{
  "id": 60054,
  "title": "free from origin?",
  "url": "/competitions/trackml-particle-identification/discussion/60054",
  "author_name": "outrunner",
  "post_date": "2018-06-29T17:26:22.188000",
  "votes": 12,
  "comment_count": 18,
  "views": 0,
  "content": "<p>Today I cross 0.88 on event000001001, but still gain more score in low \"vr range\" even though the loss in large vr (particles start far from origin) is higher. I think it is because I start searching tracks from origin, so I change the searching strategy.\n<img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350403/9730/trackML_8816.png\" alt=\"enter image description here\">\nThe results are as expected, I gain some score from high vr, and loss more in low vr. But the simple merging result is good, since it is not yet optimized. Maybe it is a good beginning toward to 0.9. </p>",
  "messages": [
    {
      "id": 350403,
      "postDate": "2018-06-29T17:26:22.190Z",
      "content": "<p>Today I cross 0.88 on event000001001, but still gain more score in low \"vr range\" even though the loss in large vr (particles start far from origin) is higher. I think it is because I start searching tracks from origin, so I change the searching strategy.\n<img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350403/9730/trackML_8816.png\" alt=\"enter image description here\">\nThe results are as expected, I gain some score from high vr, and loss more in low vr. But the simple merging result is good, since it is not yet optimized. Maybe it is a good beginning toward to 0.9. </p>",
      "rawMarkdown": "Today I cross 0.88 on event000001001, but still gain more score in low \"vr range\" even though the loss in large vr (particles start far from origin) is higher. I think it is because I start searching tracks from origin, so I change the searching strategy.\n![enter image description here][1]\nThe results are as expected, I gain some score from high vr, and loss more in low vr. But the simple merging result is good, since it is not yet optimized. Maybe it is a good beginning toward to 0.9. \n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/350403/9730/trackML_8816.png",
      "votes": 11
    },
    {
      "id": 350412,
      "postDate": "2018-06-29T18:06:17.637Z",
      "content": "<p>Thanks for the results. One question, how long does it take for you to generate results for one event?</p>",
      "rawMarkdown": "Thanks for the results. One question, how long does it take for you to generate results for one event?\n",
      "votes": 5,
      "replies": [
        {
          "id": 350564,
          "postDate": "2018-06-30T00:44:17.837Z",
          "content": "<p>10~20 hours for one event in one thread, so maybe take 10 days to make a submission. I made a \"great upgrade\" in the main stage of the pipeline to get a \"great boost\" around 0.006 these days, and it boosts the running time quadruple.</p>",
          "rawMarkdown": "10~20 hours for one event in one thread, so maybe take 10 days to make a submission. I made a \"great upgrade\" in the main stage of the pipeline to get a \"great boost\" around 0.006 these days, and it boosts the running time quadruple.",
          "votes": 9
        },
        {
          "id": 351596,
          "postDate": "2018-07-02T15:44:02.667Z",
          "content": "<p><a href=\"/outrunner\">@outrunner</a> - I'm curious what kind of setup you're using to get 10-20 hours per event</p>",
          "rawMarkdown": "@outrunner - I'm curious what kind of setup you're using to get 10-20 hours per event"
        },
        {
          "id": 351693,
          "postDate": "2018-07-02T20:14:52.587Z",
          "content": "<p>i7 + 1080ti</p>\n\n<p>It is easily to increase running time, for example, by doubling the for loop size, or do test time augmentation. </p>",
          "rawMarkdown": "i7 + 1080ti\n\nIt is easily to increase running time, for example, by doubling the for loop size, or do test time augmentation. "
        }
      ]
    },
    {
      "id": 350425,
      "postDate": "2018-06-29T18:30:24.207Z",
      "content": "<p>Amazing!  Makes me feel bad that I'm stuck at 0.6 😉</p>",
      "rawMarkdown": "Amazing!  Makes me feel bad that I'm stuck at 0.6 😉",
      "votes": 3
    },
    {
      "id": 352354,
      "postDate": "2018-07-04T06:41:41.710Z",
      "content": "<p>score distribution by pt:\n<img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/352354/9777/trackML_8863.png\" alt=\"enter image description here\">\nI think the most problem is (short) track length now.</p>",
      "rawMarkdown": "score distribution by pt:\n![enter image description here][1]\nI think the most problem is (short) track length now.\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/352354/9777/trackML_8863.png",
      "votes": 1,
      "replies": [
        {
          "id": 352389,
          "postDate": "2018-07-04T08:35:20.857Z",
          "content": "<p>Very nice analysis. A small note : what we (in physics) call pt is \"transverse momentum\" which is sqrt(px*<em>2+py</em>*2). It looks like \"your\" pt is rather our p (=modulus of momentum). We usually use pt rather than p because pt is related to the radius of curvature. Tracks with same phi and pt and different pz will approximately project on the same circle in (x,y) plane. (approximately because the magnetic field is not exactly constant in space and parallel to z axis).    vr=sqrt(vx*<em>2+vy</em>*2) is indeed the radius we use.</p>",
          "rawMarkdown": "Very nice analysis. A small note : what we (in physics) call pt is \"transverse momentum\" which is sqrt(px**2+py**2). It looks like \"your\" pt is rather our p (=modulus of momentum). We usually use pt rather than p because pt is related to the radius of curvature. Tracks with same phi and pt and different pz will approximately project on the same circle in (x,y) plane. (approximately because the magnetic field is not exactly constant in space and parallel to z axis).    vr=sqrt(vx**2+vy**2) is indeed the radius we use.",
          "votes": 2
        },
        {
          "id": 352395,
          "postDate": "2018-07-04T08:50:41.433Z",
          "content": "<p>Thanks for the explanation.</p>",
          "rawMarkdown": "Thanks for the explanation."
        },
        {
          "id": 352448,
          "postDate": "2018-07-04T12:05:59.173Z",
          "content": "<p><a href=\"/outrunner\">@outrunner</a> @David Thanks for the confirmation. I got it wrong as well, this way, even low pt tracks can be straight in the polar coordinates and can be easily found by dbscan clustering. I thought low PT particles would yield more curvy tracks and don't quite follow <code>z/r = constant</code>.  The scores of low pt and non-low-pt for this event are well distributed. </p>\n\n<pre><code>event000001001\n0.45802405 (&lt; 0.5 low PT tracks)\n0.5080315 ( &gt;=0.5 &amp; &lt;= 3.0)\n0.033944476 ( &gt; 3.0 high PT tracks)\n</code></pre>",
          "rawMarkdown": "@outrunner @David Thanks for the confirmation. I got it wrong as well, this way, even low pt tracks can be straight in the polar coordinates and can be easily found by dbscan clustering. I thought low PT particles would yield more curvy tracks and don't quite follow `z/r = constant`.  The scores of low pt and non-low-pt for this event are well distributed. \n\n    event000001001\n    0.45802405 (&lt; 0.5 low PT tracks)\n    0.5080315 ( &gt;=0.5 &amp; &lt;= 3.0)\n    0.033944476 ( &gt; 3.0 high PT tracks)\n",
          "votes": 1
        }
      ]
    },
    {
      "id": 350447,
      "postDate": "2018-06-29T19:20:28.417Z",
      "content": "<p><a href=\"/outrunner\">@outrunner</a> Do you use clustering based approch with transformations or something spicial ?if it is not a secret</p>",
      "rawMarkdown": "@outrunner Do you use clustering based approch with transformations or something spicial ?if it is not a secret",
      "votes": 1,
      "replies": [
        {
          "id": 350565,
          "postDate": "2018-06-30T00:51:13.230Z",
          "content": "<p>Seriously , I reconstruct the simulator.</p>\n\n<p>Just kidding, I use DL, but it is more like a brute-force search right now...</p>",
          "rawMarkdown": "Seriously , I reconstruct the simulator.\n\nJust kidding, I use DL, but it is more like a brute-force search right now...",
          "votes": 7
        },
        {
          "id": 350581,
          "postDate": "2018-06-30T02:23:43.413Z",
          "content": "<p>Thanks for the hint.</p>\n\n<p>So supervised deep learning is still the best! I suppose the problem is to generate candidates (which is combinatorial search) to be input the the deep network.</p>\n\n<p>I always believe that DL should work. Below show the hits of 2 events. Note the \"visual similarity\". I have better results using DL than hand-crafted approach, but have the problem of making DL run efficiently:</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350581/9734/compare.png\" alt=\"enter image description here\"></p>",
          "rawMarkdown": "Thanks for the hint.\n\nSo supervised deep learning is still the best! I suppose the problem is to generate candidates (which is combinatorial search) to be input the the deep network.\n\nI always believe that DL should work. Below show the hits of 2 events. Note the \"visual similarity\". I have better results using DL than hand-crafted approach, but have the problem of making DL run efficiently:\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/350581/9734/compare.png"
        },
        {
          "id": 350649,
          "postDate": "2018-06-30T06:46:39.493Z",
          "content": "<p>Thanks for your answer <a href=\"/outrunner\">@outrunner</a>\nBut can please help me answering one question.  I'm really just trying to understand whether I should invest my time into this or not. \nDid you use any <strong>High Energy Physics knowledge</strong> to make your submissions, or is it purely CV? </p>",
          "rawMarkdown": "Thanks for your answer @outrunner\nBut can please help me answering one question.  I'm really just trying to understand whether I should invest my time into this or not. \nDid you use any **High Energy Physics knowledge** to make your submissions, or is it purely CV? \n "
        },
        {
          "id": 350734,
          "postDate": "2018-06-30T10:39:32.987Z",
          "content": "<p>Back in 1991, I told the professors that neural networks were the future and that NN would just glance at the detector and do the analysis without any physics knowledge. They laughed. This competition has been an ample opportunity to prove them wrong, although now they are all emeritus or dead...</p>",
          "rawMarkdown": "Back in 1991, I told the professors that neural networks were the future and that NN would just glance at the detector and do the analysis without any physics knowledge. They laughed. This competition has been an ample opportunity to prove them wrong, although now they are all emeritus or dead...",
          "votes": 1
        },
        {
          "id": 350740,
          "postDate": "2018-06-30T10:52:54.473Z",
          "content": "<p>@Lasteg My High Energy Physics knowledge comes from the materials shared in this competition. As @Glimmung says, maybe the model learned something but I am not sure.</p>",
          "rawMarkdown": "@Lasteg My High Energy Physics knowledge comes from the materials shared in this competition. As @Glimmung says, maybe the model learned something but I am not sure.",
          "votes": 3
        },
        {
          "id": 351591,
          "postDate": "2018-07-02T15:33:46.510Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 350801,
      "postDate": "2018-06-30T13:37:30.363Z",
      "content": "<p>Wow,</p>\n\n<p>one week busy with something else and this competition has moved to another dimension!  Thanks for sharing, this is motivating.</p>\n\n<p>PS. This event is always giving me a much better score than the LB, I'm sure you selected this one for sharing because of that ;)  Nevertheless, even when taking this into account your local score is impressive.</p>",
      "rawMarkdown": "Wow,\n\none week busy with something else and this competition has moved to another dimension!  Thanks for sharing, this is motivating.\n\nPS. This event is always giving me a much better score than the LB, I'm sure you selected this one for sharing because of that ;)  Nevertheless, even when taking this into account your local score is impressive.",
      "replies": [
        {
          "id": 350819,
          "postDate": "2018-06-30T14:12:34.767Z",
          "content": "<p>Indeed, it is a simple event, the score is +0.006 at LB 0.807. And I only use this event to do optimization for efficiency.</p>",
          "rawMarkdown": "Indeed, it is a simple event, the score is +0.006 at LB 0.807. And I only use this event to do optimization for efficiency.",
          "votes": 1
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 350412,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-06-29T18:06:17.637000",
      "content": "<p>Thanks for the results. One question, how long does it take for you to generate results for one event?</p>",
      "votes": 5,
      "replies": [
        {
          "id": 350564,
          "author_name": "outrunner",
          "author_url": "",
          "post_date": "2018-06-30T00:44:17.837000",
          "content": "<p>10~20 hours for one event in one thread, so maybe take 10 days to make a submission. I made a \"great upgrade\" in the main stage of the pipeline to get a \"great boost\" around 0.006 these days, and it boosts the running time quadruple.</p>",
          "votes": 9,
          "replies": []
        },
        {
          "id": 351596,
          "author_name": "Poseidon",
          "author_url": "",
          "post_date": "2018-07-02T15:44:02.667000",
          "content": "<p><a href=\"/outrunner\">@outrunner</a> - I'm curious what kind of setup you're using to get 10-20 hours per event</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 351693,
          "author_name": "outrunner",
          "author_url": "",
          "post_date": "2018-07-02T20:14:52.587000",
          "content": "<p>i7 + 1080ti</p>\n\n<p>It is easily to increase running time, for example, by doubling the for loop size, or do test time augmentation. </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 350425,
      "author_name": "John Sweeney",
      "author_url": "",
      "post_date": "2018-06-29T18:30:24.207000",
      "content": "<p>Amazing!  Makes me feel bad that I'm stuck at 0.6 😉</p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 352354,
      "author_name": "outrunner",
      "author_url": "",
      "post_date": "2018-07-04T06:41:41.710000",
      "content": "<p>score distribution by pt:\n<img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/352354/9777/trackML_8863.png\" alt=\"enter image description here\">\nI think the most problem is (short) track length now.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 352389,
          "author_name": "David Rousseau",
          "author_url": "",
          "post_date": "2018-07-04T08:35:20.857000",
          "content": "<p>Very nice analysis. A small note : what we (in physics) call pt is \"transverse momentum\" which is sqrt(px*<em>2+py</em>*2). It looks like \"your\" pt is rather our p (=modulus of momentum). We usually use pt rather than p because pt is related to the radius of curvature. Tracks with same phi and pt and different pz will approximately project on the same circle in (x,y) plane. (approximately because the magnetic field is not exactly constant in space and parallel to z axis).    vr=sqrt(vx*<em>2+vy</em>*2) is indeed the radius we use.</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 352395,
          "author_name": "outrunner",
          "author_url": "",
          "post_date": "2018-07-04T08:50:41.433000",
          "content": "<p>Thanks for the explanation.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 352448,
          "author_name": "Nicole Finnie",
          "author_url": "",
          "post_date": "2018-07-04T12:05:59.173000",
          "content": "<p><a href=\"/outrunner\">@outrunner</a> @David Thanks for the confirmation. I got it wrong as well, this way, even low pt tracks can be straight in the polar coordinates and can be easily found by dbscan clustering. I thought low PT particles would yield more curvy tracks and don't quite follow <code>z/r = constant</code>.  The scores of low pt and non-low-pt for this event are well distributed. </p>\n\n<pre><code>event000001001\n0.45802405 (&lt; 0.5 low PT tracks)\n0.5080315 ( &gt;=0.5 &amp; &lt;= 3.0)\n0.033944476 ( &gt; 3.0 high PT tracks)\n</code></pre>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 350447,
      "author_name": "Arthur Stsepanenka",
      "author_url": "",
      "post_date": "2018-06-29T19:20:28.417000",
      "content": "<p><a href=\"/outrunner\">@outrunner</a> Do you use clustering based approch with transformations or something spicial ?if it is not a secret</p>",
      "votes": 1,
      "replies": [
        {
          "id": 350565,
          "author_name": "outrunner",
          "author_url": "",
          "post_date": "2018-06-30T00:51:13.230000",
          "content": "<p>Seriously , I reconstruct the simulator.</p>\n\n<p>Just kidding, I use DL, but it is more like a brute-force search right now...</p>",
          "votes": 7,
          "replies": []
        },
        {
          "id": 350581,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2018-06-30T02:23:43.413000",
          "content": "<p>Thanks for the hint.</p>\n\n<p>So supervised deep learning is still the best! I suppose the problem is to generate candidates (which is combinatorial search) to be input the the deep network.</p>\n\n<p>I always believe that DL should work. Below show the hits of 2 events. Note the \"visual similarity\". I have better results using DL than hand-crafted approach, but have the problem of making DL run efficiently:</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/350581/9734/compare.png\" alt=\"enter image description here\"></p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 350649,
          "author_name": "Vadzim Yermakou",
          "author_url": "",
          "post_date": "2018-06-30T06:46:39.493000",
          "content": "<p>Thanks for your answer <a href=\"/outrunner\">@outrunner</a>\nBut can please help me answering one question.  I'm really just trying to understand whether I should invest my time into this or not. \nDid you use any <strong>High Energy Physics knowledge</strong> to make your submissions, or is it purely CV? </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 350734,
          "author_name": "Tord Malmgren",
          "author_url": "",
          "post_date": "2018-06-30T10:39:32.987000",
          "content": "<p>Back in 1991, I told the professors that neural networks were the future and that NN would just glance at the detector and do the analysis without any physics knowledge. They laughed. This competition has been an ample opportunity to prove them wrong, although now they are all emeritus or dead...</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 350740,
          "author_name": "outrunner",
          "author_url": "",
          "post_date": "2018-06-30T10:52:54.473000",
          "content": "<p>@Lasteg My High Energy Physics knowledge comes from the materials shared in this competition. As @Glimmung says, maybe the model learned something but I am not sure.</p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 351591,
          "author_name": "",
          "author_url": "",
          "post_date": "2018-07-02T15:33:46.510000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 350801,
      "author_name": "CPMP",
      "author_url": "",
      "post_date": "2018-06-30T13:37:30.363000",
      "content": "<p>Wow,</p>\n\n<p>one week busy with something else and this competition has moved to another dimension!  Thanks for sharing, this is motivating.</p>\n\n<p>PS. This event is always giving me a much better score than the LB, I'm sure you selected this one for sharing because of that ;)  Nevertheless, even when taking this into account your local score is impressive.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 350819,
          "author_name": "outrunner",
          "author_url": "",
          "post_date": "2018-06-30T14:12:34.767000",
          "content": "<p>Indeed, it is a simple event, the score is +0.006 at LB 0.807. And I only use this event to do optimization for efficiency.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "350403": "Today I cross 0.88 on event000001001, but still gain more score in low \"vr range\" even though the loss in large vr (particles start far from origin) is higher. I think it is because I start searching tracks from origin, so I change the searching strategy.\n![enter image description here][1]\nThe results are as expected, I gain some score from high vr, and loss more in low vr. But the simple merging result is good, since it is not yet optimized. Maybe it is a good beginning toward to 0.9. \n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/350403/9730/trackML_8816.png",
    "350412": "Thanks for the results. One question, how long does it take for you to generate results for one event?\n",
    "350425": "Amazing!  Makes me feel bad that I'm stuck at 0.6 😉",
    "352354": "score distribution by pt:\n![enter image description here][1]\nI think the most problem is (short) track length now.\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/352354/9777/trackML_8863.png",
    "350447": "@outrunner Do you use clustering based approch with transformations or something spicial ?if it is not a secret",
    "350801": "Wow,\n\none week busy with something else and this competition has moved to another dimension!  Thanks for sharing, this is motivating.\n\nPS. This event is always giving me a much better score than the LB, I'm sure you selected this one for sharing because of that ;)  Nevertheless, even when taking this into account your local score is impressive."
  }
}