{
  "id": 543847,
  "title": "Combination of 1D CNN and 2D CNN on Ariel 2024 Challenge",
  "url": "/competitions/ariel-data-challenge-2024/discussion/543847",
  "author_name": "John",
  "post_date": "2024-11-01T18:49:06.128000",
  "votes": 4,
  "comment_count": 0,
  "views": 0,
  "content": "<h1>Summary</h1>\n<p>Inspired by the Host's starter notebook, we focused on the CNN approach in this competition. We utilized the 1D CNN ensemble and the shared polynomial fitting notebook to estimate the sigma of target mean. 2D CNN ensemble was used to estimate the residual of each wavelength and the uncertainty. The final result was 0.611 (67th) on the private leaderboard.  It was a great experience to learn what failed and what worked of the CNN model on this problem.</p>\n<h1>Lessons learned</h1>\n<ul>\n<li>Our 1D CNN approach on target mean (average transit depth across wavelengths) can not beat the polynomial fitting on CV set due to the model generalization issue. However, the difference of CNN prediction and polynomial fit    <br>\ngave better sigma estimation of the target prediction. </li>\n<li>Our 2D CNN on resuduals did help the overal prediction by combining the mean prediction and residual prediction. The sigma on residual was estimated by the std of a set of 2D CNN models trained by different random samples.</li>\n<li>The final sigma was estimated by the maximum of sigma from 1D and sigma 2D which is better than one sigma.</li>\n</ul>\n<h1>Acknowledgement</h1>\n<ul>\n<li>Organizers for the competition and the starter notebook</li>\n<li>Shared notebooks by Sergei Fironov, qianc, et. al.</li>\n<li>Many insightful discussions and questions by the organizers and participants</li>\n</ul>\n<p>Final notebook is here: <a href=\"https://www.kaggle.com/code/jxiesd/combination-of-1d-cnn-and-2d-cnn-on-ariel2024\" target=\"_blank\">https://www.kaggle.com/code/jxiesd/combination-of-1d-cnn-and-2d-cnn-on-ariel2024</a></p>",
  "messages": [
    {
      "id": 3034081,
      "postDate": "2024-11-01T18:49:06.127Z",
      "content": "<h1>Summary</h1>\n<p>Inspired by the Host's starter notebook, we focused on the CNN approach in this competition. We utilized the 1D CNN ensemble and the shared polynomial fitting notebook to estimate the sigma of target mean. 2D CNN ensemble was used to estimate the residual of each wavelength and the uncertainty. The final result was 0.611 (67th) on the private leaderboard.  It was a great experience to learn what failed and what worked of the CNN model on this problem.</p>\n<h1>Lessons learned</h1>\n<ul>\n<li>Our 1D CNN approach on target mean (average transit depth across wavelengths) can not beat the polynomial fitting on CV set due to the model generalization issue. However, the difference of CNN prediction and polynomial fit    <br>\ngave better sigma estimation of the target prediction. </li>\n<li>Our 2D CNN on resuduals did help the overal prediction by combining the mean prediction and residual prediction. The sigma on residual was estimated by the std of a set of 2D CNN models trained by different random samples.</li>\n<li>The final sigma was estimated by the maximum of sigma from 1D and sigma 2D which is better than one sigma.</li>\n</ul>\n<h1>Acknowledgement</h1>\n<ul>\n<li>Organizers for the competition and the starter notebook</li>\n<li>Shared notebooks by Sergei Fironov, qianc, et. al.</li>\n<li>Many insightful discussions and questions by the organizers and participants</li>\n</ul>\n<p>Final notebook is here: <a href=\"https://www.kaggle.com/code/jxiesd/combination-of-1d-cnn-and-2d-cnn-on-ariel2024\" target=\"_blank\">https://www.kaggle.com/code/jxiesd/combination-of-1d-cnn-and-2d-cnn-on-ariel2024</a></p>",
      "rawMarkdown": "# Summary\nInspired by the Host's starter notebook, we focused on the CNN approach in this competition. We utilized the 1D CNN ensemble and the shared polynomial fitting notebook to estimate the sigma of target mean. 2D CNN ensemble was used to estimate the residual of each wavelength and the uncertainty. The final result was 0.611 (67th) on the private leaderboard.  It was a great experience to learn what failed and what worked of the CNN model on this problem.\n# Lessons learned \n- Our 1D CNN approach on target mean (average transit depth across wavelengths) can not beat the polynomial fitting on CV set due to the model generalization issue. However, the difference of CNN prediction and polynomial fit    \ngave better sigma estimation of the target prediction. \n- Our 2D CNN on resuduals did help the overal prediction by combining the mean prediction and residual prediction. The sigma on residual was estimated by the std of a set of 2D CNN models trained by different random samples.\n- The final sigma was estimated by the maximum of sigma from 1D and sigma 2D which is better than one sigma.\n# Acknowledgement\n- Organizers for the competition and the starter notebook\n- Shared notebooks by Sergei Fironov, qianc, et. al.\n- Many insightful discussions and questions by the organizers and participants\n\nFinal notebook is here: https://www.kaggle.com/code/jxiesd/combination-of-1d-cnn-and-2d-cnn-on-ariel2024",
      "votes": 4
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3034081": "# Summary\nInspired by the Host's starter notebook, we focused on the CNN approach in this competition. We utilized the 1D CNN ensemble and the shared polynomial fitting notebook to estimate the sigma of target mean. 2D CNN ensemble was used to estimate the residual of each wavelength and the uncertainty. The final result was 0.611 (67th) on the private leaderboard.  It was a great experience to learn what failed and what worked of the CNN model on this problem.\n# Lessons learned \n- Our 1D CNN approach on target mean (average transit depth across wavelengths) can not beat the polynomial fitting on CV set due to the model generalization issue. However, the difference of CNN prediction and polynomial fit    \ngave better sigma estimation of the target prediction. \n- Our 2D CNN on resuduals did help the overal prediction by combining the mean prediction and residual prediction. The sigma on residual was estimated by the std of a set of 2D CNN models trained by different random samples.\n- The final sigma was estimated by the maximum of sigma from 1D and sigma 2D which is better than one sigma.\n# Acknowledgement\n- Organizers for the competition and the starter notebook\n- Shared notebooks by Sergei Fironov, qianc, et. al.\n- Many insightful discussions and questions by the organizers and participants\n\nFinal notebook is here: https://www.kaggle.com/code/jxiesd/combination-of-1d-cnn-and-2d-cnn-on-ariel2024"
  }
}