{
  "id": 251647,
  "title": "Huge regularization problems!",
  "url": "/competitions/seti-breakthrough-listen/discussion/251647",
  "author_name": "Kaeldric",
  "post_date": "2021-07-08T07:17:47.056000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hello community.<br>\nI'm having a lot of problems regulazing my network so I decided to publish my <a href=\"https://www.kaggle.com/kaeldric/aliens\" target=\"_blank\">notebook</a> hoping that someone could help me.<br>\nI am thinking of the competition data as time steps of frequency readings. So, keeping in mind one of the images we are analyzing, I stacked the images so that they can be read top-down. Each line of the image corresponds to a step of the sequence.<br>\nThen I implemented an LSTM network discovering that overfitting is easily achieved, without using particularly complex networks, but the validation set result is a disaster. Actually, the loss and the accuracy are not bad but the problem is the F1-score. Its value remain low and almost constant throughout the training phase. The confusion matrix is pretty crap too.<br>\nI have tried many combinations of regularization mechanisms, including noising data between the layers of the network, but none of these have improved the values of the score.<br>\nAt this point, I'm afraid, the error is in the way I created the data sequence.<br>\nThanks to anyone who wants to take a look at this.</p>",
  "messages": [
    {
      "id": 1380578,
      "postDate": "2021-07-08T07:17:47.057Z",
      "content": "<p>Hello community.<br>\nI'm having a lot of problems regulazing my network so I decided to publish my <a href=\"https://www.kaggle.com/kaeldric/aliens\" target=\"_blank\">notebook</a> hoping that someone could help me.<br>\nI am thinking of the competition data as time steps of frequency readings. So, keeping in mind one of the images we are analyzing, I stacked the images so that they can be read top-down. Each line of the image corresponds to a step of the sequence.<br>\nThen I implemented an LSTM network discovering that overfitting is easily achieved, without using particularly complex networks, but the validation set result is a disaster. Actually, the loss and the accuracy are not bad but the problem is the F1-score. Its value remain low and almost constant throughout the training phase. The confusion matrix is pretty crap too.<br>\nI have tried many combinations of regularization mechanisms, including noising data between the layers of the network, but none of these have improved the values of the score.<br>\nAt this point, I'm afraid, the error is in the way I created the data sequence.<br>\nThanks to anyone who wants to take a look at this.</p>",
      "rawMarkdown": "\nHello community.\n\nI'm having a lot of problems regulazing my network so I decided to publish my [notebook](https://www.kaggle.com/kaeldric/aliens) hoping that someone could help me.\n\nI am thinking of the competition data as time steps of frequency readings. So, keeping in mind one of the images we are analyzing, I stacked the images so that they can be read top-down. Each line of the image corresponds to a step of the sequence.\n\nThen I implemented an LSTM network discovering that overfitting is easily achieved, without using particularly complex networks, but the validation set result is a disaster. Actually, the loss and the accuracy are not bad but the problem is the F1-score. Its value remain low and almost constant throughout the training phase. The confusion matrix is pretty crap too.\n\nI have tried many combinations of regularization mechanisms, including noising data between the layers of the network, but none of these have improved the values of the score.\n\nAt this point, I'm afraid, the error is in the way I created the data sequence.\n\nThanks to anyone who wants to take a look at this.\n\n\n"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1380578": "\nHello community.\n\nI'm having a lot of problems regulazing my network so I decided to publish my [notebook](https://www.kaggle.com/kaeldric/aliens) hoping that someone could help me.\n\nI am thinking of the competition data as time steps of frequency readings. So, keeping in mind one of the images we are analyzing, I stacked the images so that they can be read top-down. Each line of the image corresponds to a step of the sequence.\n\nThen I implemented an LSTM network discovering that overfitting is easily achieved, without using particularly complex networks, but the validation set result is a disaster. Actually, the loss and the accuracy are not bad but the problem is the F1-score. Its value remain low and almost constant throughout the training phase. The confusion matrix is pretty crap too.\n\nI have tried many combinations of regularization mechanisms, including noising data between the layers of the network, but none of these have improved the values of the score.\n\nAt this point, I'm afraid, the error is in the way I created the data sequence.\n\nThanks to anyone who wants to take a look at this.\n\n\n"
  }
}