{
  "id": 520405,
  "title": "Kudos to everybody and some late questions",
  "url": "/competitions/leap-atmospheric-physics-ai-climsim/discussion/520405",
  "author_name": "Fernando Melo",
  "post_date": "2024-07-15T21:22:37.468000",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>First, I want to thank everybody for sharing their knowledge, especially from GreySnow and Amadeo for sharing their knowledge and great notebooks. This is probably the competition where I learned the most during my entire time on Kaggle. I hope to learn more about tricks and architectures of NNs with the revealed solutions.</p>\n<p>Second, I still have some questions about this competition (of course, just for learning purposes, as the competition itself is over). I mainly focus on shared notebooks and evolve starting from there, so I still have some concepts to fully understand. Why transform regular tabular data to seq2seq? Why do this problem and data fit this approach? In the last few days, everybody has been talking about 1D&gt;1D, 2D&gt;1D, etc. Considering the shapes (60, 25) and (60, 14), isn't it by nature a 2D&gt;2D competition? Can someone explain it to me? Why do Encoder and Encoder/Decoder fit well?</p>",
  "messages": [
    {
      "id": 2923417,
      "postDate": "2024-07-15T21:22:37.467Z",
      "content": "<p>First, I want to thank everybody for sharing their knowledge, especially from GreySnow and Amadeo for sharing their knowledge and great notebooks. This is probably the competition where I learned the most during my entire time on Kaggle. I hope to learn more about tricks and architectures of NNs with the revealed solutions.</p>\n<p>Second, I still have some questions about this competition (of course, just for learning purposes, as the competition itself is over). I mainly focus on shared notebooks and evolve starting from there, so I still have some concepts to fully understand. Why transform regular tabular data to seq2seq? Why do this problem and data fit this approach? In the last few days, everybody has been talking about 1D&gt;1D, 2D&gt;1D, etc. Considering the shapes (60, 25) and (60, 14), isn't it by nature a 2D&gt;2D competition? Can someone explain it to me? Why do Encoder and Encoder/Decoder fit well?</p>",
      "rawMarkdown": "First, I want to thank everybody for sharing their knowledge, especially from GreySnow and Amadeo for sharing their knowledge and great notebooks. This is probably the competition where I learned the most during my entire time on Kaggle. I hope to learn more about tricks and architectures of NNs with the revealed solutions.\n\nSecond, I still have some questions about this competition (of course, just for learning purposes, as the competition itself is over). I mainly focus on shared notebooks and evolve starting from there, so I still have some concepts to fully understand. Why transform regular tabular data to seq2seq? Why do this problem and data fit this approach? In the last few days, everybody has been talking about 1D>1D, 2D>1D, etc. Considering the shapes (60, 25) and (60, 14), isn't it by nature a 2D>2D competition? Can someone explain it to me? Why do Encoder and Encoder/Decoder fit well?",
      "votes": 1
    },
    {
      "id": 2923437,
      "postDate": "2024-07-15T22:06:42.493Z",
      "content": "<p>Every row in the table corresponds to a 1-dimensional atmospheric column. 2D -&gt; 1D or 2D -&gt; 2D corresponds to using multiple atmospheric columns from the same timestep to predict one or multiple columns from the same timestep.</p>",
      "rawMarkdown": "Every row in the table corresponds to a 1-dimensional atmospheric column. 2D -> 1D or 2D -> 2D corresponds to using multiple atmospheric columns from the same timestep to predict one or multiple columns from the same timestep.",
      "votes": 2
    },
    {
      "id": 2923969,
      "postDate": "2024-07-16T08:17:10.953Z",
      "content": "<p>When they are talking about shapes and dimensions, they are referring to spatial dimensions which is 60 here. They are not including features/channels (dimension with 25 inputs and 14 outputs), so 1 spatial dimension means 1D. There is also an implicit location information in data which can be added as a dimension and data can be represented as 2D -&gt; (25/14, 384, 60). Host didn't want that.</p>",
      "rawMarkdown": "When they are talking about shapes and dimensions, they are referring to spatial dimensions which is 60 here. They are not including features/channels (dimension with 25 inputs and 14 outputs), so 1 spatial dimension means 1D. There is also an implicit location information in data which can be added as a dimension and data can be represented as 2D -> (25/14, 384, 60). Host didn't want that."
    }
  ],
  "comments": [
    {
      "id": 2923437,
      "author_name": "Jerry Lin",
      "author_url": "",
      "post_date": "2024-07-15T22:06:42.493000",
      "content": "<p>Every row in the table corresponds to a 1-dimensional atmospheric column. 2D -&gt; 1D or 2D -&gt; 2D corresponds to using multiple atmospheric columns from the same timestep to predict one or multiple columns from the same timestep.</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 2923969,
      "author_name": "Gunes Evitan",
      "author_url": "",
      "post_date": "2024-07-16T08:17:10.953000",
      "content": "<p>When they are talking about shapes and dimensions, they are referring to spatial dimensions which is 60 here. They are not including features/channels (dimension with 25 inputs and 14 outputs), so 1 spatial dimension means 1D. There is also an implicit location information in data which can be added as a dimension and data can be represented as 2D -&gt; (25/14, 384, 60). Host didn't want that.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2923417": "First, I want to thank everybody for sharing their knowledge, especially from GreySnow and Amadeo for sharing their knowledge and great notebooks. This is probably the competition where I learned the most during my entire time on Kaggle. I hope to learn more about tricks and architectures of NNs with the revealed solutions.\n\nSecond, I still have some questions about this competition (of course, just for learning purposes, as the competition itself is over). I mainly focus on shared notebooks and evolve starting from there, so I still have some concepts to fully understand. Why transform regular tabular data to seq2seq? Why do this problem and data fit this approach? In the last few days, everybody has been talking about 1D>1D, 2D>1D, etc. Considering the shapes (60, 25) and (60, 14), isn't it by nature a 2D>2D competition? Can someone explain it to me? Why do Encoder and Encoder/Decoder fit well?",
    "2923437": "Every row in the table corresponds to a 1-dimensional atmospheric column. 2D -> 1D or 2D -> 2D corresponds to using multiple atmospheric columns from the same timestep to predict one or multiple columns from the same timestep.",
    "2923969": "When they are talking about shapes and dimensions, they are referring to spatial dimensions which is 60 here. They are not including features/channels (dimension with 25 inputs and 14 outputs), so 1 spatial dimension means 1D. There is also an implicit location information in data which can be added as a dimension and data can be represented as 2D -> (25/14, 384, 60). Host didn't want that."
  }
}