>",
          "rawMarkdown": "Wow! Thanks James for letting me know.  How do we fix this? need the credit!!! just kidding. Honored Jeremy is fighting for the little guy.",
          "votes": 1
        },
        {
          "id": 673894,
          "postDate": "2019-11-15T15:57:32.997Z",
          "content": "<p>Haha :) I forgot to tag him - <a href=\"/jhoward\">@jhoward</a> ? (Will only work if he has notifications on 🤔 )</p>",
          "rawMarkdown": "Haha :) I forgot to tag him - @jhoward ? (Will only work if he has notifications on 🤔 )"
        },
        {
          "id": 674075,
          "postDate": "2019-11-15T21:37:40.140Z",
          "content": "<p>How do we fix what?...</p>",
          "rawMarkdown": "How do we fix what?..."
        },
        {
          "id": 674094,
          "postDate": "2019-11-15T22:06:08.327Z",
          "content": "<blockquote>\n  <p>How do we fix what?…</p>\n</blockquote>\n\n<p>David is both 6th place in this competition <em>and</em> the author of the CutMix augmentation post you mentioned in <a href=\"https://twitter.com/jeremyphoward/status/1143364106893684737\">the linked tweet</a> - a later tweet (by sebastienddoria) in the conversation is about how he has not been credited in later papers...</p>\n\n<p>David himself did say \"just kidding\" but I think it's a valid point...</p>\n\n<p>Introductions made, my work here is done :)</p>",
          "rawMarkdown": "&gt; How do we fix what?…\n\nDavid is both 6th place in this competition *and* the author of the CutMix augmentation post you mentioned in [the linked tweet][1] - a later tweet (by sebastienddoria) in the conversation is about how he has not been credited in later papers...\n\nDavid himself did say \"just kidding\" but I think it's a valid point...\n\nIntroductions made, my work here is done :)\n\n [1]: https://twitter.com/jeremyphoward/status/1143364106893684737\n",
          "votes": 1
        },
        {
          "id": 674096,
          "postDate": "2019-11-15T22:11:13.403Z",
          "content": "<p>Oh OK - hopefully I've already done my part, by pointing it out on Twitter. Other than that, someone I guess could reach out directly to the authors of the work to point out that they're missing a citation, but they've already been published so the best that could be helped for is to update the arxiv preprint.</p>",
          "rawMarkdown": "Oh OK - hopefully I've already done my part, by pointing it out on Twitter. Other than that, someone I guess could reach out directly to the authors of the work to point out that they're missing a citation, but they've already been published so the best that could be helped for is to update the arxiv preprint.",
          "votes": 2
        }
      ]
    },
    {
      "id": 673967,
      "postDate": "2019-11-15T18:11:50.247Z",
      "content": "<p><a href=\"/godaibo\">@godaibo</a> could u throw some more ight on slicing of N window approach. I see people are mentioning about it but i am yet to understand this more. </p>",
      "rawMarkdown": "@godaibo could u throw some more ight on slicing of N window approach. I see people are mentioning about it but i am yet to understand this more. \n",
      "votes": 1
    },
    {
      "id": 673906,
      "postDate": "2019-11-15T16:23:36.077Z",
      "content": "<p>Congrats David! </p>",
      "rawMarkdown": "Congrats David! ",
      "votes": 1
    },
    {
      "id": 673779,
      "postDate": "2019-11-15T13:27:50.897Z",
      "content": "<p>Congratulations for this win and for becoming Grand master.. 👍 </p>",
      "rawMarkdown": "Congratulations for this win and for becoming Grand master.. 👍 ",
      "votes": 1
    },
    {
      "id": 673177,
      "postDate": "2019-11-14T16:05:56.737Z",
      "content": "<p>Thank you for your sharing ! <a href=\"/godaibo\">@godaibo</a> \nWhat is your CNN backborn ?\nI think that this network needs much GPU memory for training ...</p>",
      "rawMarkdown": "Thank you for your sharing ! @godaibo \nWhat is your CNN backborn ?\nI think that this network needs much GPU memory for training ...",
      "votes": 1,
      "replies": [
        {
          "id": 673185,
          "postDate": "2019-11-14T16:20:04.850Z",
          "content": "<p>We used various backbones the best was a keras pretrained resinc. Don’t need much gpu memory On a 16G v100 can train various models with image size 512 x 512 and “n” of 10 and batch size 2. Same for the 16G p100. Larger batch sizes with Smaller n. Note that the CNN is just one CNN that generates the embedding vectors, the unrolled lstm looks like there are many but it is just one CNN.</p>",
          "rawMarkdown": "We used various backbones the best was a keras pretrained resinc. Don’t need much gpu memory On a 16G v100 can train various models with image size 512 x 512 and “n” of 10 and batch size 2. Same for the 16G p100. Larger batch sizes with Smaller n. Note that the CNN is just one CNN that generates the embedding vectors, the unrolled lstm looks like there are many but it is just one CNN."
        },
        {
          "id": 673479,
          "postDate": "2019-11-15T02:59:01.233Z",
          "content": "<p>batch_size 2 is very small for me ...\nHow long did it take to train one model on v100 ?</p>",
          "rawMarkdown": "batch_size 2 is very small for me ...\nHow long did it take to train one model on v100 ?",
          "votes": 1
        }
      ]
    },
    {
      "id": 673131,
      "postDate": "2019-11-14T14:53:14.343Z",
      "content": "<p>Double congratulations <a href=\"/godaibo\">@godaibo</a>  and thanks for sharing.</p>",
      "rawMarkdown": "Double congratulations @godaibo  and thanks for sharing.",
      "votes": 1
    },
    {
      "id": 673115,
      "postDate": "2019-11-14T14:37:56.890Z",
      "content": "<p>Congratulations\nGreat Write-Up\nThanks for Sharing your Valuable Insights &amp; Approach <a href=\"/godaibo\">@godaibo</a> </p>",
      "rawMarkdown": "Congratulations\nGreat Write-Up\nThanks for Sharing your Valuable Insights &amp; Approach @godaibo ",
      "votes": 1
    },
    {
      "id": 673109,
      "postDate": "2019-11-14T14:29:45.607Z",
      "content": "<p>Congratulations, very elegant solution!</p>",
      "rawMarkdown": "Congratulations, very elegant solution!",
      "votes": 1,
      "replies": [
        {
          "id": 673111,
          "postDate": "2019-11-14T14:31:27.593Z",
          "content": "<p>thx!</p>",
          "rawMarkdown": "thx!"
        }
      ]
    },
    {
      "id": 1049995,
      "postDate": "2020-10-15T01:00:17.987Z",
      "content": "<p>Congrats! Always learning a lot from these solution sharing-s :) </p>",
      "rawMarkdown": "Congrats! Always learning a lot from these solution sharing-s :) "
    },
    {
      "id": 782084,
      "postDate": "2020-03-21T23:05:55.677Z",
      "content": "<p>Anywhere can I find the code?</p>",
      "rawMarkdown": "Anywhere can I find the code?"
    },
    {
      "id": 771026,
      "postDate": "2020-03-13T17:09:02.387Z",
      "content": "<p>Why we have used lstm can you explain ??</p>",
      "rawMarkdown": "Why we have used lstm can you explain ??"
    },
    {
      "id": 682229,
      "postDate": "2019-11-27T05:22:28.033Z",
      "content": "<p>hi <a href=\"/godaibo\">@godaibo</a>, very elegant solution. May I know whether you have make your source code public?</p>\n\n<p>And do you find training such deep model end-to-end difficult (e.g., do you apply different learning rate to different parts, or train some parts first and then train the rest)?</p>",
      "rawMarkdown": "hi @godaibo, very elegant solution. May I know whether you have make your source code public?\n\nAnd do you find training such deep model end-to-end difficult (e.g., do you apply different learning rate to different parts, or train some parts first and then train the rest)?"
    },
    {
      "id": 674533,
      "postDate": "2019-11-16T16:46:14.627Z",
      "content": "<p>This is impressive. Congrats and thanks for sharing.</p>",
      "rawMarkdown": "This is impressive. Congrats and thanks for sharing."
    },
    {
      "id": 673810,
      "postDate": "2019-11-15T14:13:01.213Z",
      "content": "<p>Congrats! Cool end2end solution 🎉 Final sub used ensemble of various backbones?</p>",
      "rawMarkdown": "Congrats! Cool end2end solution 🎉 Final sub used ensemble of various backbones?",
      "replies": [
        {
          "id": 673813,
          "postDate": "2019-11-15T14:17:24.810Z",
          "content": "<p>5 models equally weighted ensemble (xception, inception, EfficientnetB0, Resnet50, ResInc)\nbest to worst (ResInc, Xception, EfficientnetB0, inception, Resnet50)</p>\n\n<p>We might be able to match or beat our final private score with fewer models or changing the ensemble weighting, have not checked yet.</p>",
          "rawMarkdown": "5 models equally weighted ensemble (xception, inception, EfficientnetB0, Resnet50, ResInc)\nbest to worst (ResInc, Xception, EfficientnetB0, inception, Resnet50)\n\nWe might be able to match or beat our final private score with fewer models or changing the ensemble weighting, have not checked yet.\n",
          "votes": 1
        },
        {
          "id": 681357,
          "postDate": "2019-11-26T02:26:20.907Z",
          "content": "<p>Thank you!</p>",
          "rawMarkdown": "Thank you!\n"
        }
      ]
    },
    {
      "id": 673434,
      "postDate": "2019-11-15T00:46:27.017Z",
      "content": "<p>Congrats and thanks for sharing.</p>",
      "rawMarkdown": "Congrats and thanks for sharing.",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 673848,
      "author_name": "James Trotman",
      "author_url": "",
      "post_date": "2019-11-15T14:54:07.887000",
      "content": "<p>Congratulations on becoming a Kaggle GRANDMASTER! (It should always be fully capitalised ;)</p>\n\n<p>I wasn't following this competition, but am following you so this came up in my feed... I followed you because of <a href=\"https://www.kaggle.com/c/state-farm-distracted-driver-detection/discussion/22627#129848\">this great mixup idea</a> which at first I couldn't understand - working through the consequences of that idea lead me to understand ConvNets a lot better, so thank you for that!</p>\n\n<p>(It's my top example of a massively under-voted forum post :/)</p>\n\n<p>In fact, I was reminded of it recently when Jeremy Howard <a href=\"https://twitter.com/jeremyphoward/status/1143364106893684737\">tweeted a link to it</a>... Congrats on that too - not sure if you knew!</p>",
      "votes": 4,
      "replies": [
        {
          "id": 673890,
          "author_name": "DavidGbodiOdaibo",
          "author_url": "",
          "post_date": "2019-11-15T15:51:21.127000",
          "content": "<p>Wow! Thanks James for letting me know.  How do we fix this? need the credit!!! just kidding. Honored Jeremy is fighting for the little guy.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 673894,
          "author_name": "James Trotman",
          "author_url": "",
          "post_date": "2019-11-15T15:57:32.997000",
          "content": "<p>Haha :) I forgot to tag him - <a href=\"/jhoward\">@jhoward</a> ? (Will only work if he has notifications on 🤔 )</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 674075,
          "author_name": "Jeremy Howard",
          "author_url": "",
          "post_date": "2019-11-15T21:37:40.140000",
          "content": "<p>How do we fix what?...</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 674094,
          "author_name": "James Trotman",
          "author_url": "",
          "post_date": "2019-11-15T22:06:08.327000",
          "content": "<blockquote>\n  <p>How do we fix what?…</p>\n</blockquote>\n\n<p>David is both 6th place in this competition <em>and</em> the author of the CutMix augmentation post you mentioned in <a href=\"https://twitter.com/jeremyphoward/status/1143364106893684737\">the linked tweet</a> - a later tweet (by sebastienddoria) in the conversation is about how he has not been credited in later papers...</p>\n\n<p>David himself did say \"just kidding\" but I think it's a valid point...</p>\n\n<p>Introductions made, my work here is done :)</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 674096,
          "author_name": "Jeremy Howard",
          "author_url": "",
          "post_date": "2019-11-15T22:11:13.403000",
          "content": "<p>Oh OK - hopefully I've already done my part, by pointing it out on Twitter. Other than that, someone I guess could reach out directly to the authors of the work to point out that they're missing a citation, but they've already been published so the best that could be helped for is to update the arxiv preprint.</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 673967,
      "author_name": "Jaideep",
      "author_url": "",
      "post_date": "2019-11-15T18:11:50.247000",
      "content": "<p><a href=\"/godaibo\">@godaibo</a> could u throw some more ight on slicing of N window approach. I see people are mentioning about it but i am yet to understand this more. </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 673906,
      "author_name": "ThomasAnthony",
      "author_url": "",
      "post_date": "2019-11-15T16:23:36.077000",
      "content": "<p>Congrats David! </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 673779,
      "author_name": "Kishore M",
      "author_url": "",
      "post_date": "2019-11-15T13:27:50.897000",
      "content": "<p>Congratulations for this win and for becoming Grand master.. 👍 </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 673177,
      "author_name": "shimacos",
      "author_url": "",
      "post_date": "2019-11-14T16:05:56.737000",
      "content": "<p>Thank you for your sharing ! <a href=\"/godaibo\">@godaibo</a> \nWhat is your CNN backborn ?\nI think that this network needs much GPU memory for training ...</p>",
      "votes": 1,
      "replies": [
        {
          "id": 673185,
          "author_name": "DavidGbodiOdaibo",
          "author_url": "",
          "post_date": "2019-11-14T16:20:04.850000",
          "content": "<p>We used various backbones the best was a keras pretrained resinc. Don’t need much gpu memory On a 16G v100 can train various models with image size 512 x 512 and “n” of 10 and batch size 2. Same for the 16G p100. Larger batch sizes with Smaller n. Note that the CNN is just one CNN that generates the embedding vectors, the unrolled lstm looks like there are many but it is just one CNN.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 673479,
          "author_name": "shimacos",
          "author_url": "",
          "post_date": "2019-11-15T02:59:01.233000",
          "content": "<p>batch_size 2 is very small for me ...\nHow long did it take to train one model on v100 ?</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 673131,
      "author_name": "YaGana Sheriff-Hussaini",
      "author_url": "",
      "post_date": "2019-11-14T14:53:14.343000",
      "content": "<p>Double congratulations <a href=\"/godaibo\">@godaibo</a>  and thanks for sharing.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 673115,
      "author_name": "Ailurophile",
      "author_url": "",
      "post_date": "2019-11-14T14:37:56.890000",
      "content": "<p>Congratulations\nGreat Write-Up\nThanks for Sharing your Valuable Insights &amp; Approach <a href=\"/godaibo\">@godaibo</a> </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 673109,
      "author_name": "Ian Pan",
      "author_url": "",
      "post_date": "2019-11-14T14:29:45.607000",
      "content": "<p>Congratulations, very elegant solution!</p>",
      "votes": 1,
      "replies": [
        {
          "id": 673111,
          "author_name": "DavidGbodiOdaibo",
          "author_url": "",
          "post_date": "2019-11-14T14:31:27.593000",
          "content": "<p>thx!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1049995,
      "author_name": "Seungjun (Josh) Kim",
      "author_url": "",
      "post_date": "2020-10-15T01:00:17.987000",
      "content": "<p>Congrats! Always learning a lot from these solution sharing-s :) </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 782084,
      "author_name": "vicky ",
      "author_url": "",
      "post_date": "2020-03-21T23:05:55.677000",
      "content": "<p>Anywhere can I find the code?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 771026,
      "author_name": "Naveen1711",
      "author_url": "",
      "post_date": "2020-03-13T17:09:02.387000",
      "content": "<p>Why we have used lstm can you explain ??</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 682229,
      "author_name": "TC Liu",
      "author_url": "",
      "post_date": "2019-11-27T05:22:28.033000",
      "content": "<p>hi <a href=\"/godaibo\">@godaibo</a>, very elegant solution. May I know whether you have make your source code public?</p>\n\n<p>And do you find training such deep model end-to-end difficult (e.g., do you apply different learning rate to different parts, or train some parts first and then train the rest)?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 674533,
      "author_name": "sleepysleeping",
      "author_url": "",
      "post_date": "2019-11-16T16:46:14.627000",
      "content": "<p>This is impressive. Congrats and thanks for sharing.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 673810,
      "author_name": "takuoko",
      "author_url": "",
      "post_date": "2019-11-15T14:13:01.213000",
      "content": "<p>Congrats! Cool end2end solution 🎉 Final sub used ensemble of various backbones?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 673813,
          "author_name": "DavidGbodiOdaibo",
          "author_url": "",
          "post_date": "2019-11-15T14:17:24.810000",
          "content": "<p>5 models equally weighted ensemble (xception, inception, EfficientnetB0, Resnet50, ResInc)\nbest to worst (ResInc, Xception, EfficientnetB0, inception, Resnet50)</p>\n\n<p>We might be able to match or beat our final private score with fewer models or changing the ensemble weighting, have not checked yet.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 681357,
          "author_name": "takuoko",
          "author_url": "",
          "post_date": "2019-11-26T02:26:20.907000",
          "content": "<p>Thank you!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 673434,
      "author_name": "corochann",
      "author_url": "",
      "post_date": "2019-11-15T00:46:27.017000",
      "content": "<p>Congrats and thanks for sharing.</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "673104": "Congrats to all participants and the winners, and myself, I must say, for becoming a Kaggle GRANDMASTER!  Catching that elusive fifth Gold medal that I have been chasing for some time now.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F369212%2Ff3b3bd56f8e374c317bc1fadc08a0a72%2FRSNA.png?generation=1573741035979705&amp;alt=media)\n\nDown to business. We utilized an end to end sequence to sequence model. A sliding window approach was used to select a fixed window of “n” slices from the ct volume. The FIG above shows the architecture. The architecture made things nice and simple, training was end to end, no data shuffling and gymnastics. Prediction on the slices was done by the LSTM at each time step. This conveniently also enabled some nice test time augmentation (TTA) with the sliding window approach.",
    "673848": "Congratulations on becoming a Kaggle GRANDMASTER! (It should always be fully capitalised ;)\n\nI wasn't following this competition, but am following you so this came up in my feed... I followed you because of [this great mixup idea][1] which at first I couldn't understand - working through the consequences of that idea lead me to understand ConvNets a lot better, so thank you for that!\n\n(It's my top example of a massively under-voted forum post :/)\n\nIn fact, I was reminded of it recently when Jeremy Howard [tweeted a link to it][2]... Congrats on that too - not sure if you knew!\n\n [1]: https://www.kaggle.com/c/state-farm-distracted-driver-detection/discussion/22627#129848\n [2]: https://twitter.com/jeremyphoward/status/1143364106893684737\n\n",
    "673967": "@godaibo could u throw some more ight on slicing of N window approach. I see people are mentioning about it but i am yet to understand this more. \n",
    "673906": "Congrats David! ",
    "673779": "Congratulations for this win and for becoming Grand master.. 👍 ",
    "673177": "Thank you for your sharing ! @godaibo \nWhat is your CNN backborn ?\nI think that this network needs much GPU memory for training ...",
    "673131": "Double congratulations @godaibo  and thanks for sharing.",
    "673115": "Congratulations\nGreat Write-Up\nThanks for Sharing your Valuable Insights &amp; Approach @godaibo ",
    "673109": "Congratulations, very elegant solution!",
    "1049995": "Congrats! Always learning a lot from these solution sharing-s :) ",
    "782084": "Anywhere can I find the code?",
    "771026": "Why we have used lstm can you explain ??",
    "682229": "hi @godaibo, very elegant solution. May I know whether you have make your source code public?\n\nAnd do you find training such deep model end-to-end difficult (e.g., do you apply different learning rate to different parts, or train some parts first and then train the rest)?",
    "674533": "This is impressive. Congrats and thanks for sharing.",
    "673810": "Congrats! Cool end2end solution 🎉 Final sub used ensemble of various backbones?",
    "673434": "Congrats and thanks for sharing."
  }
}