{
  "id": 114978,
  "title": "Rules Question",
  "url": "/competitions/Kannada-MNIST/discussion/114978",
  "author_name": "",
  "post_date": "2019-10-30T13:21:53.113212500Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi, I currently have a model that have a score of 0.99 and have  good place on the leader board. When I submit a solution the model will run again and produce slightly different results every time becuase of the random varibles in the model. </p>\n\n<p>What I wonder is it allowed to resubmit the same solution multiple times without new commits to get a good training run?.</p>\n\n<p>I think this is better than just change a few characters and waste 2 hours of GPU time on commiting a equivalent solution!</p>",
  "messages": [
    {
      "id": "661570",
      "postDate": "10/30/2019 13:21:53",
      "content": "<p>Hi, I currently have a model that have a score of 0.99 and have  good place on the leader board. When I submit a solution the model will run again and produce slightly different results every time becuase of the random varibles in the model. </p>\n\n<p>What I wonder is it allowed to resubmit the same solution multiple times without new commits to get a good training run?.</p>\n\n<p>I think this is better than just change a few characters and waste 2 hours of GPU time on commiting a equivalent solution!</p>",
      "rawMarkdown": "Hi, I currently have a model that have a score of 0.99 and have  good place on the leader board. When I submit a solution the model will run again and produce slightly different results every time becuase of the random varibles in the model. \n\nWhat I wonder is it allowed to resubmit the same solution multiple times without new commits to get a good training run?.\n\nI think this is better than just change a few characters and waste 2 hours of GPU time on commiting a equivalent solution!",
      "votes": null
    },
    {
      "id": "661592",
      "postDate": "10/30/2019 13:46:08",
      "content": "<p>If I do this i am probably just going to overfit on the public test set. </p>",
      "rawMarkdown": "If I do this i am probably just going to overfit on the public test set.",
      "votes": null
    },
    {
      "id": "661596",
      "postDate": "10/30/2019 13:48:36",
      "content": "<p>May I suggest using random.seed() so you can initialize the random seed and don't get different results? Or also you can define in many models a random_state variable. In that way you can control better your output.</p>",
      "rawMarkdown": "May I suggest using random.seed() so you can initialize the random seed and don't get different results? Or also you can define in many models a random_state variable. In that way you can control better your output.",
      "votes": null
    },
    {
      "id": "661602",
      "postDate": "10/30/2019 13:57:01",
      "content": "<p>That is a very good idea, since I have had a pretty random experience between runs. </p>",
      "rawMarkdown": "That is a very good idea, since I have had a pretty random experience between runs.",
      "votes": null
    },
    {
      "id": "663591",
      "postDate": "11/02/2019 11:37:33",
      "content": "<p>That is definitely a risk of doing it the way you described.</p>",
      "rawMarkdown": "That is definitely a risk of doing it the way you described.",
      "votes": null
    },
    {
      "id": "680213",
      "postDate": "11/24/2019 08:34:25",
      "content": "<p>hi, I would save the weight set after training and when submitting a solution I would reload the set and do a no-training but predict only kernel. please find a sample <a href=\"https://www.kaggle.com/nandor65/kmnist-trio\">here</a>. runtime is less than 60 sec without GPU. hope this helps.</p>",
      "rawMarkdown": "hi, I would save the weight set after training and when submitting a solution I would reload the set and do a no-training but predict only kernel. please find a sample [here](https://www.kaggle.com/nandor65/kmnist-trio). runtime is less than 60 sec without GPU. hope this helps.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 661592,
      "author_name": "rincewind007",
      "author_url": "",
      "post_date": "10/30/2019 13:46:08",
      "content": "<p>If I do this i am probably just going to overfit on the public test set. </p>",
      "votes": null,
      "replies": [
        {
          "id": 663591,
          "author_name": "datahobbit",
          "author_url": "",
          "post_date": "11/02/2019 11:37:33",
          "content": "<p>That is definitely a risk of doing it the way you described.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 661596,
      "author_name": "sebastianmpalacio",
      "author_url": "",
      "post_date": "10/30/2019 13:48:36",
      "content": "<p>May I suggest using random.seed() so you can initialize the random seed and don't get different results? Or also you can define in many models a random_state variable. In that way you can control better your output.</p>",
      "votes": null,
      "replies": [
        {
          "id": 661602,
          "author_name": "rincewind007",
          "author_url": "",
          "post_date": "10/30/2019 13:57:01",
          "content": "<p>That is a very good idea, since I have had a pretty random experience between runs. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 680213,
      "author_name": "nandor65",
      "author_url": "",
      "post_date": "11/24/2019 08:34:25",
      "content": "<p>hi, I would save the weight set after training and when submitting a solution I would reload the set and do a no-training but predict only kernel. please find a sample <a href=\"https://www.kaggle.com/nandor65/kmnist-trio\">here</a>. runtime is less than 60 sec without GPU. hope this helps.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "661570": "Hi, I currently have a model that have a score of 0.99 and have  good place on the leader board. When I submit a solution the model will run again and produce slightly different results every time becuase of the random varibles in the model. \n\nWhat I wonder is it allowed to resubmit the same solution multiple times without new commits to get a good training run?.\n\nI think this is better than just change a few characters and waste 2 hours of GPU time on commiting a equivalent solution!",
    "661592": "If I do this i am probably just going to overfit on the public test set.",
    "661596": "May I suggest using random.seed() so you can initialize the random seed and don't get different results? Or also you can define in many models a random_state variable. In that way you can control better your output.",
    "661602": "That is a very good idea, since I have had a pretty random experience between runs.",
    "663591": "That is definitely a risk of doing it the way you described.",
    "680213": "hi, I would save the weight set after training and when submitting a solution I would reload the set and do a no-training but predict only kernel. please find a sample [here](https://www.kaggle.com/nandor65/kmnist-trio). runtime is less than 60 sec without GPU. hope this helps."
  },
  "source": "meta"
}