{
  "id": 59181,
  "title": "Problems training Point Net",
  "url": "/competitions/trackml-particle-identification/discussion/59181",
  "author_name": "bilal2vec",
  "post_date": "2018-06-19T14:06:26.827000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi,</p>\n\n<p>I implemented a Point Net similar to <a href=\"https://www.kaggle.com/c/trackml-particle-identification/discussion/58782\">this discussion post</a> and tried training it to segment mostly straight tracks (cone slice at 75 degrees) by giving it groups of 16 hits (x, y, z) to predict the one-hot encoded track id of each hit. Unfortunately, when training the model with the cross entropy loss and the SGD, the loss stays stuck at ~12. Does anyone know why this might be happening?</p>\n\n<p>Thanks</p>",
  "messages": [
    {
      "id": 345254,
      "postDate": "2018-06-19T14:06:26.827Z",
      "content": "<p>Hi,</p>\n\n<p>I implemented a Point Net similar to <a href=\"https://www.kaggle.com/c/trackml-particle-identification/discussion/58782\">this discussion post</a> and tried training it to segment mostly straight tracks (cone slice at 75 degrees) by giving it groups of 16 hits (x, y, z) to predict the one-hot encoded track id of each hit. Unfortunately, when training the model with the cross entropy loss and the SGD, the loss stays stuck at ~12. Does anyone know why this might be happening?</p>\n\n<p>Thanks</p>",
      "rawMarkdown": "Hi,\n\nI implemented a Point Net similar to [this discussion post][1] and tried training it to segment mostly straight tracks (cone slice at 75 degrees) by giving it groups of 16 hits (x, y, z) to predict the one-hot encoded track id of each hit. Unfortunately, when training the model with the cross entropy loss and the SGD, the loss stays stuck at ~12. Does anyone know why this might be happening?\n\nThanks\n\n[1]: https://www.kaggle.com/c/trackml-particle-identification/discussion/58782"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "345254": "Hi,\n\nI implemented a Point Net similar to [this discussion post][1] and tried training it to segment mostly straight tracks (cone slice at 75 degrees) by giving it groups of 16 hits (x, y, z) to predict the one-hot encoded track id of each hit. Unfortunately, when training the model with the cross entropy loss and the SGD, the loss stays stuck at ~12. Does anyone know why this might be happening?\n\nThanks\n\n[1]: https://www.kaggle.com/c/trackml-particle-identification/discussion/58782"
  }
}