{
  "id": 417717,
  "title": "3rd Place Solution: Transformer+GRU",
  "url": "/competitions/tlvmc-parkinsons-freezing-gait-prediction/writeups/stochoshi-3rd-place-solution-transformer-gru",
  "author_name": "",
  "post_date": "2023-06-18T19:05:31.343Z",
  "votes": 7,
  "comment_count": 1,
  "views": 0,
  "content": "<p>The first and second place solution have much better fits to the individual defog/tdcs datasets;  this solution fit one model for both.</p>\n<p>Architectures: Deberta/VisionTransformer/VisionTransformerRelPos -&gt; LSTM/GRU<br>\n(typically 2-4 layers, with single-layer RNN)</p>\n<p>Patch sizes are 7-13, with sequences of 192-384 patches.</p>\n<p>All sequences have heavy augmentation--stretching, cropping, ablation, accumulated Gaussian noise, etc.</p>\n<p>--</p>\n<p>See data.py and model.py for details:</p>\n<p>Inference Code: <a href=\"https://www.kaggle.com/code/stochoshi/fork-of-walk3\" target=\"_blank\">https://www.kaggle.com/code/stochoshi/fork-of-walk3</a><br>\nAdditional Code: <a href=\"https://www.kaggle.com/datasets/stochoshi/walkdata4\" target=\"_blank\">https://www.kaggle.com/datasets/stochoshi/walkdata4</a></p>",
  "messages": [
    {
      "id": "2305729",
      "postDate": "06/16/2023 22:01:24",
      "content": "<p>The first and second place solution have much better fits to the individual defog/tdcs datasets;  this solution fit one model for both.</p>\n<p>Architectures: Deberta/VisionTransformer/VisionTransformerRelPos -&gt; LSTM/GRU<br>\n(typically 2-4 layers, with single-layer RNN)</p>\n<p>Patch sizes are 7-13, with sequences of 192-384 patches.</p>\n<p>All sequences have heavy augmentation--stretching, cropping, ablation, accumulated Gaussian noise, etc.</p>\n<p>--</p>\n<p>See data.py and model.py for details:</p>\n<p>Inference Code: <a href=\"https://www.kaggle.com/code/stochoshi/fork-of-walk3\" target=\"_blank\">https://www.kaggle.com/code/stochoshi/fork-of-walk3</a><br>\nAdditional Code: <a href=\"https://www.kaggle.com/datasets/stochoshi/walkdata4\" target=\"_blank\">https://www.kaggle.com/datasets/stochoshi/walkdata4</a></p>",
      "rawMarkdown": "The first and second place solution have much better fits to the individual defog/tdcs datasets;  this solution fit one model for both.\n\nArchitectures: Deberta/VisionTransformer/VisionTransformerRelPos -> LSTM/GRU\n(typically 2-4 layers, with single-layer RNN)\n\nPatch sizes are 7-13, with sequences of 192-384 patches.\n\nAll sequences have heavy augmentation--stretching, cropping, ablation, accumulated Gaussian noise, etc.\n\n--\n\nSee data.py and model.py for details:\n\nInference Code: https://www.kaggle.com/code/stochoshi/fork-of-walk3\nAdditional Code: https://www.kaggle.com/datasets/stochoshi/walkdata4",
      "votes": null
    },
    {
      "id": "2742069",
      "postDate": "04/08/2024 18:17:31",
      "content": "<p>Thanks for the clear code.<br>\nInspiring and elegant approach!</p>",
      "rawMarkdown": "Thanks for the clear code.\nInspiring and elegant approach!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2742069,
      "author_name": "liornis",
      "author_url": "",
      "post_date": "04/08/2024 18:17:31",
      "content": "<p>Thanks for the clear code.<br>\nInspiring and elegant approach!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2305729": "The first and second place solution have much better fits to the individual defog/tdcs datasets;  this solution fit one model for both.\n\nArchitectures: Deberta/VisionTransformer/VisionTransformerRelPos -> LSTM/GRU\n(typically 2-4 layers, with single-layer RNN)\n\nPatch sizes are 7-13, with sequences of 192-384 patches.\n\nAll sequences have heavy augmentation--stretching, cropping, ablation, accumulated Gaussian noise, etc.\n\n--\n\nSee data.py and model.py for details:\n\nInference Code: https://www.kaggle.com/code/stochoshi/fork-of-walk3\nAdditional Code: https://www.kaggle.com/datasets/stochoshi/walkdata4",
    "2742069": "Thanks for the clear code.\nInspiring and elegant approach!"
  },
  "source": "meta"
}