Congratulations to all participants of this amazing competition! As organizers it was fantastic to watch the quick evolution of the score and the forum interactions as well as the posted kernels. We are looking forward to learning more details on how you did it.
\n\nThe (temporary) private leaderboard is now revealed and winners are: #1 Top Quarks (icecuber+ersol), #2 outrunner, #3 Sergey Gorbunov. Bravo!
\n\nWe are happy to see that the private leaderboard is identical to the public one for ranks 1 to 19. There has been no luck factor!
\n\nPlease check our last post on https://twitter.com/trackmllhc with some interesting plots (but all other relevant information is in this topic).
\n\nWhat happens now:
\n\nTop scorers will be invited to submit their software (or better making it public, because organizers will do this anyway) with an open source license within 2 weeks, and to release some structured documentation on their software (precise indication how to do so in my reply below). This is mandatory to be able to claim the monetary prizes 12.000$ for #1, 8.000$ for #2, 5.000$ for #3.
Anyone (even not top scorers) who thinks they made valuable contributions with innovative algorithms (in particular if they have shared insights on the forum) will be welcome to release publicly their code with an open source license and lightweight structured documentation also within two weeks (instructions in my reply below). This will allow to compete for the « HEP meets ML » jury prizes: a NVidia V100 GPU, and two invitations to NIPS 2018 or the spring 2019 CERN workshop.
2b : The survey indicated in my reply below can also be filled by other participants, as a mean for feedback to the competition
\n\nThe second ''Throughput\" phase of the competition will start early September on Codalab platform (this is an official NIPS competition). The dataset will be very similar, with correction of bug/features (most notable one being the Z spread of the origin of the particles which increases from 5.5mm to 55mm). The scoring will combine accuracy score and speed. Accuracy score is the same as for the first phase, except that only particles originating from a narrow cylinder around the z axis will be accounted for (so that we will focus on the « easy » particles). Speed is measured on a single core CPU. Participation to the Throughput phase is open to anyone. Using software released by Accuracy phase participants is of course authorized.\nWinners of that phase will share 15k$ monetary prize, and there will be a second set of « HEP meets ML » jury prizes, another Nvidia V100 GPU, and 2 invitation to CERN or NIPS.
Meanwhile the international jury with both physicists and computer scientists will select by end September the most interesting submissions of the first phase to attribute the special \"HEP meet ML\" prizes.
Congratulations to all participants of this amazing competition! As organizers it was fantastic to watch the quick evolution of the score and the forum interactions as well as the posted kernels. We are looking forward to learning more details on how you did it.
\n\nThe (temporary) private leaderboard is now revealed and winners are: #1 Top Quarks (icecuber+ersol), #2 outrunner, #3 Sergey Gorbunov. Bravo!
\n\nWe are happy to see that the private leaderboard is identical to the public one for ranks 1 to 19. There has been no luck factor!
\n\nPlease check our last post on https://twitter.com/trackmllhc with some interesting plots (but all other relevant information is in this topic).
\n\nWhat happens now:
\n\nTop scorers will be invited to submit their software (or better making it public, because organizers will do this anyway) with an open source license within 2 weeks, and to release some structured documentation on their software (precise indication how to do so in my reply below). This is mandatory to be able to claim the monetary prizes 12.000$ for #1, 8.000$ for #2, 5.000$ for #3.
Anyone (even not top scorers) who thinks they made valuable contributions with innovative algorithms (in particular if they have shared insights on the forum) will be welcome to release publicly their code with an open source license and lightweight structured documentation also within two weeks (instructions in my reply below). This will allow to compete for the « HEP meets ML » jury prizes: a NVidia V100 GPU, and two invitations to NIPS 2018 or the spring 2019 CERN workshop.
2b : The survey indicated in my reply below can also be filled by other participants, as a mean for feedback to the competition
\n\nThe second ''Throughput\" phase of the competition will start early September on Codalab platform (this is an official NIPS competition). The dataset will be very similar, with correction of bug/features (most notable one being the Z spread of the origin of the particles which increases from 5.5mm to 55mm). The scoring will combine accuracy score and speed. Accuracy score is the same as for the first phase, except that only particles originating from a narrow cylinder around the z axis will be accounted for (so that we will focus on the « easy » particles). Speed is measured on a single core CPU. Participation to the Throughput phase is open to anyone. Using software released by Accuracy phase participants is of course authorized.\nWinners of that phase will share 15k$ monetary prize, and there will be a second set of « HEP meets ML » jury prizes, another Nvidia V100 GPU, and 2 invitation to CERN or NIPS.
Meanwhile the international jury with both physicists and computer scientists will select by end September the most interesting submissions of the first phase to attribute the special \"HEP meet ML\" prizes.
Congratulations to winners and all other participants)
", "rawMarkdown": "Congratulations to winners and all other participants)", "votes": null }, { "id": "370586", "postDate": "08/15/2018 04:35:18", "content": "Thanks for these detailed instruction.
\n\n\n\n\nwith correction of bug/features (most notable one being the Z spread of the origin of the particles which increases from 5.5mm to 55mm).
\n
The magnetic field variation impacts trajectories way more than what is described in the documents you shared. Indeed, it reads that departure from perfect helix should not exceed few millimeters. In reality departure from perfect helix can be way larger. I'm pointing this out to make sure you do not have a scaling issue for the magnetic field direction as well.
\n\n\n\n\nThe scoring will combine accuracy score and speed. Accuracy score is the same as for the first phase, except that only particles originating from a narrow cylinder around the z axis will be accounted for (so that we will focus on the « easy » particles).
\n
I wish this was the case for the first phase as clustering approaches would be way closer to the top score ! ;)
", "rawMarkdown": "Thanks for these detailed instruction.\n\n> with correction of bug/features (most notable one being the Z spread of the origin of the particles which increases from 5.5mm to 55mm). \n\nThe magnetic field variation impacts trajectories way more than what is described in the documents you shared. Indeed, it reads that departure from perfect helix should not exceed few millimeters. In reality departure from perfect helix can be way larger. I'm pointing this out to make sure you do not have a scaling issue for the magnetic field direction as well.\n\n> The scoring will combine accuracy score and speed. Accuracy score is the same as for the first phase, except that only particles originating from a narrow cylinder around the z axis will be accounted for (so that we will focus on the « easy » particles). \n\nI wish this was the case for the first phase as clustering approaches would be way closer to the top score ! ;)", "votes": null }, { "id": "370693", "postDate": "08/15/2018 09:53:31", "content": "\n\n\nIndeed, it reads that departure from perfect helix should not exceed few millimeters.
\n
Judging by the intersection displacements learned by my algorithm (see the \"Pretty pictures\" post), this is mostly true if one considers only the systematic deviations due to the magnetic field. Only in the end cap regions do the deviations become of the order of 10 mm or so, if I remember correctly. The consideration of these systematic deviations therefore gave me only a small score improvement (~1%).
", "rawMarkdown": "> Indeed, it reads that departure from perfect helix should not exceed few millimeters.\n\nJudging by the intersection displacements learned by my algorithm (see the \"Pretty pictures\" post), this is mostly true if one considers only the *systematic* deviations due to the magnetic field. Only in the end cap regions do the deviations become of the order of 10 mm or so, if I remember correctly. The consideration of these systematic deviations therefore gave me only a small score improvement (~1%).", "votes": null }, { "id": "370825", "postDate": "08/15/2018 14:02:08", "content": "My correction of the unrolling angle gave me 0.03 score improvement, at least.
", "rawMarkdown": "My correction of the unrolling angle gave me 0.03 score improvement, at least.", "votes": null }, { "id": "371144", "postDate": "08/16/2018 05:15:27", "content": "So participants #1 #2 and #3, and all participants applying to HEP meets ML jury prize (don't be shy!), should *before Monday 27th August 11:59PM UTC:
\n\nWe are also welcoming the feedback of other participants by filling in the survey (without releaseing software or document).
\n\nQuestions by replying here, or using trackml.contact@gmail.com
", "rawMarkdown": "So participants #1 #2 and #3, and all participants applying to HEP meets ML jury prize (don't be shy!), should *before Monday 27th August 11:59PM UTC:\n\n 1. release their software with an open source license as indicated in section B of the attached document (github or the like is preferred)\n 2. write a short pdf document detailing their methods following template in section A of the attached document. (many of you have already posted detailed info on their method on the forum, which is good of course, but please take the time to reformat the information according to the template, this will simplify considerably the work of the jury)\n 3. post the two public links to the forum (under the topic you might have already created) or a new one\n 4. fill the survey https://goo.gl/forms/hADGfrOjKw2ws91G3 which will only be seen by Kaggle, the TrackML organizers and the jury\n\nWe are also welcoming the feedback of other participants by filling in the survey (without releaseing software or document).\n\nQuestions by replying here, or using trackml.contact@gmail.com", "votes": null }, { "id": "371228", "postDate": "08/16/2018 09:41:43", "content": "One suggestion if I may: if you use conda or anaconda, then you can export your environment to a yaml file that contains all the packages you use with their version, see https://conda.io/docs/user-guide/tasks/manage-environments.html#sharing-an-environment . You can then easily recreate the same environment from the file, see https://conda.io/docs/user-guide/tasks/manage-environments.html#creating-an-environment-from-an-environment-yml-file
\n\nThis is a standard way to describe the Python dependencies you have, and I would recommend you accept it as an alternative to requirements.txt .
", "rawMarkdown": "One suggestion if I may: if you use conda or anaconda, then you can export your environment to a yaml file that contains all the packages you use with their version, see https://conda.io/docs/user-guide/tasks/manage-environments.html#sharing-an-environment . You can then easily recreate the same environment from the file, see https://conda.io/docs/user-guide/tasks/manage-environments.html#creating-an-environment-from-an-environment-yml-file\n\nThis is a standard way to describe the Python dependencies you have, and I would recommend you accept it as an alternative to requirements.txt .", "votes": null }, { "id": "371323", "postDate": "08/16/2018 14:39:49", "content": "Agreed
", "rawMarkdown": "Agreed", "votes": null }, { "id": "372825", "postDate": "08/20/2018 12:27:24", "content": "I do not know if this was posted before, but could we have an idea of the improvement on the score between organizers's initial solution and the final, winning one ? \nThis is always very interesting to have an idea of the use of this kind of competition.
\n\nThanks!
", "rawMarkdown": "I do not know if this was posted before, but could we have an idea of the improvement on the score between organizers's initial solution and the final, winning one ? \nThis is always very interesting to have an idea of the use of this kind of competition. \n\nThanks!", "votes": null }, { "id": "373339", "postDate": "08/21/2018 07:45:10", "content": "The goal of the Trackml challenge is to uncover new efficient and fast tracking algorithm. The first phase here on Kaggle had no CPU incentive, the goal being to uncover efficient algorithms, without worrying too much on the speed (except for praticality). The second phase to run on codalab in September will have a strong speed incentive. So it is much too early to conclude. \nStill to answer more directly your question: this dataset was made specially for this competition, in an unusual form (absolute coordinates rather than local one, more intuitive for newcomers but unusual for physicists ) we don't have a reference we can say this is the best we can do. However #3 and #4 are physicists with tracking expertise so they can be used as reference. The fact that we have different algorithms in the same ball park (say above 0.8) is already a success. Even people with lower score have had nice ideas as it appear from the forum.
", "rawMarkdown": "The goal of the Trackml challenge is to uncover new efficient and fast tracking algorithm. The first phase here on Kaggle had no CPU incentive, the goal being to uncover efficient algorithms, without worrying too much on the speed (except for praticality). The second phase to run on codalab in September will have a strong speed incentive. So it is much too early to conclude. \nStill to answer more directly your question: this dataset was made specially for this competition, in an unusual form (absolute coordinates rather than local one, more intuitive for newcomers but unusual for physicists ) we don't have a reference we can say this is the best we can do. However #3 and #4 are physicists with tracking expertise so they can be used as reference. The fact that we have different algorithms in the same ball park (say above 0.8) is already a success. Even people with lower score have had nice ideas as it appear from the forum.", "votes": null }, { "id": "373347", "postDate": "08/21/2018 07:55:35", "content": "Kind reminder, just one more week to provide the document and fill in the questionnaire. Please take the time to do it, it will only a fraction of the time you've invested so far and will increase the impact on the community significantly.
", "rawMarkdown": "Kind reminder, just one more week to provide the document and fill in the questionnaire. Please take the time to do it, it will only a fraction of the time you've invested so far and will increase the impact on the community significantly.", "votes": null }, { "id": "373833", "postDate": "08/22/2018 01:52:13", "content": "In order to prepare for the second phase, may I ask how can you measure the speed of a code? Can we use \"pre-trained\" models and \"pre-defined\" parameters and then just plug them in? Or the speed will be measured by only allowing the code implementing directly from the given input data? If your answer is the former case, then there would be thousands of ways to increase speed, including using rules, dictionaries... Is there any constraint?
", "rawMarkdown": "In order to prepare for the second phase, may I ask how can you measure the speed of a code? Can we use \"pre-trained\" models and \"pre-defined\" parameters and then just plug them in? Or the speed will be measured by only allowing the code implementing directly from the given input data? If your answer is the former case, then there would be thousands of ways to increase speed, including using rules, dictionaries... Is there any constraint?", "votes": null }, { "id": "373999", "postDate": "08/22/2018 08:37:09", "content": "We will take into account only the evaluation / track finding time. Participant train their model on their own resources, then upload the trained model to Codalab platform. They will need to plug their evaluation in a skeleton (we provide) taking care of the I/O and calling participant code event by event. Only the time spent in the participant code is measured.
", "rawMarkdown": "We will take into account only the evaluation / track finding time. Participant train their model on their own resources, then upload the trained model to Codalab platform. They will need to plug their evaluation in a skeleton (we provide) taking care of the I/O and calling participant code event by event. Only the time spent in the participant code is measured.", "votes": null }, { "id": "382839", "postDate": "09/07/2018 07:48:12", "content": "Please find attached the zip file with the 7 documents we have received, as well as pointer to the released software.\nWith contributions from #1 @icecuber and @erlinsol #2 @outrunner #3 @sgorbuno Sergey Gorbunov #7 @yuval6967 Yuval R and @trian2018 #9 @cpmpml #11 @andri27 Andrea Lonza #12 @nicolefinnie and @jliamfinnie\nEnjoy.
\n\n(In case you are wondering, we are in the final tests of the second Throughput phase on Codalab, to be opened in a few days)
", "rawMarkdown": "Please find attached the zip file with the 7 documents we have received, as well as pointer to the released software.\nWith contributions from #1 @icecuber and @erlinsol #2 @outrunner #3 @sgorbuno Sergey Gorbunov #7 @yuval6967 Yuval R and @trian2018 #9 @cpmpml #11 @andri27 Andrea Lonza #12 @nicolefinnie and @jliamfinnie\nEnjoy.\n\n(In case you are wondering, we are in the final tests of the second Throughput phase on Codalab, to be opened in a few days)", "votes": null }, { "id": "385944", "postDate": "09/11/2018 20:55:58", "content": "The second \"Throughput\" phase of the competition is online, see this post
", "rawMarkdown": "The second \"Throughput\" phase of the competition is online, see [this post][1]\n\n\n [1]: https://www.kaggle.com/c/trackml-particle-identification/discussion/65525", "votes": null } ], "comments": [ { "id": 370418, "author_name": "bolkonsky", "author_url": "", "post_date": "08/14/2018 19:34:22", "content": "Congratulations to winners and all other participants)
", "votes": null, "replies": [] }, { "id": 370586, "author_name": "cpmpml", "author_url": "", "post_date": "08/15/2018 04:35:18", "content": "Thanks for these detailed instruction.
\n\n\n\n\nwith correction of bug/features (most notable one being the Z spread of the origin of the particles which increases from 5.5mm to 55mm).
\n
The magnetic field variation impacts trajectories way more than what is described in the documents you shared. Indeed, it reads that departure from perfect helix should not exceed few millimeters. In reality departure from perfect helix can be way larger. I'm pointing this out to make sure you do not have a scaling issue for the magnetic field direction as well.
\n\n\n\n\nThe scoring will combine accuracy score and speed. Accuracy score is the same as for the first phase, except that only particles originating from a narrow cylinder around the z axis will be accounted for (so that we will focus on the « easy » particles).
\n
I wish this was the case for the first phase as clustering approaches would be way closer to the top score ! ;)
", "votes": null, "replies": [ { "id": 370693, "author_name": "edwinst", "author_url": "", "post_date": "08/15/2018 09:53:31", "content": "\n\n\nIndeed, it reads that departure from perfect helix should not exceed few millimeters.
\n
Judging by the intersection displacements learned by my algorithm (see the \"Pretty pictures\" post), this is mostly true if one considers only the systematic deviations due to the magnetic field. Only in the end cap regions do the deviations become of the order of 10 mm or so, if I remember correctly. The consideration of these systematic deviations therefore gave me only a small score improvement (~1%).
", "votes": null, "replies": [] }, { "id": 370825, "author_name": "cpmpml", "author_url": "", "post_date": "08/15/2018 14:02:08", "content": "My correction of the unrolling angle gave me 0.03 score improvement, at least.
", "votes": null, "replies": [] } ] }, { "id": 371144, "author_name": "droussea", "author_url": "", "post_date": "08/16/2018 05:15:27", "content": "So participants #1 #2 and #3, and all participants applying to HEP meets ML jury prize (don't be shy!), should *before Monday 27th August 11:59PM UTC:
\n\nWe are also welcoming the feedback of other participants by filling in the survey (without releaseing software or document).
\n\nQuestions by replying here, or using trackml.contact@gmail.com
", "votes": null, "replies": [ { "id": 371228, "author_name": "cpmpml", "author_url": "", "post_date": "08/16/2018 09:41:43", "content": "One suggestion if I may: if you use conda or anaconda, then you can export your environment to a yaml file that contains all the packages you use with their version, see https://conda.io/docs/user-guide/tasks/manage-environments.html#sharing-an-environment . You can then easily recreate the same environment from the file, see https://conda.io/docs/user-guide/tasks/manage-environments.html#creating-an-environment-from-an-environment-yml-file
\n\nThis is a standard way to describe the Python dependencies you have, and I would recommend you accept it as an alternative to requirements.txt .
", "votes": null, "replies": [] }, { "id": 371323, "author_name": "droussea", "author_url": "", "post_date": "08/16/2018 14:39:49", "content": "Agreed
", "votes": null, "replies": [] } ] }, { "id": 372825, "author_name": "equiplane", "author_url": "", "post_date": "08/20/2018 12:27:24", "content": "I do not know if this was posted before, but could we have an idea of the improvement on the score between organizers's initial solution and the final, winning one ? \nThis is always very interesting to have an idea of the use of this kind of competition.
\n\nThanks!
", "votes": null, "replies": [ { "id": 373339, "author_name": "droussea", "author_url": "", "post_date": "08/21/2018 07:45:10", "content": "The goal of the Trackml challenge is to uncover new efficient and fast tracking algorithm. The first phase here on Kaggle had no CPU incentive, the goal being to uncover efficient algorithms, without worrying too much on the speed (except for praticality). The second phase to run on codalab in September will have a strong speed incentive. So it is much too early to conclude. \nStill to answer more directly your question: this dataset was made specially for this competition, in an unusual form (absolute coordinates rather than local one, more intuitive for newcomers but unusual for physicists ) we don't have a reference we can say this is the best we can do. However #3 and #4 are physicists with tracking expertise so they can be used as reference. The fact that we have different algorithms in the same ball park (say above 0.8) is already a success. Even people with lower score have had nice ideas as it appear from the forum.
", "votes": null, "replies": [] } ] }, { "id": 373347, "author_name": "droussea", "author_url": "", "post_date": "08/21/2018 07:55:35", "content": "Kind reminder, just one more week to provide the document and fill in the questionnaire. Please take the time to do it, it will only a fraction of the time you've invested so far and will increase the impact on the community significantly.
", "votes": null, "replies": [ { "id": 373833, "author_name": "khahuras", "author_url": "", "post_date": "08/22/2018 01:52:13", "content": "In order to prepare for the second phase, may I ask how can you measure the speed of a code? Can we use \"pre-trained\" models and \"pre-defined\" parameters and then just plug them in? Or the speed will be measured by only allowing the code implementing directly from the given input data? If your answer is the former case, then there would be thousands of ways to increase speed, including using rules, dictionaries... Is there any constraint?
", "votes": null, "replies": [] }, { "id": 373999, "author_name": "droussea", "author_url": "", "post_date": "08/22/2018 08:37:09", "content": "We will take into account only the evaluation / track finding time. Participant train their model on their own resources, then upload the trained model to Codalab platform. They will need to plug their evaluation in a skeleton (we provide) taking care of the I/O and calling participant code event by event. Only the time spent in the participant code is measured.
", "votes": null, "replies": [] } ] }, { "id": 382839, "author_name": "droussea", "author_url": "", "post_date": "09/07/2018 07:48:12", "content": "Please find attached the zip file with the 7 documents we have received, as well as pointer to the released software.\nWith contributions from #1 @icecuber and @erlinsol #2 @outrunner #3 @sgorbuno Sergey Gorbunov #7 @yuval6967 Yuval R and @trian2018 #9 @cpmpml #11 @andri27 Andrea Lonza #12 @nicolefinnie and @jliamfinnie\nEnjoy.
\n\n(In case you are wondering, we are in the final tests of the second Throughput phase on Codalab, to be opened in a few days)
", "votes": null, "replies": [] }, { "id": 385944, "author_name": "droussea", "author_url": "", "post_date": "09/11/2018 20:55:58", "content": "The second \"Throughput\" phase of the competition is online, see this post
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "370224": "Congratulations to all participants of this amazing competition! As organizers it was fantastic to watch the quick evolution of the score and the forum interactions as well as the posted kernels. We are looking forward to learning more details on how you did it.\n\nThe (temporary) private leaderboard is now revealed and winners are: #1 Top Quarks (icecuber+ersol), #2 outrunner, #3 Sergey Gorbunov. Bravo!\n\nWe are happy to see that the private leaderboard is identical to the public one for ranks 1 to 19. There has been no luck factor!\n\nPlease check our last post on https://twitter.com/trackmllhc with some interesting plots (but all other relevant information is in this topic).\n\nWhat happens now:\n\n 1. Top scorers will be invited to submit their software (or better making it public, because organizers will do this anyway) with an open source license within 2 weeks, and to release some structured documentation on their software (precise indication how to do so in my reply below). This is mandatory to be able to claim the monetary prizes 12.000$ for #1, 8.000$ for #2, 5.000$ for #3. \n\n 2. Anyone (even not top scorers) who thinks they made valuable contributions with innovative algorithms (in particular if they have shared insights on the forum) will be welcome to release publicly their code with an open source license and lightweight structured documentation also within two weeks (instructions in my reply below). This will allow to compete for the « HEP meets ML » jury prizes: a NVidia V100 GPU, and two invitations to NIPS 2018 or the spring 2019 CERN workshop.\n\n2b : The survey indicated in my reply below can also be filled by other participants, as a mean for feedback to the competition\n\n 3. The second ''Throughput\" phase of the competition will start early September on Codalab platform (this is an official NIPS competition). The dataset will be very similar, with correction of bug/features (most notable one being the Z spread of the origin of the particles which increases from 5.5mm to 55mm). The scoring will combine accuracy score and speed. Accuracy score is the same as for the first phase, except that only particles originating from a narrow cylinder around the z axis will be accounted for (so that we will focus on the « easy » particles). Speed is measured on a single core CPU. Participation to the Throughput phase is open to anyone. Using software released by Accuracy phase participants is of course authorized.\nWinners of that phase will share 15k$ monetary prize, and there will be a second set of « HEP meets ML » jury prizes, another Nvidia V100 GPU, and 2 invitation to CERN or NIPS.\n\n 4. Meanwhile the international jury with both physicists and computer scientists will select by end September the most interesting submissions of the first phase to attribute the special \"HEP meet ML\" prizes.", "370418": "Congratulations to winners and all other participants)", "370586": "Thanks for these detailed instruction.\n\n> with correction of bug/features (most notable one being the Z spread of the origin of the particles which increases from 5.5mm to 55mm). \n\nThe magnetic field variation impacts trajectories way more than what is described in the documents you shared. Indeed, it reads that departure from perfect helix should not exceed few millimeters. In reality departure from perfect helix can be way larger. I'm pointing this out to make sure you do not have a scaling issue for the magnetic field direction as well.\n\n> The scoring will combine accuracy score and speed. Accuracy score is the same as for the first phase, except that only particles originating from a narrow cylinder around the z axis will be accounted for (so that we will focus on the « easy » particles). \n\nI wish this was the case for the first phase as clustering approaches would be way closer to the top score ! ;)", "370693": "> Indeed, it reads that departure from perfect helix should not exceed few millimeters.\n\nJudging by the intersection displacements learned by my algorithm (see the \"Pretty pictures\" post), this is mostly true if one considers only the *systematic* deviations due to the magnetic field. Only in the end cap regions do the deviations become of the order of 10 mm or so, if I remember correctly. The consideration of these systematic deviations therefore gave me only a small score improvement (~1%).", "370825": "My correction of the unrolling angle gave me 0.03 score improvement, at least.", "371144": "So participants #1 #2 and #3, and all participants applying to HEP meets ML jury prize (don't be shy!), should *before Monday 27th August 11:59PM UTC:\n\n 1. release their software with an open source license as indicated in section B of the attached document (github or the like is preferred)\n 2. write a short pdf document detailing their methods following template in section A of the attached document. (many of you have already posted detailed info on their method on the forum, which is good of course, but please take the time to reformat the information according to the template, this will simplify considerably the work of the jury)\n 3. post the two public links to the forum (under the topic you might have already created) or a new one\n 4. fill the survey https://goo.gl/forms/hADGfrOjKw2ws91G3 which will only be seen by Kaggle, the TrackML organizers and the jury\n\nWe are also welcoming the feedback of other participants by filling in the survey (without releaseing software or document).\n\nQuestions by replying here, or using trackml.contact@gmail.com", "371228": "One suggestion if I may: if you use conda or anaconda, then you can export your environment to a yaml file that contains all the packages you use with their version, see https://conda.io/docs/user-guide/tasks/manage-environments.html#sharing-an-environment . You can then easily recreate the same environment from the file, see https://conda.io/docs/user-guide/tasks/manage-environments.html#creating-an-environment-from-an-environment-yml-file\n\nThis is a standard way to describe the Python dependencies you have, and I would recommend you accept it as an alternative to requirements.txt .", "371323": "Agreed", "372825": "I do not know if this was posted before, but could we have an idea of the improvement on the score between organizers's initial solution and the final, winning one ? \nThis is always very interesting to have an idea of the use of this kind of competition. \n\nThanks!", "373339": "The goal of the Trackml challenge is to uncover new efficient and fast tracking algorithm. The first phase here on Kaggle had no CPU incentive, the goal being to uncover efficient algorithms, without worrying too much on the speed (except for praticality). The second phase to run on codalab in September will have a strong speed incentive. So it is much too early to conclude. \nStill to answer more directly your question: this dataset was made specially for this competition, in an unusual form (absolute coordinates rather than local one, more intuitive for newcomers but unusual for physicists ) we don't have a reference we can say this is the best we can do. However #3 and #4 are physicists with tracking expertise so they can be used as reference. The fact that we have different algorithms in the same ball park (say above 0.8) is already a success. Even people with lower score have had nice ideas as it appear from the forum.", "373347": "Kind reminder, just one more week to provide the document and fill in the questionnaire. Please take the time to do it, it will only a fraction of the time you've invested so far and will increase the impact on the community significantly.", "373833": "In order to prepare for the second phase, may I ask how can you measure the speed of a code? Can we use \"pre-trained\" models and \"pre-defined\" parameters and then just plug them in? Or the speed will be measured by only allowing the code implementing directly from the given input data? If your answer is the former case, then there would be thousands of ways to increase speed, including using rules, dictionaries... Is there any constraint?", "373999": "We will take into account only the evaluation / track finding time. Participant train their model on their own resources, then upload the trained model to Codalab platform. They will need to plug their evaluation in a skeleton (we provide) taking care of the I/O and calling participant code event by event. Only the time spent in the participant code is measured.", "382839": "Please find attached the zip file with the 7 documents we have received, as well as pointer to the released software.\nWith contributions from #1 @icecuber and @erlinsol #2 @outrunner #3 @sgorbuno Sergey Gorbunov #7 @yuval6967 Yuval R and @trian2018 #9 @cpmpml #11 @andri27 Andrea Lonza #12 @nicolefinnie and @jliamfinnie\nEnjoy.\n\n(In case you are wondering, we are in the final tests of the second Throughput phase on Codalab, to be opened in a few days)", "385944": "The second \"Throughput\" phase of the competition is online, see [this post][1]\n\n\n [1]: https://www.kaggle.com/c/trackml-particle-identification/discussion/65525" }, "source": "meta" }