{
  "id": 651809,
  "title": "Spatio-temporal Transformer model",
  "url": "/competitions/MABe-mouse-behavior-detection/discussion/651809",
  "author_name": "Kh0a",
  "post_date": "2025-12-05T11:40:36.590000",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I've joined this competition a bit late and have been attempting to implement a Spatio-temporal Transformer model. Unfortunately, I am finding it quite challenging to fit all the data for training due to the significant computational cost.</p>\n<p>My input to model is (node_features,edge_features,static_features)</p>\n<p>node_features: (B, T, N, F_node)</p>\n<p>edge_features: (B, T, N, N, F_edge)</p>\n<p>static_features: (B, N, F_static)</p>\n<p>And output is (B, T, agent, target, A)</p>\n<p>In contrast, the public kernels are mostly tree models, and their scores are very strong when compared to my own CV score from a CNN model.</p>\n<p>I would be highly appreciative if you could share your experiments and successful techniques from this competition.</p>\n<p>Thanks!  🤗</p>",
  "messages": [
    {
      "id": 3362831,
      "postDate": "2025-12-05T11:40:36.590Z",
      "content": "<p>I've joined this competition a bit late and have been attempting to implement a Spatio-temporal Transformer model. Unfortunately, I am finding it quite challenging to fit all the data for training due to the significant computational cost.</p>\n<p>My input to model is (node_features,edge_features,static_features)</p>\n<p>node_features: (B, T, N, F_node)</p>\n<p>edge_features: (B, T, N, N, F_edge)</p>\n<p>static_features: (B, N, F_static)</p>\n<p>And output is (B, T, agent, target, A)</p>\n<p>In contrast, the public kernels are mostly tree models, and their scores are very strong when compared to my own CV score from a CNN model.</p>\n<p>I would be highly appreciative if you could share your experiments and successful techniques from this competition.</p>\n<p>Thanks!  🤗</p>",
      "rawMarkdown": "I've joined this competition a bit late and have been attempting to implement a Spatio-temporal Transformer model. Unfortunately, I am finding it quite challenging to fit all the data for training due to the significant computational cost.\n\nMy input to model is (node_features,edge_features,static_features)\n\nnode_features: (B, T, N, F_node)\n\nedge_features: (B, T, N, N, F_edge)\n\nstatic_features: (B, N, F_static)\n\nAnd output is (B, T, agent, target, A)\n\nIn contrast, the public kernels are mostly tree models, and their scores are very strong when compared to my own CV score from a CNN model.\n\nI would be highly appreciative if you could share your experiments and successful techniques from this competition.\n\nThanks!  🤗",
      "votes": 2
    },
    {
      "id": 3369836,
      "postDate": "2025-12-10T11:38:43.193Z",
      "content": "<p>Tried a TCN plus ST GCN </p>",
      "rawMarkdown": "Tried a TCN plus ST GCN "
    },
    {
      "id": 3364275,
      "postDate": "2025-12-06T11:33:58.627Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 3369836,
      "author_name": "LUKORITO",
      "author_url": "",
      "post_date": "2025-12-10T11:38:43.193000",
      "content": "<p>Tried a TCN plus ST GCN </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3364275,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-12-06T11:33:58.627000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3362831": "I've joined this competition a bit late and have been attempting to implement a Spatio-temporal Transformer model. Unfortunately, I am finding it quite challenging to fit all the data for training due to the significant computational cost.\n\nMy input to model is (node_features,edge_features,static_features)\n\nnode_features: (B, T, N, F_node)\n\nedge_features: (B, T, N, N, F_edge)\n\nstatic_features: (B, N, F_static)\n\nAnd output is (B, T, agent, target, A)\n\nIn contrast, the public kernels are mostly tree models, and their scores are very strong when compared to my own CV score from a CNN model.\n\nI would be highly appreciative if you could share your experiments and successful techniques from this competition.\n\nThanks!  🤗",
    "3369836": "Tried a TCN plus ST GCN ",
    "3364275": ""
  }
}