{
  "id": 74877,
  "title": "Late coding",
  "url": "/competitions/PLAsTiCC-2018/discussion/74877",
  "author_name": "",
  "post_date": "2018-12-17T02:00:01.941478300Z",
  "votes": 14,
  "comment_count": 25,
  "views": 0,
  "content": "<p>This moment when you think you have found an idea, validated on CV, then you realize that computing test features will take about 6 days on a 20 core machine and there are 24 hours left...</p>\n\n<p>This is when optimizing code makes sense ;)</p>\n\n<p>This is also why I'm still awake at 3 am (:</p>\n\n<p>We'll see tomorrow if the code runs fast enough for a sub.  And if this is not another false hope.  </p>\n\n<p>This competition has been full of surprises, good and bad,.  Many proven practice have been detrimental here.  I am eager to see what the top teams have been doing that we are missing!</p>\n\n<p>Well, there is almost a day left, let's keep pushing!</p>",
  "messages": [
    {
      "id": "440072",
      "postDate": "12/17/2018 02:00:01",
      "content": "<p>This moment when you think you have found an idea, validated on CV, then you realize that computing test features will take about 6 days on a 20 core machine and there are 24 hours left...</p>\n\n<p>This is when optimizing code makes sense ;)</p>\n\n<p>This is also why I'm still awake at 3 am (:</p>\n\n<p>We'll see tomorrow if the code runs fast enough for a sub.  And if this is not another false hope.  </p>\n\n<p>This competition has been full of surprises, good and bad,.  Many proven practice have been detrimental here.  I am eager to see what the top teams have been doing that we are missing!</p>\n\n<p>Well, there is almost a day left, let's keep pushing!</p>",
      "rawMarkdown": "This moment when you think you have found an idea, validated on CV, then you realize that computing test features will take about 6 days on a 20 core machine and there are 24 hours left...\n\nThis is when optimizing code makes sense ;)\n\nThis is also why I'm still awake at 3 am (:\n\nWe'll see tomorrow if the code runs fast enough for a sub.  And if this is not another false hope.  \n\nThis competition has been full of surprises, good and bad,.  Many proven practice have been detrimental here.  I am eager to see what the top teams have been doing that we are missing!\n\nWell, there is almost a day left, let's keep pushing!",
      "votes": null
    },
    {
      "id": "440075",
      "postDate": "12/17/2018 02:06:17",
      "content": "<p>oh i am experiencing it now. i finally found something that improved my CV. but alas, it takes 1 hour to calculate just the train.</p>",
      "rawMarkdown": "oh i am experiencing it now. i finally found something that improved my CV. but alas, it takes 1 hour to calculate just the train.",
      "votes": null
    },
    {
      "id": "440078",
      "postDate": "12/17/2018 02:09:20",
      "content": "<p>This is what it takes to stay true to your best practices advice.</p>",
      "rawMarkdown": "This is what it takes to stay true to your best practices advice.",
      "votes": null
    },
    {
      "id": "440098",
      "postDate": "12/17/2018 03:09:45",
      "content": "<p>I can relate - though I don't have a 20 core machine.  Doing everything in the Kaggle kernels.</p>",
      "rawMarkdown": "I can relate - though I don't have a 20 core machine.  Doing everything in the Kaggle kernels.",
      "votes": null
    },
    {
      "id": "440102",
      "postDate": "12/17/2018 03:26:53",
      "content": "<p>Hi Sir, when you are done with this idea then share it separately how you optimized the code after competition is over</p>",
      "rawMarkdown": "Hi Sir, when you are done with this idea then share it separately how you optimized the code after competition is over",
      "votes": null
    },
    {
      "id": "440124",
      "postDate": "12/17/2018 04:54:35",
      "content": "<p>I experienced something similar. I initially tried to use libraries like dask which I read could handle large csv files, but it was still proving slow, my estimate said it would take me 150 days to compute test features. So I did what any sane person would do, which is to divide the single test file containing 3.5 million objects into 3.5 million small files and process them in parallel. That bought the computation time down to 2 days on a 16-core EC2 instance.</p>",
      "rawMarkdown": "I experienced something similar. I initially tried to use libraries like dask which I read could handle large csv files, but it was still proving slow, my estimate said it would take me 150 days to compute test features. So I did what any sane person would do, which is to divide the single test file containing 3.5 million objects into 3.5 million small files and process them in parallel. That bought the computation time down to 2 days on a 16-core EC2 instance.",
      "votes": null
    },
    {
      "id": "440134",
      "postDate": "12/17/2018 05:38:38",
      "content": "<p>Different on score, but it's the same feeling that we shared.\nBon courage!\n-- Best wishes from a guy coding at 2 am</p>",
      "rawMarkdown": "Different on score, but it's the same feeling that we shared.\nBon courage!\n-- Best wishes from a guy coding at 2 am",
      "votes": null
    },
    {
      "id": "440167",
      "postDate": "12/17/2018 07:07:36",
      "content": "<p>There is a thread about it, see <a href=\"https://www.kaggle.com/c/PLAsTiCC-2018/discussion/71398#latest-430428\">https://www.kaggle.com/c/PLAsTiCC-2018/discussion/71398#latest-430428</a>  But the most effective in my case is to replace slowest pandas piece by numpy base code.  I'll be more specific tomorrow ;)</p>",
      "rawMarkdown": "There is a thread about it, see https://www.kaggle.com/c/PLAsTiCC-2018/discussion/71398#latest-430428  But the most effective in my case is to replace slowest pandas piece by numpy base code.  I'll be more specific tomorrow ;)",
      "votes": null
    },
    {
      "id": "440168",
      "postDate": "12/17/2018 07:07:57",
      "content": "<p>Is there a limit on the number of kernels you can run?</p>",
      "rawMarkdown": "Is there a limit on the number of kernels you can run?",
      "votes": null
    },
    {
      "id": "440170",
      "postDate": "12/17/2018 07:09:58",
      "content": "<p>Thanks !</p>",
      "rawMarkdown": "Thanks !",
      "votes": null
    },
    {
      "id": "440178",
      "postDate": "12/17/2018 07:14:32",
      "content": "<p>I guess no because I am currently running 8 kernels at a time. If there is limit then it might be more than 10. kernels are saviour for guys like me who have poor computational machines:)</p>",
      "rawMarkdown": "I guess no because I am currently running 8 kernels at a time. If there is limit then it might be more than 10. kernels are saviour for guys like me who have poor computational machines:)",
      "votes": null
    },
    {
      "id": "440542",
      "postDate": "12/17/2018 17:17:53",
      "content": "<p>For those who desperately wanted to know the end of the story, the code ran fast enough, but it is not useful for improving LB score...</p>",
      "rawMarkdown": "For those who desperately wanted to know the end of the story, the code ran fast enough, but it is not useful for improving LB score...",
      "votes": null
    },
    {
      "id": "440546",
      "postDate": "12/17/2018 17:24:43",
      "content": "<p>I ran up to 20 at a time without issue.  </p>",
      "rawMarkdown": "I ran up to 20 at a time without issue.",
      "votes": null
    },
    {
      "id": "440548",
      "postDate": "12/17/2018 17:26:31",
      "content": "<p>@cpmp,\nSir, you can run upto 25 kernals at a time with CPU and \n15 with GPU</p>",
      "rawMarkdown": "cpmp,\nSir, you can run upto 25 kernals at a time with CPU and \n15 with GPU",
      "votes": null
    },
    {
      "id": "440550",
      "postDate": "12/17/2018 17:28:57",
      "content": "<p>Don't fell bad. Today I improved my CV from 0.44 to 0.34 and LB went from 0.85 to 1.02  :-P</p>",
      "rawMarkdown": "Don't fell bad. Today I improved my CV from 0.44 to 0.34 and LB went from 0.85 to 1.02  :-P",
      "votes": null
    },
    {
      "id": "440552",
      "postDate": "12/17/2018 17:29:49",
      "content": "<p>FYI,\ni have run 25 kernals at a time, in fact currently running ;)</p>",
      "rawMarkdown": "FYI,\ni have run 25 kernals at a time, in fact currently running ;)",
      "votes": null
    },
    {
      "id": "440556",
      "postDate": "12/17/2018 17:34:01",
      "content": "<p>if it takes 1 hrs on train then its possible to compute all the features on test in about 5 -10 hrs ,\nbreak the test file into small pices of pickle file (i did 90 pices) and run in parallel on all the kernals you can use( you can use upto 25 kernals)</p>",
      "rawMarkdown": "if it takes 1 hrs on train then its possible to compute all the features on test in about 5 -10 hrs ,\nbreak the test file into small pices of pickle file (i did 90 pices) and run in parallel on all the kernals you can use( you can use upto 25 kernals)",
      "votes": null
    },
    {
      "id": "440576",
      "postDate": "12/17/2018 18:12:12",
      "content": "<p>Did you share screen with me? ;)</p>",
      "rawMarkdown": "Did you share screen with me? ;)",
      "votes": null
    },
    {
      "id": "440791",
      "postDate": "12/18/2018 01:23:10",
      "content": "<p>When you find your good feature, calculate it in one hour... get to gold,  and then get out just &lt;0.001 from gold on private... after all this hard work while holding 2 moths-old baby on hands... writing it and crying...  this is life... </p>",
      "rawMarkdown": "When you find your good feature, calculate it in one hour... get to gold,  and then get out just &lt;0.001 from gold on private... after all this hard work while holding 2 moths-old baby on hands... writing it and crying...  this is life...",
      "votes": null
    },
    {
      "id": "440918",
      "postDate": "12/18/2018 05:03:21",
      "content": "<p>Blonde, little bit of bad luck to miss it by such small margin.  Keep up the spirits. You'll get plenty of golds in future competitions. You are an amazingly fast learner. </p>",
      "rawMarkdown": "Blonde, little bit of bad luck to miss it by such small margin.  Keep up the spirits. You'll get plenty of golds in future competitions. You are an amazingly fast learner.",
      "votes": null
    },
    {
      "id": "440989",
      "postDate": "12/18/2018 06:39:20",
      "content": "<p>@blondika This is a very impressive performance still, you should be very proud of it.</p>",
      "rawMarkdown": "blondika This is a very impressive performance still, you should be very proud of it.",
      "votes": null
    },
    {
      "id": "441108",
      "postDate": "12/18/2018 09:16:43",
      "content": "<p>Very Inspiring and I am waiting to look at your strategy.</p>",
      "rawMarkdown": "Very Inspiring and I am waiting to look at your strategy.",
      "votes": null
    },
    {
      "id": "441177",
      "postDate": "12/18/2018 11:25:24",
      "content": "<p>Some times, tiny last minute changes, makes the difference. The two last submission we submitted 2h to the deadline are the ones that gave us the 3rd place on the Private Leader board. \nAnd as @CPMP summarized it \"let's keep pushing\"  </p>",
      "rawMarkdown": "Some times, tiny last minute changes, makes the difference. The two last submission we submitted 2h to the deadline are the ones that gave us the 3rd place on the Private Leader board. \nAnd as @CPMP summarized it \"let's keep pushing\"",
      "votes": null
    },
    {
      "id": "441193",
      "postDate": "12/18/2018 11:45:47",
      "content": "<p>Congratulations, looking forward to your solutions</p>",
      "rawMarkdown": "Congratulations, looking forward to your solutions",
      "votes": null
    },
    {
      "id": "441207",
      "postDate": "12/18/2018 12:07:32",
      "content": "<p>Same to us, last two submits during last 40 min of competitions were the best... well, not place 3, but 13... last good feature calculated for test around 6 h before the end. That was the day! I did not manage even to get food, lol</p>",
      "rawMarkdown": "Same to us, last two submits during last 40 min of competitions were the best... well, not place 3, but 13... last good feature calculated for test around 6 h before the end. That was the day! I did not manage even to get food, lol",
      "votes": null
    },
    {
      "id": "441290",
      "postDate": "12/18/2018 13:55:26",
      "content": "<p>We will post a short description of our solution later today</p>",
      "rawMarkdown": "We will post a short description of our solution later today",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 440075,
      "author_name": "niclasdoce",
      "author_url": "",
      "post_date": "12/17/2018 02:06:17",
      "content": "<p>oh i am experiencing it now. i finally found something that improved my CV. but alas, it takes 1 hour to calculate just the train.</p>",
      "votes": null,
      "replies": [
        {
          "id": 440556,
          "author_name": "mks2192",
          "author_url": "",
          "post_date": "12/17/2018 17:34:01",
          "content": "<p>if it takes 1 hrs on train then its possible to compute all the features on test in about 5 -10 hrs ,\nbreak the test file into small pices of pickle file (i did 90 pices) and run in parallel on all the kernals you can use( you can use upto 25 kernals)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 440078,
      "author_name": "iprapas",
      "author_url": "",
      "post_date": "12/17/2018 02:09:20",
      "content": "<p>This is what it takes to stay true to your best practices advice.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 440098,
      "author_name": "jimpsull",
      "author_url": "",
      "post_date": "12/17/2018 03:09:45",
      "content": "<p>I can relate - though I don't have a 20 core machine.  Doing everything in the Kaggle kernels.</p>",
      "votes": null,
      "replies": [
        {
          "id": 440168,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "12/17/2018 07:07:57",
          "content": "<p>Is there a limit on the number of kernels you can run?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440178,
          "author_name": "aashish7936",
          "author_url": "",
          "post_date": "12/17/2018 07:14:32",
          "content": "<p>I guess no because I am currently running 8 kernels at a time. If there is limit then it might be more than 10. kernels are saviour for guys like me who have poor computational machines:)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440546,
          "author_name": "jimpsull",
          "author_url": "",
          "post_date": "12/17/2018 17:24:43",
          "content": "<p>I ran up to 20 at a time without issue.  </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440548,
          "author_name": "mks2192",
          "author_url": "",
          "post_date": "12/17/2018 17:26:31",
          "content": "<p>@cpmp,\nSir, you can run upto 25 kernals at a time with CPU and \n15 with GPU</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440552,
          "author_name": "mks2192",
          "author_url": "",
          "post_date": "12/17/2018 17:29:49",
          "content": "<p>FYI,\ni have run 25 kernals at a time, in fact currently running ;)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 440102,
      "author_name": "mks2192",
      "author_url": "",
      "post_date": "12/17/2018 03:26:53",
      "content": "<p>Hi Sir, when you are done with this idea then share it separately how you optimized the code after competition is over</p>",
      "votes": null,
      "replies": [
        {
          "id": 440167,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "12/17/2018 07:07:36",
          "content": "<p>There is a thread about it, see <a href=\"https://www.kaggle.com/c/PLAsTiCC-2018/discussion/71398#latest-430428\">https://www.kaggle.com/c/PLAsTiCC-2018/discussion/71398#latest-430428</a>  But the most effective in my case is to replace slowest pandas piece by numpy base code.  I'll be more specific tomorrow ;)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440170,
          "author_name": "mks2192",
          "author_url": "",
          "post_date": "12/17/2018 07:09:58",
          "content": "<p>Thanks !</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 440124,
      "author_name": "pankajb64",
      "author_url": "",
      "post_date": "12/17/2018 04:54:35",
      "content": "<p>I experienced something similar. I initially tried to use libraries like dask which I read could handle large csv files, but it was still proving slow, my estimate said it would take me 150 days to compute test features. So I did what any sane person would do, which is to divide the single test file containing 3.5 million objects into 3.5 million small files and process them in parallel. That bought the computation time down to 2 days on a 16-core EC2 instance.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 440134,
      "author_name": "alexpartisan",
      "author_url": "",
      "post_date": "12/17/2018 05:38:38",
      "content": "<p>Different on score, but it's the same feeling that we shared.\nBon courage!\n-- Best wishes from a guy coding at 2 am</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 440542,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "12/17/2018 17:17:53",
      "content": "<p>For those who desperately wanted to know the end of the story, the code ran fast enough, but it is not useful for improving LB score...</p>",
      "votes": null,
      "replies": [
        {
          "id": 440550,
          "author_name": "titericz",
          "author_url": "",
          "post_date": "12/17/2018 17:28:57",
          "content": "<p>Don't fell bad. Today I improved my CV from 0.44 to 0.34 and LB went from 0.85 to 1.02  :-P</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440576,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "12/17/2018 18:12:12",
          "content": "<p>Did you share screen with me? ;)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 440791,
      "author_name": "blondinka",
      "author_url": "",
      "post_date": "12/18/2018 01:23:10",
      "content": "<p>When you find your good feature, calculate it in one hour... get to gold,  and then get out just &lt;0.001 from gold on private... after all this hard work while holding 2 moths-old baby on hands... writing it and crying...  this is life... </p>",
      "votes": null,
      "replies": [
        {
          "id": 440918,
          "author_name": "subrahmanyamv",
          "author_url": "",
          "post_date": "12/18/2018 05:03:21",
          "content": "<p>Blonde, little bit of bad luck to miss it by such small margin.  Keep up the spirits. You'll get plenty of golds in future competitions. You are an amazingly fast learner. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 440989,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "12/18/2018 06:39:20",
          "content": "<p>@blondika This is a very impressive performance still, you should be very proud of it.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 441108,
      "author_name": "karthik7395",
      "author_url": "",
      "post_date": "12/18/2018 09:16:43",
      "content": "<p>Very Inspiring and I am waiting to look at your strategy.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 441177,
      "author_name": "yuval6967",
      "author_url": "",
      "post_date": "12/18/2018 11:25:24",
      "content": "<p>Some times, tiny last minute changes, makes the difference. The two last submission we submitted 2h to the deadline are the ones that gave us the 3rd place on the Private Leader board. \nAnd as @CPMP summarized it \"let's keep pushing\"  </p>",
      "votes": null,
      "replies": [
        {
          "id": 441193,
          "author_name": "mks2192",
          "author_url": "",
          "post_date": "12/18/2018 11:45:47",
          "content": "<p>Congratulations, looking forward to your solutions</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 441207,
          "author_name": "blondinka",
          "author_url": "",
          "post_date": "12/18/2018 12:07:32",
          "content": "<p>Same to us, last two submits during last 40 min of competitions were the best... well, not place 3, but 13... last good feature calculated for test around 6 h before the end. That was the day! I did not manage even to get food, lol</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 441290,
          "author_name": "yuval6967",
          "author_url": "",
          "post_date": "12/18/2018 13:55:26",
          "content": "<p>We will post a short description of our solution later today</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "440072": "This moment when you think you have found an idea, validated on CV, then you realize that computing test features will take about 6 days on a 20 core machine and there are 24 hours left...\n\nThis is when optimizing code makes sense ;)\n\nThis is also why I'm still awake at 3 am (:\n\nWe'll see tomorrow if the code runs fast enough for a sub.  And if this is not another false hope.  \n\nThis competition has been full of surprises, good and bad,.  Many proven practice have been detrimental here.  I am eager to see what the top teams have been doing that we are missing!\n\nWell, there is almost a day left, let's keep pushing!",
    "440075": "oh i am experiencing it now. i finally found something that improved my CV. but alas, it takes 1 hour to calculate just the train.",
    "440078": "This is what it takes to stay true to your best practices advice.",
    "440098": "I can relate - though I don't have a 20 core machine.  Doing everything in the Kaggle kernels.",
    "440102": "Hi Sir, when you are done with this idea then share it separately how you optimized the code after competition is over",
    "440124": "I experienced something similar. I initially tried to use libraries like dask which I read could handle large csv files, but it was still proving slow, my estimate said it would take me 150 days to compute test features. So I did what any sane person would do, which is to divide the single test file containing 3.5 million objects into 3.5 million small files and process them in parallel. That bought the computation time down to 2 days on a 16-core EC2 instance.",
    "440134": "Different on score, but it's the same feeling that we shared.\nBon courage!\n-- Best wishes from a guy coding at 2 am",
    "440167": "There is a thread about it, see https://www.kaggle.com/c/PLAsTiCC-2018/discussion/71398#latest-430428  But the most effective in my case is to replace slowest pandas piece by numpy base code.  I'll be more specific tomorrow ;)",
    "440168": "Is there a limit on the number of kernels you can run?",
    "440170": "Thanks !",
    "440178": "I guess no because I am currently running 8 kernels at a time. If there is limit then it might be more than 10. kernels are saviour for guys like me who have poor computational machines:)",
    "440542": "For those who desperately wanted to know the end of the story, the code ran fast enough, but it is not useful for improving LB score...",
    "440546": "I ran up to 20 at a time without issue.",
    "440548": "cpmp,\nSir, you can run upto 25 kernals at a time with CPU and \n15 with GPU",
    "440550": "Don't fell bad. Today I improved my CV from 0.44 to 0.34 and LB went from 0.85 to 1.02  :-P",
    "440552": "FYI,\ni have run 25 kernals at a time, in fact currently running ;)",
    "440556": "if it takes 1 hrs on train then its possible to compute all the features on test in about 5 -10 hrs ,\nbreak the test file into small pices of pickle file (i did 90 pices) and run in parallel on all the kernals you can use( you can use upto 25 kernals)",
    "440576": "Did you share screen with me? ;)",
    "440791": "When you find your good feature, calculate it in one hour... get to gold,  and then get out just &lt;0.001 from gold on private... after all this hard work while holding 2 moths-old baby on hands... writing it and crying...  this is life...",
    "440918": "Blonde, little bit of bad luck to miss it by such small margin.  Keep up the spirits. You'll get plenty of golds in future competitions. You are an amazingly fast learner.",
    "440989": "blondika This is a very impressive performance still, you should be very proud of it.",
    "441108": "Very Inspiring and I am waiting to look at your strategy.",
    "441177": "Some times, tiny last minute changes, makes the difference. The two last submission we submitted 2h to the deadline are the ones that gave us the 3rd place on the Private Leader board. \nAnd as @CPMP summarized it \"let's keep pushing\"",
    "441193": "Congratulations, looking forward to your solutions",
    "441207": "Same to us, last two submits during last 40 min of competitions were the best... well, not place 3, but 13... last good feature calculated for test around 6 h before the end. That was the day! I did not manage even to get food, lol",
    "441290": "We will post a short description of our solution later today"
  },
  "source": "meta"
}