{
  "id": 58705,
  "title": "Open position in our team",
  "url": "/competitions/avito-demand-prediction/discussion/58705",
  "author_name": "",
  "post_date": "2018-06-12T17:40:48.320930300Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi, we still have room in our team.</p>\n\n<p>We have few good models (lgb 0.2200 LB, NN 0.2196 LB, xgb and catboost both around 0.2210) but lack feature engineering skills and computational power to train more models for ensemble. Additionally we did not put much effort in ensembling yet, just simple 1:1 averaging.  Thinking also of future competitions, I don't see the point in simply blending our models without learning a thing or two from each other.</p>\n\n<p>So, if someone experienced wants to join our team feel free to send a pm or post here.</p>",
  "messages": [
    {
      "id": "342008",
      "postDate": "06/12/2018 17:40:48",
      "content": "<p>Hi, we still have room in our team.</p>\n\n<p>We have few good models (lgb 0.2200 LB, NN 0.2196 LB, xgb and catboost both around 0.2210) but lack feature engineering skills and computational power to train more models for ensemble. Additionally we did not put much effort in ensembling yet, just simple 1:1 averaging.  Thinking also of future competitions, I don't see the point in simply blending our models without learning a thing or two from each other.</p>\n\n<p>So, if someone experienced wants to join our team feel free to send a pm or post here.</p>",
      "rawMarkdown": "Hi, we still have room in our team.\n\nWe have few good models (lgb 0.2200 LB, NN 0.2196 LB, xgb and catboost both around 0.2210) but lack feature engineering skills and computational power to train more models for ensemble. Additionally we did not put much effort in ensembling yet, just simple 1:1 averaging.  Thinking also of future competitions, I don't see the point in simply blending our models without learning a thing or two from each other.\n\nSo, if someone experienced wants to join our team feel free to send a pm or post here.",
      "votes": null
    },
    {
      "id": "342040",
      "postDate": "06/12/2018 18:34:02",
      "content": "<p>Are you using images?</p>",
      "rawMarkdown": "Are you using images?",
      "votes": null
    },
    {
      "id": "342045",
      "postDate": "06/12/2018 18:46:50",
      "content": "<p>Not the raw pixels if you mean that. We extracted some features from pretrained models and feed those as numerical feature</p>",
      "rawMarkdown": "Not the raw pixels if you mean that. We extracted some features from pretrained models and feed those as numerical feature",
      "votes": null
    },
    {
      "id": "342091",
      "postDate": "06/12/2018 20:12:59",
      "content": "<p>Thanks. Im working on a NN which combines image pixels plus other features (non-engineering so far). </p>\n\n<p>I haven't yet submitted anything but w/o using active/period csvs I'm getting 0.2222 local CV (just 1 fold, trained stopped prematurely b/c im still testing... 4-5 epochs). I want to see how far I get w/ a NN. I plan to spend the next couple of days adding active/period CSVs plus other engineered features. Training w/ pixels takes a lot of time even w/ GPU but all things being equal it improves CV (one epoch w/o pixels takes ~3-4 minutes and with pixels ~2-3 hours per epoch).</p>",
      "rawMarkdown": "Thanks. Im working on a NN which combines image pixels plus other features (non-engineering so far). \n\nI haven't yet submitted anything but w/o using active/period csvs I'm getting 0.2222 local CV (just 1 fold, trained stopped prematurely b/c im still testing... 4-5 epochs). I want to see how far I get w/ a NN. I plan to spend the next couple of days adding active/period CSVs plus other engineered features. Training w/ pixels takes a lot of time even w/ GPU but all things being equal it improves CV (one epoch w/o pixels takes ~3-4 minutes and with pixels ~2-3 hours per epoch).",
      "votes": null
    },
    {
      "id": "342181",
      "postDate": "06/13/2018 03:35:41",
      "content": "<p>I have sent a PM message please check.</p>",
      "rawMarkdown": "I have sent a PM message please check.",
      "votes": null
    },
    {
      "id": "342545",
      "postDate": "06/13/2018 17:17:55",
      "content": "<p>I'm doing the similar thing now, but it's so time-consuming, even I'm using 11G GTX 1080 ti</p>",
      "rawMarkdown": "I'm doing the similar thing now, but it's so time-consuming, even I'm using 11G GTX 1080 ti",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 342040,
      "author_name": "antorsae",
      "author_url": "",
      "post_date": "06/12/2018 18:34:02",
      "content": "<p>Are you using images?</p>",
      "votes": null,
      "replies": [
        {
          "id": 342045,
          "author_name": "christofhenkel",
          "author_url": "",
          "post_date": "06/12/2018 18:46:50",
          "content": "<p>Not the raw pixels if you mean that. We extracted some features from pretrained models and feed those as numerical feature</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 342091,
          "author_name": "antorsae",
          "author_url": "",
          "post_date": "06/12/2018 20:12:59",
          "content": "<p>Thanks. Im working on a NN which combines image pixels plus other features (non-engineering so far). </p>\n\n<p>I haven't yet submitted anything but w/o using active/period csvs I'm getting 0.2222 local CV (just 1 fold, trained stopped prematurely b/c im still testing... 4-5 epochs). I want to see how far I get w/ a NN. I plan to spend the next couple of days adding active/period CSVs plus other engineered features. Training w/ pixels takes a lot of time even w/ GPU but all things being equal it improves CV (one epoch w/o pixels takes ~3-4 minutes and with pixels ~2-3 hours per epoch).</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 342545,
          "author_name": "creatrol",
          "author_url": "",
          "post_date": "06/13/2018 17:17:55",
          "content": "<p>I'm doing the similar thing now, but it's so time-consuming, even I'm using 11G GTX 1080 ti</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 342181,
      "author_name": "adityakumarsinha",
      "author_url": "",
      "post_date": "06/13/2018 03:35:41",
      "content": "<p>I have sent a PM message please check.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "342008": "Hi, we still have room in our team.\n\nWe have few good models (lgb 0.2200 LB, NN 0.2196 LB, xgb and catboost both around 0.2210) but lack feature engineering skills and computational power to train more models for ensemble. Additionally we did not put much effort in ensembling yet, just simple 1:1 averaging.  Thinking also of future competitions, I don't see the point in simply blending our models without learning a thing or two from each other.\n\nSo, if someone experienced wants to join our team feel free to send a pm or post here.",
    "342040": "Are you using images?",
    "342045": "Not the raw pixels if you mean that. We extracted some features from pretrained models and feed those as numerical feature",
    "342091": "Thanks. Im working on a NN which combines image pixels plus other features (non-engineering so far). \n\nI haven't yet submitted anything but w/o using active/period csvs I'm getting 0.2222 local CV (just 1 fold, trained stopped prematurely b/c im still testing... 4-5 epochs). I want to see how far I get w/ a NN. I plan to spend the next couple of days adding active/period CSVs plus other engineered features. Training w/ pixels takes a lot of time even w/ GPU but all things being equal it improves CV (one epoch w/o pixels takes ~3-4 minutes and with pixels ~2-3 hours per epoch).",
    "342181": "I have sent a PM message please check.",
    "342545": "I'm doing the similar thing now, but it's so time-consuming, even I'm using 11G GTX 1080 ti"
  },
  "source": "meta"
}