{
  "id": 55832,
  "title": "Pytorch starter kit!",
  "url": "/competitions/trackml-particle-identification/discussion/55832",
  "author_name": "hengck23",
  "post_date": "2018-05-02T09:22:12.202000",
  "votes": 52,
  "comment_count": 19,
  "views": 0,
  "content": "<p>Baisc method: lstm network to repalce kalman filter based methods.\nPlease let me know if this method is suitable or not. Thanks!</p>\n\n<p>reference code:  <a href=\"https://github.com/HEPTrkX/heptrkx-ctd/blob/master/hit_classification/lstm_toy2D.ipynb\">https://github.com/HEPTrkX/heptrkx-ctd/blob/master/hit_classification/lstm_toy2D.ipynb</a></p>\n\n<p>paper: The HEP.TrkX Project: deep neural networks for HL-LHC online\nand offline tracking</p>\n\n<hr>\n\n<p>updated:</p>\n\n<ol>\n<li>initial tripletNet for seeding. please see folder \"0515\" and 0515.pptx. This is initial experiments. It does not make submission csv.</li>\n</ol>\n\n<p>link of code and results:</p>\n\n<p><a href=\"https://drive.google.com/open?id=1NXrPxRDyYldl3Ha_sGl_yJujdhmcTX6H\">https://drive.google.com/open?id=1NXrPxRDyYldl3Ha_sGl_yJujdhmcTX6H</a></p>",
  "messages": [
    {
      "id": 321994,
      "postDate": "2018-05-02T09:22:12.203Z",
      "content": "<p>Baisc method: lstm network to repalce kalman filter based methods.\nPlease let me know if this method is suitable or not. Thanks!</p>\n\n<p>reference code:  <a href=\"https://github.com/HEPTrkX/heptrkx-ctd/blob/master/hit_classification/lstm_toy2D.ipynb\">https://github.com/HEPTrkX/heptrkx-ctd/blob/master/hit_classification/lstm_toy2D.ipynb</a></p>\n\n<p>paper: The HEP.TrkX Project: deep neural networks for HL-LHC online\nand offline tracking</p>\n\n<hr>\n\n<p>updated:</p>\n\n<ol>\n<li>initial tripletNet for seeding. please see folder \"0515\" and 0515.pptx. This is initial experiments. It does not make submission csv.</li>\n</ol>\n\n<p>link of code and results:</p>\n\n<p><a href=\"https://drive.google.com/open?id=1NXrPxRDyYldl3Ha_sGl_yJujdhmcTX6H\">https://drive.google.com/open?id=1NXrPxRDyYldl3Ha_sGl_yJujdhmcTX6H</a></p>",
      "rawMarkdown": "\nBaisc method: lstm network to repalce kalman filter based methods.\nPlease let me know if this method is suitable or not. Thanks!\n\nreference code:  https://github.com/HEPTrkX/heptrkx-ctd/blob/master/hit_classification/lstm_toy2D.ipynb\n\npaper: The HEP.TrkX Project: deep neural networks for HL-LHC online\nand offline tracking\n\n---\n\n\nupdated:\n\n1.  initial tripletNet for seeding. please see folder \"0515\" and 0515.pptx. This is initial experiments. It does not make submission csv.\n\n\nlink of code and results:\n\nhttps://drive.google.com/open?id=1NXrPxRDyYldl3Ha_sGl_yJujdhmcTX6H\n\n\n",
      "votes": 52
    },
    {
      "id": 323808,
      "postDate": "2018-05-06T09:31:27.013Z",
      "content": "<p>how the ground truth look like</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/323808/9367/output1.gif\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "how the ground truth look like\n\n  ![enter image description here][1]\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/323808/9367/output1.gif",
      "votes": 10
    },
    {
      "id": 329306,
      "postDate": "2018-05-16T07:35:55.050Z",
      "content": "<p>extending the track:\n(colored: hit, black:fp)</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/329306/9430/extension.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "extending the track:\n(colored: hit, black:fp)\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/329306/9430/extension.png",
      "votes": 1,
      "replies": [
        {
          "id": 332028,
          "postDate": "2018-05-22T11:32:21.713Z",
          "content": "<p>@Heng are you doing something similar to this paper ? \n<a href=\"https://arxiv.org/pdf/1606.04990.pdf\">https://arxiv.org/pdf/1606.04990.pdf</a></p>",
          "rawMarkdown": "@Heng are you doing something similar to this paper ? \nhttps://arxiv.org/pdf/1606.04990.pdf"
        }
      ]
    },
    {
      "id": 328945,
      "postDate": "2018-05-15T12:21:19.307Z",
      "content": "<p>other information:</p>\n\n<p>paper: The HEP.TrkX Project: deep neural networks for HL-LHC online and offline tracking</p>\n\n<p>other information:</p>\n\n<p><a href=\"https://indico.fnal.gov/event/13497/session/1/contribution/18/material/slides/0.pdf\">https://indico.fnal.gov/event/13497/session/1/contribution/18/material/slides/0.pdf</a></p>\n\n<p><a href=\"https://indico.cern.ch/event/595059/contributions/2498118/attachments/1431635/2199380/03222017heptrkx_IML.pdf\">https://indico.cern.ch/event/595059/contributions/2498118/attachments/1431635/2199380/03222017heptrkx_IML.pdf</a></p>\n\n<p><a href=\"https://heptrkx.github.io/\">https://heptrkx.github.io/</a></p>\n\n<p><a href=\"https://dl4physicalsciences.github.io/files/nips_dlps_2017_28.pdf\">https://dl4physicalsciences.github.io/files/nips_dlps_2017_28.pdf</a></p>\n\n<p><a href=\"https://www.youtube.com/watch?v=QMuCSLWsks4\">https://www.youtube.com/watch?v=QMuCSLWsks4</a></p>",
      "rawMarkdown": "other information:\n\npaper: The HEP.TrkX Project: deep neural networks for HL-LHC online and offline tracking\n\nother information:\n\nhttps://indico.fnal.gov/event/13497/session/1/contribution/18/material/slides/0.pdf\n\nhttps://indico.cern.ch/event/595059/contributions/2498118/attachments/1431635/2199380/03222017heptrkx_IML.pdf\n\nhttps://heptrkx.github.io/\n\nhttps://dl4physicalsciences.github.io/files/nips_dlps_2017_28.pdf\n\nhttps://www.youtube.com/watch?v=QMuCSLWsks4",
      "votes": 1
    },
    {
      "id": 322710,
      "postDate": "2018-05-03T13:36:08.320Z",
      "content": "<p>some promising results on fitting train data:\n(different colors dots are from different detectors)</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322710/9358/results.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322710/9359/single.png\" alt=\"enter image description here\">\n   (highest scoring track)</p>",
      "rawMarkdown": "some promising results on fitting train data:\n(different colors dots are from different detectors)\n\n  ![enter image description here][1]\n\n  ![enter image description here][2]\n   (highest scoring track)\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322710/9358/results.png\n  [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322710/9359/single.png",
      "votes": 1
    },
    {
      "id": 330719,
      "postDate": "2018-05-19T14:23:30.053Z",
      "content": "<p>linking results. It is possible to divide 3d space into \"x&gt;0, y&gt;0, z&gt;0\", etc and use a single deep net to classify triplets</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/330719/9457/animated.gif\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "linking results. It is possible to divide 3d space into \"x&gt;0, y&gt;0, z&gt;0\", etc and use a single deep net to classify triplets\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/330719/9457/animated.gif",
      "votes": 2
    },
    {
      "id": 329369,
      "postDate": "2018-05-16T10:45:30.110Z",
      "content": "<p>some important tips in training deep networks (and also other learning machines):</p>\n\n<ul>\n<li><p>the measurements are in 1000's (x,y z hit values)</p></li>\n<li><p>however, the \"discriminative  values\" are small, e.g a hit of x=1000 to connected to another hit x=1001 but not  one with x=1002. To take care of this, either scale your data \"correctly\" or use \"appropriate learning rate\". You need network parameters to capture such subtle  \"discriminative  values\". e.g. if your learning rate is too large, the parameter cannot change from 123.1 to 123.10001.</p></li>\n</ul>\n\n<p>detecting different quadruples using a single network:</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/329369/9436/output.gif\" alt=\"enter image description here\"> </p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/329369/9435/00007.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "some important tips in training deep networks (and also other learning machines):\n\n  -  the measurements are in 1000's (x,y z hit values)\n  \n  - however, the \"discriminative  values\" are small, e.g a hit of x=1000 to connected to another hit x=1001 but not  one with x=1002. To take care of this, either scale your data \"correctly\" or use \"appropriate learning rate\". You need network parameters to capture such subtle  \"discriminative  values\". e.g. if your learning rate is too large, the parameter cannot change from 123.1 to 123.10001.\n\n\ndetecting different quadruples using a single network:\n\n  ![enter image description here][1] \n\n  ![enter image description here][2]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/329369/9436/output.gif\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/329369/9435/00007.png",
      "votes": 2
    },
    {
      "id": 328163,
      "postDate": "2018-05-13T14:15:59.467Z",
      "content": "<p>Dynamic Graph CNN for Learning on Point Clouds</p>\n\n<p><a href=\"https://arxiv.org/pdf/1801.07829.pdf\">https://arxiv.org/pdf/1801.07829.pdf</a></p>\n\n<p><a href=\"https://github.com/WangYueFt/dgcnn\">https://github.com/WangYueFt/dgcnn</a></p>",
      "rawMarkdown": "Dynamic Graph CNN for Learning on Point Clouds\n\nhttps://arxiv.org/pdf/1801.07829.pdf\n\nhttps://github.com/WangYueFt/dgcnn",
      "votes": 2
    },
    {
      "id": 329542,
      "postDate": "2018-05-16T17:20:12.940Z",
      "content": "<p>Hi, Heng, \nAre you using supervised classification for your LSTM/CNN models?\ndo you think using CNN auto encoder can work?</p>",
      "rawMarkdown": "Hi, Heng, \nAre you using supervised classification for your LSTM/CNN models?\ndo you think using CNN auto encoder can work?"
    },
    {
      "id": 328178,
      "postDate": "2018-05-13T15:21:04.230Z",
      "content": "<p>so cool</p>",
      "rawMarkdown": "so cool"
    },
    {
      "id": 328165,
      "postDate": "2018-05-13T14:22:47.357Z",
      "content": "<p>Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on Graphs</p>\n\n<p><a href=\"https://arxiv.org/pdf/1704.02901.pdf\">https://arxiv.org/pdf/1704.02901.pdf</a></p>\n\n<p><a href=\"https://github.com/mys007/ecc\">https://github.com/mys007/ecc</a>.</p>",
      "rawMarkdown": "Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on Graphs\n\nhttps://arxiv.org/pdf/1704.02901.pdf\n\n\nhttps://github.com/mys007/ecc.\n"
    },
    {
      "id": 328015,
      "postDate": "2018-05-13T06:42:08.463Z",
      "content": "<p>Point cnn</p>\n\n<p><a href=\"https://arxiv.org/pdf/1801.07791.pdf\">https://arxiv.org/pdf/1801.07791.pdf</a></p>\n\n<p><a href=\"https://github.com/yangyanli/PointCNN\">https://github.com/yangyanli/PointCNN</a></p>",
      "rawMarkdown": "Point cnn\n\nhttps://arxiv.org/pdf/1801.07791.pdf\n\nhttps://github.com/yangyanli/PointCNN"
    },
    {
      "id": 323268,
      "postDate": "2018-05-04T18:21:20.783Z",
      "content": "<p>Thanks for all your post, it is very helpful!! </p>",
      "rawMarkdown": "Thanks for all your post, it is very helpful!! \n"
    },
    {
      "id": 322716,
      "postDate": "2018-05-03T13:50:38.380Z",
      "content": "<p>i think this is a better method</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322716/9360/hough_net.png\" alt=\"enter image description here\"></p>\n\n<p>see also \"A neural implementation of the Hough transform and the advantages of\nexplaining away\" (<a href=\"https://nms.kcl.ac.uk/michael.spratling/Doc/pcbc_hough.pdf\">https://nms.kcl.ac.uk/michael.spratling/Doc/pcbc_hough.pdf</a>)</p>",
      "rawMarkdown": "i think this is a better method\n\n  ![enter image description here][1]\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322716/9360/hough_net.png\n\nsee also \"A neural implementation of the Hough transform and the advantages of\nexplaining away\" (https://nms.kcl.ac.uk/michael.spratling/Doc/pcbc_hough.pdf)"
    },
    {
      "id": 322486,
      "postDate": "2018-05-03T04:08:54.373Z",
      "content": "<p>at first i will just link up the hits. Then i will use lstm to parse them into separate tracks in later stage</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322486/9356/plan.png\" alt=\"enter image description here\"></p>\n\n<p>(graph convolution is also a possible candidate)</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322486/9357/plan2.png\" alt=\"enter image description here\"></p>\n\n<p>if the input space is too big, we can work on random subsample of points. this allows for future parallel processing.\ni.e.</p>\n\n<ol>\n<li><p>extract all hits in some volumes. Random divide into set1, set2 set3 ....</p></li>\n<li><p>apply connectNet to set1, set2, set3, ... independently.</p></li>\n<li><p>Ensemble results</p></li>\n</ol>",
      "rawMarkdown": "at first i will just link up the hits. Then i will use lstm to parse them into separate tracks in later stage\n\n  ![enter image description here][1]\n\n  \n\n(graph convolution is also a possible candidate)\n\n\n\n  ![enter image description here][2]\n\n\nif the input space is too big, we can work on random subsample of points. this allows for future parallel processing.\ni.e.\n\n1. extract all hits in some volumes. Random divide into set1, set2 set3 ....\n\n2. apply connectNet to set1, set2, set3, ... independently.\n\n3. Ensemble results\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322486/9356/plan.png\n  [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322486/9357/plan2.png"
    },
    {
      "id": 322274,
      "postDate": "2018-05-02T16:48:57.323Z",
      "content": "<p>problem setup</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322274/9341/problem.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "problem setup\n\n  ![enter image description here][1]\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322274/9341/problem.png"
    },
    {
      "id": 322871,
      "postDate": "2018-05-03T20:26:12.283Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 328976,
      "postDate": "2018-05-15T13:15:52.713Z",
      "content": "<p>very helpful thread. Thanks a lot</p>",
      "rawMarkdown": "very helpful thread. Thanks a lot"
    },
    {
      "id": 324253,
      "postDate": "2018-05-07T13:34:00.977Z",
      "content": "<p>thanks !!</p>",
      "rawMarkdown": "thanks !!"
    }
  ],
  "comments": [
    {
      "id": 323808,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-06T09:31:27.013000",
      "content": "<p>how the ground truth look like</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/323808/9367/output1.gif\" alt=\"enter image description here\"></p>",
      "votes": 10,
      "replies": []
    },
    {
      "id": 329306,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-16T07:35:55.050000",
      "content": "<p>extending the track:\n(colored: hit, black:fp)</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/329306/9430/extension.png\" alt=\"enter image description here\"></p>",
      "votes": 1,
      "replies": [
        {
          "id": 332028,
          "author_name": "AncientMonk",
          "author_url": "",
          "post_date": "2018-05-22T11:32:21.713000",
          "content": "<p>@Heng are you doing something similar to this paper ? \n<a href=\"https://arxiv.org/pdf/1606.04990.pdf\">https://arxiv.org/pdf/1606.04990.pdf</a></p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 328945,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-15T12:21:19.307000",
      "content": "<p>other information:</p>\n\n<p>paper: The HEP.TrkX Project: deep neural networks for HL-LHC online and offline tracking</p>\n\n<p>other information:</p>\n\n<p><a href=\"https://indico.fnal.gov/event/13497/session/1/contribution/18/material/slides/0.pdf\">https://indico.fnal.gov/event/13497/session/1/contribution/18/material/slides/0.pdf</a></p>\n\n<p><a href=\"https://indico.cern.ch/event/595059/contributions/2498118/attachments/1431635/2199380/03222017heptrkx_IML.pdf\">https://indico.cern.ch/event/595059/contributions/2498118/attachments/1431635/2199380/03222017heptrkx_IML.pdf</a></p>\n\n<p><a href=\"https://heptrkx.github.io/\">https://heptrkx.github.io/</a></p>\n\n<p><a href=\"https://dl4physicalsciences.github.io/files/nips_dlps_2017_28.pdf\">https://dl4physicalsciences.github.io/files/nips_dlps_2017_28.pdf</a></p>\n\n<p><a href=\"https://www.youtube.com/watch?v=QMuCSLWsks4\">https://www.youtube.com/watch?v=QMuCSLWsks4</a></p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 322710,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-03T13:36:08.320000",
      "content": "<p>some promising results on fitting train data:\n(different colors dots are from different detectors)</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322710/9358/results.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322710/9359/single.png\" alt=\"enter image description here\">\n   (highest scoring track)</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 330719,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-19T14:23:30.053000",
      "content": "<p>linking results. It is possible to divide 3d space into \"x&gt;0, y&gt;0, z&gt;0\", etc and use a single deep net to classify triplets</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/330719/9457/animated.gif\" alt=\"enter image description here\"></p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 329369,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-16T10:45:30.110000",
      "content": "<p>some important tips in training deep networks (and also other learning machines):</p>\n\n<ul>\n<li><p>the measurements are in 1000's (x,y z hit values)</p></li>\n<li><p>however, the \"discriminative  values\" are small, e.g a hit of x=1000 to connected to another hit x=1001 but not  one with x=1002. To take care of this, either scale your data \"correctly\" or use \"appropriate learning rate\". You need network parameters to capture such subtle  \"discriminative  values\". e.g. if your learning rate is too large, the parameter cannot change from 123.1 to 123.10001.</p></li>\n</ul>\n\n<p>detecting different quadruples using a single network:</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/329369/9436/output.gif\" alt=\"enter image description here\"> </p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/329369/9435/00007.png\" alt=\"enter image description here\"></p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 328163,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-13T14:15:59.467000",
      "content": "<p>Dynamic Graph CNN for Learning on Point Clouds</p>\n\n<p><a href=\"https://arxiv.org/pdf/1801.07829.pdf\">https://arxiv.org/pdf/1801.07829.pdf</a></p>\n\n<p><a href=\"https://github.com/WangYueFt/dgcnn\">https://github.com/WangYueFt/dgcnn</a></p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 329542,
      "author_name": "cosmojo",
      "author_url": "",
      "post_date": "2018-05-16T17:20:12.940000",
      "content": "<p>Hi, Heng, \nAre you using supervised classification for your LSTM/CNN models?\ndo you think using CNN auto encoder can work?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 328178,
      "author_name": "joejiong",
      "author_url": "",
      "post_date": "2018-05-13T15:21:04.230000",
      "content": "<p>so cool</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 328165,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-13T14:22:47.357000",
      "content": "<p>Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on Graphs</p>\n\n<p><a href=\"https://arxiv.org/pdf/1704.02901.pdf\">https://arxiv.org/pdf/1704.02901.pdf</a></p>\n\n<p><a href=\"https://github.com/mys007/ecc\">https://github.com/mys007/ecc</a>.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 328015,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-13T06:42:08.463000",
      "content": "<p>Point cnn</p>\n\n<p><a href=\"https://arxiv.org/pdf/1801.07791.pdf\">https://arxiv.org/pdf/1801.07791.pdf</a></p>\n\n<p><a href=\"https://github.com/yangyanli/PointCNN\">https://github.com/yangyanli/PointCNN</a></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 323268,
      "author_name": "aparrot",
      "author_url": "",
      "post_date": "2018-05-04T18:21:20.783000",
      "content": "<p>Thanks for all your post, it is very helpful!! </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 322716,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-03T13:50:38.380000",
      "content": "<p>i think this is a better method</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322716/9360/hough_net.png\" alt=\"enter image description here\"></p>\n\n<p>see also \"A neural implementation of the Hough transform and the advantages of\nexplaining away\" (<a href=\"https://nms.kcl.ac.uk/michael.spratling/Doc/pcbc_hough.pdf\">https://nms.kcl.ac.uk/michael.spratling/Doc/pcbc_hough.pdf</a>)</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 322486,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-03T04:08:54.373000",
      "content": "<p>at first i will just link up the hits. Then i will use lstm to parse them into separate tracks in later stage</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322486/9356/plan.png\" alt=\"enter image description here\"></p>\n\n<p>(graph convolution is also a possible candidate)</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322486/9357/plan2.png\" alt=\"enter image description here\"></p>\n\n<p>if the input space is too big, we can work on random subsample of points. this allows for future parallel processing.\ni.e.</p>\n\n<ol>\n<li><p>extract all hits in some volumes. Random divide into set1, set2 set3 ....</p></li>\n<li><p>apply connectNet to set1, set2, set3, ... independently.</p></li>\n<li><p>Ensemble results</p></li>\n</ol>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 322274,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-05-02T16:48:57.323000",
      "content": "<p>problem setup</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/322274/9341/problem.png\" alt=\"enter image description here\"></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 322871,
      "author_name": "",
      "author_url": "",
      "post_date": "2018-05-03T20:26:12.283000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 328976,
      "author_name": "rahul yadav",
      "author_url": "",
      "post_date": "2018-05-15T13:15:52.713000",
      "content": "<p>very helpful thread. Thanks a lot</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 324253,
      "author_name": "eksingh",
      "author_url": "",
      "post_date": "2018-05-07T13:34:00.977000",
      "content": "<p>thanks !!</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "321994": "\nBaisc method: lstm network to repalce kalman filter based methods.\nPlease let me know if this method is suitable or not. Thanks!\n\nreference code:  https://github.com/HEPTrkX/heptrkx-ctd/blob/master/hit_classification/lstm_toy2D.ipynb\n\npaper: The HEP.TrkX Project: deep neural networks for HL-LHC online\nand offline tracking\n\n---\n\n\nupdated:\n\n1.  initial tripletNet for seeding. please see folder \"0515\" and 0515.pptx. This is initial experiments. It does not make submission csv.\n\n\nlink of code and results:\n\nhttps://drive.google.com/open?id=1NXrPxRDyYldl3Ha_sGl_yJujdhmcTX6H\n\n\n",
    "323808": "how the ground truth look like\n\n  ![enter image description here][1]\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/323808/9367/output1.gif",
    "329306": "extending the track:\n(colored: hit, black:fp)\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/329306/9430/extension.png",
    "328945": "other information:\n\npaper: The HEP.TrkX Project: deep neural networks for HL-LHC online and offline tracking\n\nother information:\n\nhttps://indico.fnal.gov/event/13497/session/1/contribution/18/material/slides/0.pdf\n\nhttps://indico.cern.ch/event/595059/contributions/2498118/attachments/1431635/2199380/03222017heptrkx_IML.pdf\n\nhttps://heptrkx.github.io/\n\nhttps://dl4physicalsciences.github.io/files/nips_dlps_2017_28.pdf\n\nhttps://www.youtube.com/watch?v=QMuCSLWsks4",
    "322710": "some promising results on fitting train data:\n(different colors dots are from different detectors)\n\n  ![enter image description here][1]\n\n  ![enter image description here][2]\n   (highest scoring track)\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322710/9358/results.png\n  [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322710/9359/single.png",
    "330719": "linking results. It is possible to divide 3d space into \"x&gt;0, y&gt;0, z&gt;0\", etc and use a single deep net to classify triplets\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/330719/9457/animated.gif",
    "329369": "some important tips in training deep networks (and also other learning machines):\n\n  -  the measurements are in 1000's (x,y z hit values)\n  \n  - however, the \"discriminative  values\" are small, e.g a hit of x=1000 to connected to another hit x=1001 but not  one with x=1002. To take care of this, either scale your data \"correctly\" or use \"appropriate learning rate\". You need network parameters to capture such subtle  \"discriminative  values\". e.g. if your learning rate is too large, the parameter cannot change from 123.1 to 123.10001.\n\n\ndetecting different quadruples using a single network:\n\n  ![enter image description here][1] \n\n  ![enter image description here][2]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/329369/9436/output.gif\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/329369/9435/00007.png",
    "328163": "Dynamic Graph CNN for Learning on Point Clouds\n\nhttps://arxiv.org/pdf/1801.07829.pdf\n\nhttps://github.com/WangYueFt/dgcnn",
    "329542": "Hi, Heng, \nAre you using supervised classification for your LSTM/CNN models?\ndo you think using CNN auto encoder can work?",
    "328178": "so cool",
    "328165": "Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on Graphs\n\nhttps://arxiv.org/pdf/1704.02901.pdf\n\n\nhttps://github.com/mys007/ecc.\n",
    "328015": "Point cnn\n\nhttps://arxiv.org/pdf/1801.07791.pdf\n\nhttps://github.com/yangyanli/PointCNN",
    "323268": "Thanks for all your post, it is very helpful!! \n",
    "322716": "i think this is a better method\n\n  ![enter image description here][1]\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322716/9360/hough_net.png\n\nsee also \"A neural implementation of the Hough transform and the advantages of\nexplaining away\" (https://nms.kcl.ac.uk/michael.spratling/Doc/pcbc_hough.pdf)",
    "322486": "at first i will just link up the hits. Then i will use lstm to parse them into separate tracks in later stage\n\n  ![enter image description here][1]\n\n  \n\n(graph convolution is also a possible candidate)\n\n\n\n  ![enter image description here][2]\n\n\nif the input space is too big, we can work on random subsample of points. this allows for future parallel processing.\ni.e.\n\n1. extract all hits in some volumes. Random divide into set1, set2 set3 ....\n\n2. apply connectNet to set1, set2, set3, ... independently.\n\n3. Ensemble results\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322486/9356/plan.png\n  [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322486/9357/plan2.png",
    "322274": "problem setup\n\n  ![enter image description here][1]\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/322274/9341/problem.png",
    "322871": "",
    "328976": "very helpful thread. Thanks a lot",
    "324253": "thanks !!"
  }
}