{
  "id": 357900,
  "title": "Predicting Next Frame",
  "url": "/competitions/tabular-playground-series-oct-2022/discussion/357900",
  "author_name": "",
  "post_date": "2022-10-06T04:35:48.599385500Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I think intuitively I'm clear on how it might be useful to embed the positioning of all players and ball to one another, but I'm not entirely sure I have the implementation details ironed out in a way that makes sense. For that reason, I've created a public notebook that others can reference and hopefully use as a starting point to make improvements on: <a href=\"https://www.kaggle.com/code/ryancaldwell/cnn-predict-next-frame\" target=\"_blank\">https://www.kaggle.com/code/ryancaldwell/cnn-predict-next-frame</a></p>\n<p>Ideally we could take in a row from the test set and extract the representation of the field based on location and velocity of players and ball. That could be part of the input space to the classifier of which team scores within the next 10 seconds.</p>",
  "messages": [
    {
      "id": "1974106",
      "postDate": "10/06/2022 04:35:48",
      "content": "<p>I think intuitively I'm clear on how it might be useful to embed the positioning of all players and ball to one another, but I'm not entirely sure I have the implementation details ironed out in a way that makes sense. For that reason, I've created a public notebook that others can reference and hopefully use as a starting point to make improvements on: <a href=\"https://www.kaggle.com/code/ryancaldwell/cnn-predict-next-frame\" target=\"_blank\">https://www.kaggle.com/code/ryancaldwell/cnn-predict-next-frame</a></p>\n<p>Ideally we could take in a row from the test set and extract the representation of the field based on location and velocity of players and ball. That could be part of the input space to the classifier of which team scores within the next 10 seconds.</p>",
      "rawMarkdown": "I think intuitively I'm clear on how it might be useful to embed the positioning of all players and ball to one another, but I'm not entirely sure I have the implementation details ironed out in a way that makes sense. For that reason, I've created a public notebook that others can reference and hopefully use as a starting point to make improvements on: https://www.kaggle.com/code/ryancaldwell/cnn-predict-next-frame\n\nIdeally we could take in a row from the test set and extract the representation of the field based on location and velocity of players and ball. That could be part of the input space to the classifier of which team scores within the next 10 seconds.",
      "votes": null
    },
    {
      "id": "1974365",
      "postDate": "10/06/2022 07:55:43",
      "content": "<p>Just a random idea, could you try KNN with k=1? In theory the next frame will be 'closer' than all other frames</p>",
      "rawMarkdown": "Just a random idea, could you try KNN with k=1? In theory the next frame will be 'closer' than all other frames",
      "votes": null
    },
    {
      "id": "1974793",
      "postDate": "10/06/2022 13:05:05",
      "content": "<p>The idea of this notebook is to predict/reconstruct the next frame as a pre-task to create pos/vel embeddings, not necessarily to identify the next frame in the dataset. I guess you could use these embeddings and KNN on the test set if you were interested in ordering/pairing test records like they are in the training data. And if you were able to do so, then maybe a sequence model of each frame could be an option.</p>",
      "rawMarkdown": "The idea of this notebook is to predict/reconstruct the next frame as a pre-task to create pos/vel embeddings, not necessarily to identify the next frame in the dataset. I guess you could use these embeddings and KNN on the test set if you were interested in ordering/pairing test records like they are in the training data. And if you were able to do so, then maybe a sequence model of each frame could be an option.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1974365,
      "author_name": "samuelcortinhas",
      "author_url": "",
      "post_date": "10/06/2022 07:55:43",
      "content": "<p>Just a random idea, could you try KNN with k=1? In theory the next frame will be 'closer' than all other frames</p>",
      "votes": null,
      "replies": [
        {
          "id": 1974793,
          "author_name": "ryancaldwell",
          "author_url": "",
          "post_date": "10/06/2022 13:05:05",
          "content": "<p>The idea of this notebook is to predict/reconstruct the next frame as a pre-task to create pos/vel embeddings, not necessarily to identify the next frame in the dataset. I guess you could use these embeddings and KNN on the test set if you were interested in ordering/pairing test records like they are in the training data. And if you were able to do so, then maybe a sequence model of each frame could be an option.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1974106": "I think intuitively I'm clear on how it might be useful to embed the positioning of all players and ball to one another, but I'm not entirely sure I have the implementation details ironed out in a way that makes sense. For that reason, I've created a public notebook that others can reference and hopefully use as a starting point to make improvements on: https://www.kaggle.com/code/ryancaldwell/cnn-predict-next-frame\n\nIdeally we could take in a row from the test set and extract the representation of the field based on location and velocity of players and ball. That could be part of the input space to the classifier of which team scores within the next 10 seconds.",
    "1974365": "Just a random idea, could you try KNN with k=1? In theory the next frame will be 'closer' than all other frames",
    "1974793": "The idea of this notebook is to predict/reconstruct the next frame as a pre-task to create pos/vel embeddings, not necessarily to identify the next frame in the dataset. I guess you could use these embeddings and KNN on the test set if you were interested in ordering/pairing test records like they are in the training data. And if you were able to do so, then maybe a sequence model of each frame could be an option."
  },
  "source": "meta"
}