{
  "id": 209548,
  "title": "Did you write a train loader to mimic to the test for your transformer-based models?",
  "url": "/competitions/riiid-test-answer-prediction/discussion/209548",
  "author_name": "",
  "post_date": "2021-01-07T21:29:47.815384200Z",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I tried to split the sequence using <code>timestamp</code> for a user:<br>\n<a href=\"https://www.kaggle.com/scaomath/riiid-sakt-fine-tuning-mimicking-the-test\" target=\"_blank\">https://www.kaggle.com/scaomath/riiid-sakt-fine-tuning-mimicking-the-test</a></p>\n<p>However, the improvement is only marginal. Those of you who are high on the LB, did you do this or use just a simple trainloader but encode the <code>timestamp</code> or <code>lag</code> info as embedding?</p>",
  "messages": [
    {
      "id": "1143365",
      "postDate": "01/07/2021 21:29:47",
      "content": "<p>I tried to split the sequence using <code>timestamp</code> for a user:<br>\n<a href=\"https://www.kaggle.com/scaomath/riiid-sakt-fine-tuning-mimicking-the-test\" target=\"_blank\">https://www.kaggle.com/scaomath/riiid-sakt-fine-tuning-mimicking-the-test</a></p>\n<p>However, the improvement is only marginal. Those of you who are high on the LB, did you do this or use just a simple trainloader but encode the <code>timestamp</code> or <code>lag</code> info as embedding?</p>",
      "rawMarkdown": "I tried to split the sequence using `timestamp` for a user:\nhttps://www.kaggle.com/scaomath/riiid-sakt-fine-tuning-mimicking-the-test\n\nHowever, the improvement is only marginal. Those of you who are high on the LB, did you do this or use just a simple trainloader but encode the `timestamp` or `lag` info as embedding?",
      "votes": null
    },
    {
      "id": "1143388",
      "postDate": "01/07/2021 21:58:08",
      "content": "<p>I actually write the validation dataloader to mimic the inference. I split based on the notebook of <a href=\"https://www.kaggle.com/yihdarshieh\" target=\"_blank\">@yihdarshieh</a>. I am gonna share the notebook in an hour if you have interest and time, you can check out :D. I got only 0.786 with a single SAINT, so a lot of things would be missing, I will appreciate if you could suggest some ideas to improve.</p>",
      "rawMarkdown": "I actually write the validation dataloader to mimic the inference. I split based on the notebook of @yihdarshieh. I am gonna share the notebook in an hour if you have interest and time, you can check out :D. I got only 0.786 with a single SAINT, so a lot of things would be missing, I will appreciate if you could suggest some ideas to improve.",
      "votes": null
    },
    {
      "id": "1143396",
      "postDate": "01/07/2021 22:08:07",
      "content": "<p>I also tried a prototype BERT-like model, however, due to the time constraint…I cannot finish. I suspect some top solutions are using BERT-like structures.</p>",
      "rawMarkdown": "I also tried a prototype BERT-like model, however, due to the time constraint...I cannot finish. I suspect some top solutions are using BERT-like structures.",
      "votes": null
    },
    {
      "id": "1143406",
      "postDate": "01/07/2021 22:22:25",
      "content": "<p>great idea, I am also looking forward to top-LB solutions.</p>",
      "rawMarkdown": "great idea, I am also looking forward to top-LB solutions.",
      "votes": null
    },
    {
      "id": "1143436",
      "postDate": "01/07/2021 22:46:05",
      "content": "<p>Im using a bert like model, but there are issues because I only use an encoder</p>",
      "rawMarkdown": "Im using a bert like model, but there are issues because I only use an encoder",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1143388,
      "author_name": "shinomoriaoshi",
      "author_url": "",
      "post_date": "01/07/2021 21:58:08",
      "content": "<p>I actually write the validation dataloader to mimic the inference. I split based on the notebook of <a href=\"https://www.kaggle.com/yihdarshieh\" target=\"_blank\">@yihdarshieh</a>. I am gonna share the notebook in an hour if you have interest and time, you can check out :D. I got only 0.786 with a single SAINT, so a lot of things would be missing, I will appreciate if you could suggest some ideas to improve.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1143396,
          "author_name": "scaomath",
          "author_url": "",
          "post_date": "01/07/2021 22:08:07",
          "content": "<p>I also tried a prototype BERT-like model, however, due to the time constraint…I cannot finish. I suspect some top solutions are using BERT-like structures.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1143406,
          "author_name": "shinomoriaoshi",
          "author_url": "",
          "post_date": "01/07/2021 22:22:25",
          "content": "<p>great idea, I am also looking forward to top-LB solutions.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1143436,
          "author_name": "shujun717",
          "author_url": "",
          "post_date": "01/07/2021 22:46:05",
          "content": "<p>Im using a bert like model, but there are issues because I only use an encoder</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1143365": "I tried to split the sequence using `timestamp` for a user:\nhttps://www.kaggle.com/scaomath/riiid-sakt-fine-tuning-mimicking-the-test\n\nHowever, the improvement is only marginal. Those of you who are high on the LB, did you do this or use just a simple trainloader but encode the `timestamp` or `lag` info as embedding?",
    "1143388": "I actually write the validation dataloader to mimic the inference. I split based on the notebook of @yihdarshieh. I am gonna share the notebook in an hour if you have interest and time, you can check out :D. I got only 0.786 with a single SAINT, so a lot of things would be missing, I will appreciate if you could suggest some ideas to improve.",
    "1143396": "I also tried a prototype BERT-like model, however, due to the time constraint...I cannot finish. I suspect some top solutions are using BERT-like structures.",
    "1143406": "great idea, I am also looking forward to top-LB solutions.",
    "1143436": "Im using a bert like model, but there are issues because I only use an encoder"
  },
  "source": "meta"
}