{
  "id": 239341,
  "title": "Simple EDA and starter [LB: 0.95] + some observations",
  "url": "/competitions/seti-breakthrough-listen/discussion/239341",
  "author_name": "ilovescience",
  "post_date": "2021-05-15T21:33:23.372000",
  "votes": 14,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I have posted two notebooks, one is an EDA notebook and the other is a quick starter.</p>\n<h2>SETI simple EDA to help you get started! 👽</h2>\n<p><a href=\"https://www.kaggle.com/tanlikesmath/seti-simple-eda-to-help-you-get-started\" target=\"_blank\">https://www.kaggle.com/tanlikesmath/seti-simple-eda-to-help-you-get-started</a></p>\n<h2>SETI ET signal detection - a simple CNN starter</h2>\n<p><a href=\"https://www.kaggle.com/tanlikesmath/seti-et-signal-detection-a-simple-cnn-starter\" target=\"_blank\">https://www.kaggle.com/tanlikesmath/seti-et-signal-detection-a-simple-cnn-starter</a></p>\n<h2>Some observations</h2>\n<p>As others have done as well, it became very clear to me that this would turn out to actually be a computer vision competition. Sure, other signal processing approaches may be useful, but I think the main model of any solution will likely be a CNN.</p>\n<p>One aspect that should be examined further is the significant imbalance in the dataset. Techniques like oversampling, focal loss, etc. might be worth investigating.</p>\n<p>I am interested by the superior performance of the spatial method proposed by others. Switching to the spatial method resulted in much improved AUC-ROC score (with less epochs, I might add).</p>\n<p>Fastai makes it trivial to set up and train models, especially for fine-tuning scenarios. Even with a custom PyTorch Dataset, it's easy to pass them into a <code>Learner</code> class and start fitting a model. The major benefit is that we don't need to write training loop, mixed precision support, LR schedules, etc. from scratch. </p>\n<p>Fine-tuning a simple ResNext50-32x4d achieved a score of 0.95. A single model! It's very clear that this competition is going to end with very high scores! </p>\n<p>I hope my notebooks are useful! Also I would appreciate any suggestions or feedback!</p>\n<p>Good luck everyone! 👽🙂</p>",
  "messages": [
    {
      "id": 1309365,
      "postDate": "2021-05-15T21:33:23.373Z",
      "content": "<p>I have posted two notebooks, one is an EDA notebook and the other is a quick starter.</p>\n<h2>SETI simple EDA to help you get started! 👽</h2>\n<p><a href=\"https://www.kaggle.com/tanlikesmath/seti-simple-eda-to-help-you-get-started\" target=\"_blank\">https://www.kaggle.com/tanlikesmath/seti-simple-eda-to-help-you-get-started</a></p>\n<h2>SETI ET signal detection - a simple CNN starter</h2>\n<p><a href=\"https://www.kaggle.com/tanlikesmath/seti-et-signal-detection-a-simple-cnn-starter\" target=\"_blank\">https://www.kaggle.com/tanlikesmath/seti-et-signal-detection-a-simple-cnn-starter</a></p>\n<h2>Some observations</h2>\n<p>As others have done as well, it became very clear to me that this would turn out to actually be a computer vision competition. Sure, other signal processing approaches may be useful, but I think the main model of any solution will likely be a CNN.</p>\n<p>One aspect that should be examined further is the significant imbalance in the dataset. Techniques like oversampling, focal loss, etc. might be worth investigating.</p>\n<p>I am interested by the superior performance of the spatial method proposed by others. Switching to the spatial method resulted in much improved AUC-ROC score (with less epochs, I might add).</p>\n<p>Fastai makes it trivial to set up and train models, especially for fine-tuning scenarios. Even with a custom PyTorch Dataset, it's easy to pass them into a <code>Learner</code> class and start fitting a model. The major benefit is that we don't need to write training loop, mixed precision support, LR schedules, etc. from scratch. </p>\n<p>Fine-tuning a simple ResNext50-32x4d achieved a score of 0.95. A single model! It's very clear that this competition is going to end with very high scores! </p>\n<p>I hope my notebooks are useful! Also I would appreciate any suggestions or feedback!</p>\n<p>Good luck everyone! 👽🙂</p>",
      "rawMarkdown": "I have posted two notebooks, one is an EDA notebook and the other is a quick starter.\n\n## SETI simple EDA to help you get started! 👽\nhttps://www.kaggle.com/tanlikesmath/seti-simple-eda-to-help-you-get-started\n\n## SETI ET signal detection - a simple CNN starter\nhttps://www.kaggle.com/tanlikesmath/seti-et-signal-detection-a-simple-cnn-starter\n\n\n## Some observations\nAs others have done as well, it became very clear to me that this would turn out to actually be a computer vision competition. Sure, other signal processing approaches may be useful, but I think the main model of any solution will likely be a CNN.\n\nOne aspect that should be examined further is the significant imbalance in the dataset. Techniques like oversampling, focal loss, etc. might be worth investigating.\n\nI am interested by the superior performance of the spatial method proposed by others. Switching to the spatial method resulted in much improved AUC-ROC score (with less epochs, I might add).\n\nFastai makes it trivial to set up and train models, especially for fine-tuning scenarios. Even with a custom PyTorch Dataset, it's easy to pass them into a `Learner` class and start fitting a model. The major benefit is that we don't need to write training loop, mixed precision support, LR schedules, etc. from scratch. \n\nFine-tuning a simple ResNext50-32x4d achieved a score of 0.95. A single model! It's very clear that this competition is going to end with very high scores! \n\nI hope my notebooks are useful! Also I would appreciate any suggestions or feedback!\n\nGood luck everyone! 👽🙂\n\n\n",
      "votes": 14
    },
    {
      "id": 1309377,
      "postDate": "2021-05-15T22:11:32.063Z",
      "content": "<p>Thanks for sharing. Are you a fastai promoter? 😁</p>",
      "rawMarkdown": "Thanks for sharing. Are you a fastai promoter? 😁",
      "votes": 1,
      "replies": [
        {
          "id": 1309380,
          "postDate": "2021-05-15T22:16:19.653Z",
          "content": "<p>Haha, I am an avid user and contributor of fastai… </p>\n<p>but I believe that it's important to use the tools that make your life easier, and in this case fastai certainly achieves that (but that isn't always true for all deep learning projects).</p>",
          "rawMarkdown": "Haha, I am an avid user and contributor of fastai... \n\nbut I believe that it's important to use the tools that make your life easier, and in this case fastai certainly achieves that (but that isn't always true for all deep learning projects).",
          "votes": 2
        },
        {
          "id": 1309393,
          "postDate": "2021-05-15T22:37:24.753Z",
          "content": "<p>I also started DL with fastai like many. I remember when I first built a model in several lines; the feeling was great. It was def a meaningful tool for me. What is your opinion on fastai’s source code? With my limited engineering skills, I find it hard to read.</p>",
          "rawMarkdown": "I also started DL with fastai like many. I remember when I first built a model in several lines; the feeling was great. It was def a meaningful tool for me. What is your opinion on fastai’s source code? With my limited engineering skills, I find it hard to read."
        },
        {
          "id": 1311084,
          "postDate": "2021-05-17T06:40:42.957Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1309377,
      "author_name": "sin",
      "author_url": "",
      "post_date": "2021-05-15T22:11:32.063000",
      "content": "<p>Thanks for sharing. Are you a fastai promoter? 😁</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1309380,
          "author_name": "ilovescience",
          "author_url": "",
          "post_date": "2021-05-15T22:16:19.653000",
          "content": "<p>Haha, I am an avid user and contributor of fastai… </p>\n<p>but I believe that it's important to use the tools that make your life easier, and in this case fastai certainly achieves that (but that isn't always true for all deep learning projects).</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1309393,
          "author_name": "sin",
          "author_url": "",
          "post_date": "2021-05-15T22:37:24.753000",
          "content": "<p>I also started DL with fastai like many. I remember when I first built a model in several lines; the feeling was great. It was def a meaningful tool for me. What is your opinion on fastai’s source code? With my limited engineering skills, I find it hard to read.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1311084,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-05-17T06:40:42.957000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1309365": "I have posted two notebooks, one is an EDA notebook and the other is a quick starter.\n\n## SETI simple EDA to help you get started! 👽\nhttps://www.kaggle.com/tanlikesmath/seti-simple-eda-to-help-you-get-started\n\n## SETI ET signal detection - a simple CNN starter\nhttps://www.kaggle.com/tanlikesmath/seti-et-signal-detection-a-simple-cnn-starter\n\n\n## Some observations\nAs others have done as well, it became very clear to me that this would turn out to actually be a computer vision competition. Sure, other signal processing approaches may be useful, but I think the main model of any solution will likely be a CNN.\n\nOne aspect that should be examined further is the significant imbalance in the dataset. Techniques like oversampling, focal loss, etc. might be worth investigating.\n\nI am interested by the superior performance of the spatial method proposed by others. Switching to the spatial method resulted in much improved AUC-ROC score (with less epochs, I might add).\n\nFastai makes it trivial to set up and train models, especially for fine-tuning scenarios. Even with a custom PyTorch Dataset, it's easy to pass them into a `Learner` class and start fitting a model. The major benefit is that we don't need to write training loop, mixed precision support, LR schedules, etc. from scratch. \n\nFine-tuning a simple ResNext50-32x4d achieved a score of 0.95. A single model! It's very clear that this competition is going to end with very high scores! \n\nI hope my notebooks are useful! Also I would appreciate any suggestions or feedback!\n\nGood luck everyone! 👽🙂\n\n\n",
    "1309377": "Thanks for sharing. Are you a fastai promoter? 😁"
  }
}