{
  "id": 58292,
  "title": "pytorch starter kit for pointnet for learning link",
  "url": "/competitions/trackml-particle-identification/discussion/58292",
  "author_name": "hengck23",
  "post_date": "2018-06-05T16:49:45.331000",
  "votes": 5,
  "comment_count": 9,
  "views": 0,
  "content": "<p>Note:</p>\n\n<ul>\n<li><p>this is not complete training code. It is mean as an illustration to see how pointnet can be used.</p></li>\n<li><p>the model structure, etc are not well tuned.</p></li>\n<li><p>the results shown is for over-fitting a train sample (NOT on test sample).</p></li>\n<li><p>please refer to attachment code for detail</p></li>\n<li><p>no submission to LB yet.</p></li>\n<li><p>results:</p></li>\n</ul>\n\n<p>gray: input hits and ground truth track</p>\n\n<p>red: estimated track</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/338725/9577/graph_pointnet.png\" alt=\"enter image description here\"></p>",
  "messages": [
    {
      "id": 338725,
      "postDate": "2018-06-05T16:49:45.330Z",
      "content": "<p>Note:</p>\n\n<ul>\n<li><p>this is not complete training code. It is mean as an illustration to see how pointnet can be used.</p></li>\n<li><p>the model structure, etc are not well tuned.</p></li>\n<li><p>the results shown is for over-fitting a train sample (NOT on test sample).</p></li>\n<li><p>please refer to attachment code for detail</p></li>\n<li><p>no submission to LB yet.</p></li>\n<li><p>results:</p></li>\n</ul>\n\n<p>gray: input hits and ground truth track</p>\n\n<p>red: estimated track</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/338725/9577/graph_pointnet.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "Note:\n\n- this is not complete training code. It is mean as an illustration to see how pointnet can be used.\n\n- the model structure, etc are not well tuned.\n\n- the results shown is for over-fitting a train sample (NOT on test sample).\n\n- please refer to attachment code for detail\n\n- no submission to LB yet.\n\n- results:\n\ngray: input hits and ground truth track\n\nred: estimated track\n\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/338725/9577/graph_pointnet.png",
      "votes": 5
    },
    {
      "id": 338817,
      "postDate": "2018-06-05T19:42:53.303Z",
      "content": "<p>Sorry for the off-topic question, but why do you choose pytorch over keras or tf?  I've seen that many top kagglers have made this choice.  Are there real advantages, or is it just what you are comfortable with?</p>\n\n<p>Thanks,\nJohn</p>",
      "rawMarkdown": "Sorry for the off-topic question, but why do you choose pytorch over keras or tf?  I've seen that many top kagglers have made this choice.  Are there real advantages, or is it just what you are comfortable with?\n\nThanks,\nJohn",
      "votes": 3
    },
    {
      "id": 339779,
      "postDate": "2018-06-07T16:12:54.420Z",
      "content": "<p>triplet generation sample code</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/339779/9585/animated.gif\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/339779/9584/Slide3.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "triplet generation sample code\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/339779/9585/animated.gif\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/339779/9584/Slide3.png",
      "votes": 4
    },
    {
      "id": 340323,
      "postDate": "2018-06-08T22:36:20.753Z",
      "content": "<p>check this:</p>\n\n<p><a href=\"https://indico.in2p3.fr/event/17295/contributions/61320/attachments/47567/59832/Neural_Network_Tracking_for_LHCb_Vertex_Detector.pdf\">https://indico.in2p3.fr/event/17295/contributions/61320/attachments/47567/59832/Neural_Network_Tracking_for_LHCb_Vertex_Detector.pdf</a></p>\n\n<p>So for each hit, we want to predict if the nearest 32 hits (phi-sorted) belong to the same track</p>",
      "rawMarkdown": "check this:\n\nhttps://indico.in2p3.fr/event/17295/contributions/61320/attachments/47567/59832/Neural_Network_Tracking_for_LHCb_Vertex_Detector.pdf\n\nSo for each hit, we want to predict if the nearest 32 hits (phi-sorted) belong to the same track",
      "votes": 2
    },
    {
      "id": 341211,
      "postDate": "2018-06-11T08:00:31.207Z",
      "content": "<p>this may be useful:</p>\n\n<p>pairwise distance + spectral clustering </p>\n\n<p>you can use deep learning to do spectral clustering. you just need to add a spectral clustering loss for back propagation</p>",
      "rawMarkdown": "this may be useful:\n\npairwise distance + spectral clustering \n\nyou can use deep learning to do spectral clustering. you just need to add a spectral clustering loss for back propagation"
    },
    {
      "id": 339225,
      "postDate": "2018-06-06T14:40:07.837Z",
      "content": "<p>improved results (over fitting on train set, not test set):</p>\n\n<p>gray: hits and ground truth tracks</p>\n\n<p>red: predicted pairwise link </p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/339225/9582/improved_train_fitting.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "improved results (over fitting on train set, not test set):\n\ngray: hits and ground truth tracks\n\nred: predicted pairwise link \n\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/339225/9582/improved_train_fitting.png",
      "replies": [
        {
          "id": 339242,
          "postDate": "2018-06-06T15:20:11.657Z",
          "content": "<p>Wow - impressive work!  What are the axis on your 2-d graph? z (0-900) vs phi (0-pi/2)?\nBTW - you are so productive...When do you sleep???????</p>",
          "rawMarkdown": "Wow - impressive work!  What are the axis on your 2-d graph? z (0-900) vs phi (0-pi/2)?\nBTW - you are so productive...When do you sleep???????"
        },
        {
          "id": 339244,
          "postDate": "2018-06-06T15:25:40.903Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 339451,
          "postDate": "2018-06-07T01:39:54.667Z",
          "content": "<p>basically it works. but there are still many issues to be solved, especially, generation of pairs for testing and training (e.g.  hard negative model). Actually accuracy of triplets is much better, but generation of triplets is more difficult.</p>\n\n<p>in faster rcnn, it uses 2 stages process, proposal generation first and then classification. it is end-to-end.</p>\n\n<p>i am trying pair generation first, and then extension to triplets in a end-to-end manner</p>",
          "rawMarkdown": "basically it works. but there are still many issues to be solved, especially, generation of pairs for testing and training (e.g.  hard negative model). Actually accuracy of triplets is much better, but generation of triplets is more difficult.\n\nin faster rcnn, it uses 2 stages process, proposal generation first and then classification. it is end-to-end.\n\ni am trying pair generation first, and then extension to triplets in a end-to-end manner\n\n ",
          "votes": 2
        },
        {
          "id": 339787,
          "postDate": "2018-06-07T16:40:30.007Z",
          "content": "<p>Would this be intended as a final step after clustering and extension, or as a replacement for extension? </p>\n\n<p>Cpmp recently commented that track extension improves his score less, possibility alluding to the fact that extension helps the shortcomings of the culstering, and might possibly be omitted in a final solution?</p>\n\n<p>Do you think pointnet could possibly be a standalone solution?</p>\n\n<p>Always a treat to see your work.</p>",
          "rawMarkdown": "Would this be intended as a final step after clustering and extension, or as a replacement for extension? \n\nCpmp recently commented that track extension improves his score less, possibility alluding to the fact that extension helps the shortcomings of the culstering, and might possibly be omitted in a final solution?\n\nDo you think pointnet could possibly be a standalone solution?\n\nAlways a treat to see your work.",
          "votes": 1
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 338817,
      "author_name": "John Sweeney",
      "author_url": "",
      "post_date": "2018-06-05T19:42:53.303000",
      "content": "<p>Sorry for the off-topic question, but why do you choose pytorch over keras or tf?  I've seen that many top kagglers have made this choice.  Are there real advantages, or is it just what you are comfortable with?</p>\n\n<p>Thanks,\nJohn</p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 339779,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-06-07T16:12:54.420000",
      "content": "<p>triplet generation sample code</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/339779/9585/animated.gif\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/339779/9584/Slide3.png\" alt=\"enter image description here\"></p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 340323,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-06-08T22:36:20.753000",
      "content": "<p>check this:</p>\n\n<p><a href=\"https://indico.in2p3.fr/event/17295/contributions/61320/attachments/47567/59832/Neural_Network_Tracking_for_LHCb_Vertex_Detector.pdf\">https://indico.in2p3.fr/event/17295/contributions/61320/attachments/47567/59832/Neural_Network_Tracking_for_LHCb_Vertex_Detector.pdf</a></p>\n\n<p>So for each hit, we want to predict if the nearest 32 hits (phi-sorted) belong to the same track</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 341211,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-06-11T08:00:31.207000",
      "content": "<p>this may be useful:</p>\n\n<p>pairwise distance + spectral clustering </p>\n\n<p>you can use deep learning to do spectral clustering. you just need to add a spectral clustering loss for back propagation</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 339225,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-06-06T14:40:07.837000",
      "content": "<p>improved results (over fitting on train set, not test set):</p>\n\n<p>gray: hits and ground truth tracks</p>\n\n<p>red: predicted pairwise link </p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/339225/9582/improved_train_fitting.png\" alt=\"enter image description here\"></p>",
      "votes": 0,
      "replies": [
        {
          "id": 339242,
          "author_name": "John Sweeney",
          "author_url": "",
          "post_date": "2018-06-06T15:20:11.657000",
          "content": "<p>Wow - impressive work!  What are the axis on your 2-d graph? z (0-900) vs phi (0-pi/2)?\nBTW - you are so productive...When do you sleep???????</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 339244,
          "author_name": "",
          "author_url": "",
          "post_date": "2018-06-06T15:25:40.903000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 339451,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2018-06-07T01:39:54.667000",
          "content": "<p>basically it works. but there are still many issues to be solved, especially, generation of pairs for testing and training (e.g.  hard negative model). Actually accuracy of triplets is much better, but generation of triplets is more difficult.</p>\n\n<p>in faster rcnn, it uses 2 stages process, proposal generation first and then classification. it is end-to-end.</p>\n\n<p>i am trying pair generation first, and then extension to triplets in a end-to-end manner</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 339787,
          "author_name": "macfarll",
          "author_url": "",
          "post_date": "2018-06-07T16:40:30.007000",
          "content": "<p>Would this be intended as a final step after clustering and extension, or as a replacement for extension? </p>\n\n<p>Cpmp recently commented that track extension improves his score less, possibility alluding to the fact that extension helps the shortcomings of the culstering, and might possibly be omitted in a final solution?</p>\n\n<p>Do you think pointnet could possibly be a standalone solution?</p>\n\n<p>Always a treat to see your work.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "338725": "Note:\n\n- this is not complete training code. It is mean as an illustration to see how pointnet can be used.\n\n- the model structure, etc are not well tuned.\n\n- the results shown is for over-fitting a train sample (NOT on test sample).\n\n- please refer to attachment code for detail\n\n- no submission to LB yet.\n\n- results:\n\ngray: input hits and ground truth track\n\nred: estimated track\n\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/338725/9577/graph_pointnet.png",
    "338817": "Sorry for the off-topic question, but why do you choose pytorch over keras or tf?  I've seen that many top kagglers have made this choice.  Are there real advantages, or is it just what you are comfortable with?\n\nThanks,\nJohn",
    "339779": "triplet generation sample code\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/339779/9585/animated.gif\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/339779/9584/Slide3.png",
    "340323": "check this:\n\nhttps://indico.in2p3.fr/event/17295/contributions/61320/attachments/47567/59832/Neural_Network_Tracking_for_LHCb_Vertex_Detector.pdf\n\nSo for each hit, we want to predict if the nearest 32 hits (phi-sorted) belong to the same track",
    "341211": "this may be useful:\n\npairwise distance + spectral clustering \n\nyou can use deep learning to do spectral clustering. you just need to add a spectral clustering loss for back propagation",
    "339225": "improved results (over fitting on train set, not test set):\n\ngray: hits and ground truth tracks\n\nred: predicted pairwise link \n\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/339225/9582/improved_train_fitting.png"
  }
}