{
  "id": 21792,
  "title": "What tools/setup do people with double digit submission use?",
  "url": "/competitions/facebook-v-predicting-check-ins/discussion/21792",
  "author_name": "",
  "post_date": "2016-06-19T23:45:14.910Z",
  "votes": null,
  "comment_count": 6,
  "views": 719,
  "content": "<p>This is my first Kaggle. Not trying to steal anyone's competitive advantage, I just want to get an idea of the state of the art.  </p>\n\n<p>I'm running xgboost on EC2 instances wherein each submission takes a few days. When I look at the leader board, it appears that some people can generate a submission within a few hours. Can someone please share a peak of their tools/setup? Spark?  Access to high performance computing infrastructure (via job, academia..)? </p>",
  "messages": [
    {
      "id": "124564",
      "postDate": "06/19/2016 23:45:14",
      "content": "<p>This is my first Kaggle. Not trying to steal anyone's competitive advantage, I just want to get an idea of the state of the art.  </p>\n\n<p>I'm running xgboost on EC2 instances wherein each submission takes a few days. When I look at the leader board, it appears that some people can generate a submission within a few hours. Can someone please share a peak of their tools/setup? Spark?  Access to high performance computing infrastructure (via job, academia..)? </p>",
      "rawMarkdown": "This is my first Kaggle. Not trying to steal anyone's competitive advantage, I just want to get an idea of the state of the art.  \r\n\r\nI'm running xgboost on EC2 instances wherein each submission takes a few days. When I look at the leader board, it appears that some people can generate a submission within a few hours. Can someone please share a peak of their tools/setup? Spark?  Access to high performance computing infrastructure (via job, academia..)?",
      "votes": null
    },
    {
      "id": "124566",
      "postDate": "06/20/2016 00:39:59",
      "content": "<p>Speed could happen in many ways:</p>\n\n<p>Fast algorithms (KNN, QDA, LR...)\nProblem shrinkage (grids, keep most common classes...)\nGeneral coding efficiency (initializing, vector operations, ...)</p>\n\n<p>You don't necessarily need a high-end computer or cluster setup to solve these problems but yes they can speed things up.</p>\n\n<p>For me, focusing on the first three, and borrowing some coding structure from the scripts, helped a lot. </p>",
      "rawMarkdown": "Speed could happen in many ways:\r\n\r\nFast algorithms (KNN, QDA, LR...)\r\nProblem shrinkage (grids, keep most common classes...)\r\nGeneral coding efficiency (initializing, vector operations, ...)\r\n\r\nYou don't necessarily need a high-end computer or cluster setup to solve these problems but yes they can speed things up.\r\n\r\nFor me, focusing on the first three, and borrowing some coding structure from the scripts, helped a lot.",
      "votes": null
    },
    {
      "id": "124579",
      "postDate": "06/20/2016 04:14:49",
      "content": "<p>You can submit on a subset of the dataset (just put all other row_ids as a place that isn't common in the test set). Furthermore I think there are some tricks that let you quickly do well on this dataset, some of the people at 0.6+ claimed their model only took 40 minutes to run. </p>",
      "rawMarkdown": "You can submit on a subset of the dataset (just put all other row_ids as a place that isn't common in the test set). Furthermore I think there are some tricks that let you quickly do well on this dataset, some of the people at 0.6+ claimed their model only took 40 minutes to run.",
      "votes": null
    },
    {
      "id": "124609",
      "postDate": "06/20/2016 11:24:45",
      "content": "<p>I do everything on my personal laptop. As Jared says, you can always submit on a subset of the dataset for testing purposes. Not to mention that the KNN script runs in under 20 minutes. In this competition, there's zero advantage for using fewer submissions so it's not a bad idea to test ideas with methods that can be run in a short amount of time and also translate to complete runs. </p>",
      "rawMarkdown": "I do everything on my personal laptop. As Jared says, you can always submit on a subset of the dataset for testing purposes. Not to mention that the KNN script runs in under 20 minutes. In this competition, there's zero advantage for using fewer submissions so it's not a bad idea to test ideas with methods that can be run in a short amount of time and also translate to complete runs.",
      "votes": null
    },
    {
      "id": "124676",
      "postDate": "06/21/2016 04:36:31",
      "content": "<p>[quote=Qingchen;124609]</p>\n\n<p>I do everything on my personal laptop. As Jared says, you can always submit on a subset of the dataset for testing purposes. Not to mention that the KNN script runs in under 20 minutes. In this competition, there's zero advantage for using fewer submissions so it's not a bad idea to test ideas with methods that can be run in a short amount of time and also translate to complete runs. </p>\n\n<p>[/quote]</p>\n\n<p>Oh, I didn't know we can submit a subset of test.csv. Clearly this is a good way to quickly test an idea. Will keep this in mind for next competitions. Thanks.</p>",
      "rawMarkdown": "[quote=Qingchen;124609]\r\n\r\nI do everything on my personal laptop. As Jared says, you can always submit on a subset of the dataset for testing purposes. Not to mention that the KNN script runs in under 20 minutes. In this competition, there's zero advantage for using fewer submissions so it's not a bad idea to test ideas with methods that can be run in a short amount of time and also translate to complete runs. \r\n\r\n[/quote]\r\n\r\nOh, I didn't know we can submit a subset of test.csv. Clearly this is a good way to quickly test an idea. Will keep this in mind for next competitions. Thanks.",
      "votes": null
    },
    {
      "id": "124709",
      "postDate": "06/21/2016 14:03:58",
      "content": "<p>[quote=Khanh;124676]</p>\n\n<p>Oh, I didn't know we can submit a subset of test.csv. Clearly this is a good way to quickly test an idea. Will keep this in mind for next competitions. Thanks.</p>\n\n<p>[/quote]</p>\n\n<p>You can't submit exactly a subset, but you can submit the sample submission and only change the rows for the subset that you predicted.</p>",
      "rawMarkdown": "[quote=Khanh;124676]\r\n\r\nOh, I didn't know we can submit a subset of test.csv. Clearly this is a good way to quickly test an idea. Will keep this in mind for next competitions. Thanks.\r\n\r\n[/quote]\r\n\r\nYou can't submit exactly a subset, but you can submit the sample submission and only change the rows for the subset that you predicted.",
      "votes": null
    },
    {
      "id": "124712",
      "postDate": "06/21/2016 14:54:13",
      "content": "<p>I am also new this and had the same problem with my first run taking over 24 hours. I am now down to less than an hour. </p>\n\n<p>Main way to speed up was to vectorize everything in pandas. I thought this was not possible in some cases but that was just because I did not understand pandas well enough. If you have any loops then you are doing it wrong!</p>",
      "rawMarkdown": "I am also new this and had the same problem with my first run taking over 24 hours. I am now down to less than an hour. \r\n\r\nMain way to speed up was to vectorize everything in pandas. I thought this was not possible in some cases but that was just because I did not understand pandas well enough. If you have any loops then you are doing it wrong!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 124566,
      "author_name": "fireiceearth",
      "author_url": "",
      "post_date": "06/20/2016 00:39:59",
      "content": "<p>Speed could happen in many ways:</p>\n\n<p>Fast algorithms (KNN, QDA, LR...)\nProblem shrinkage (grids, keep most common classes...)\nGeneral coding efficiency (initializing, vector operations, ...)</p>\n\n<p>You don't necessarily need a high-end computer or cluster setup to solve these problems but yes they can speed things up.</p>\n\n<p>For me, focusing on the first three, and borrowing some coding structure from the scripts, helped a lot. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 124579,
      "author_name": "jturkewitz",
      "author_url": "",
      "post_date": "06/20/2016 04:14:49",
      "content": "<p>You can submit on a subset of the dataset (just put all other row_ids as a place that isn't common in the test set). Furthermore I think there are some tricks that let you quickly do well on this dataset, some of the people at 0.6+ claimed their model only took 40 minutes to run. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 124609,
      "author_name": "qwang88",
      "author_url": "",
      "post_date": "06/20/2016 11:24:45",
      "content": "<p>I do everything on my personal laptop. As Jared says, you can always submit on a subset of the dataset for testing purposes. Not to mention that the KNN script runs in under 20 minutes. In this competition, there's zero advantage for using fewer submissions so it's not a bad idea to test ideas with methods that can be run in a short amount of time and also translate to complete runs. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 124676,
      "author_name": "knguyen",
      "author_url": "",
      "post_date": "06/21/2016 04:36:31",
      "content": "<p>[quote=Qingchen;124609]</p>\n\n<p>I do everything on my personal laptop. As Jared says, you can always submit on a subset of the dataset for testing purposes. Not to mention that the KNN script runs in under 20 minutes. In this competition, there's zero advantage for using fewer submissions so it's not a bad idea to test ideas with methods that can be run in a short amount of time and also translate to complete runs. </p>\n\n<p>[/quote]</p>\n\n<p>Oh, I didn't know we can submit a subset of test.csv. Clearly this is a good way to quickly test an idea. Will keep this in mind for next competitions. Thanks.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 124709,
      "author_name": "qwang88",
      "author_url": "",
      "post_date": "06/21/2016 14:03:58",
      "content": "<p>[quote=Khanh;124676]</p>\n\n<p>Oh, I didn't know we can submit a subset of test.csv. Clearly this is a good way to quickly test an idea. Will keep this in mind for next competitions. Thanks.</p>\n\n<p>[/quote]</p>\n\n<p>You can't submit exactly a subset, but you can submit the sample submission and only change the rows for the subset that you predicted.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 124712,
      "author_name": "simonm3",
      "author_url": "",
      "post_date": "06/21/2016 14:54:13",
      "content": "<p>I am also new this and had the same problem with my first run taking over 24 hours. I am now down to less than an hour. </p>\n\n<p>Main way to speed up was to vectorize everything in pandas. I thought this was not possible in some cases but that was just because I did not understand pandas well enough. If you have any loops then you are doing it wrong!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "124564": "This is my first Kaggle. Not trying to steal anyone's competitive advantage, I just want to get an idea of the state of the art.  \r\n\r\nI'm running xgboost on EC2 instances wherein each submission takes a few days. When I look at the leader board, it appears that some people can generate a submission within a few hours. Can someone please share a peak of their tools/setup? Spark?  Access to high performance computing infrastructure (via job, academia..)?",
    "124566": "Speed could happen in many ways:\r\n\r\nFast algorithms (KNN, QDA, LR...)\r\nProblem shrinkage (grids, keep most common classes...)\r\nGeneral coding efficiency (initializing, vector operations, ...)\r\n\r\nYou don't necessarily need a high-end computer or cluster setup to solve these problems but yes they can speed things up.\r\n\r\nFor me, focusing on the first three, and borrowing some coding structure from the scripts, helped a lot.",
    "124579": "You can submit on a subset of the dataset (just put all other row_ids as a place that isn't common in the test set). Furthermore I think there are some tricks that let you quickly do well on this dataset, some of the people at 0.6+ claimed their model only took 40 minutes to run.",
    "124609": "I do everything on my personal laptop. As Jared says, you can always submit on a subset of the dataset for testing purposes. Not to mention that the KNN script runs in under 20 minutes. In this competition, there's zero advantage for using fewer submissions so it's not a bad idea to test ideas with methods that can be run in a short amount of time and also translate to complete runs.",
    "124676": "[quote=Qingchen;124609]\r\n\r\nI do everything on my personal laptop. As Jared says, you can always submit on a subset of the dataset for testing purposes. Not to mention that the KNN script runs in under 20 minutes. In this competition, there's zero advantage for using fewer submissions so it's not a bad idea to test ideas with methods that can be run in a short amount of time and also translate to complete runs. \r\n\r\n[/quote]\r\n\r\nOh, I didn't know we can submit a subset of test.csv. Clearly this is a good way to quickly test an idea. Will keep this in mind for next competitions. Thanks.",
    "124709": "[quote=Khanh;124676]\r\n\r\nOh, I didn't know we can submit a subset of test.csv. Clearly this is a good way to quickly test an idea. Will keep this in mind for next competitions. Thanks.\r\n\r\n[/quote]\r\n\r\nYou can't submit exactly a subset, but you can submit the sample submission and only change the rows for the subset that you predicted.",
    "124712": "I am also new this and had the same problem with my first run taking over 24 hours. I am now down to less than an hour. \r\n\r\nMain way to speed up was to vectorize everything in pandas. I thought this was not possible in some cases but that was just because I did not understand pandas well enough. If you have any loops then you are doing it wrong!"
  },
  "source": "meta"
}