{
  "id": 189242,
  "title": "67th Place Solution but Just ElasticNet performed way better!",
  "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/189242",
  "author_name": "Jagadish Sivakumar",
  "post_date": "2020-10-07T03:55:25.968000",
  "votes": 9,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Thanks to OSIC and Kaggle for hosting this competition. It was lot of fun exploring and learning from public kernels, discussion forums and testing new methods</p>\n<p><strong>Brief Summary</strong><br>\nOur Solution is an ensemble of EffNet B5 with competition metrics for evaluating CV performance. Quantile Regression with modified loss functions and ElasticNet</p>\n<ul>\n<li>EffNet B5 -&gt; (520x520) -&gt; Linear Decay based on EffNet for coefficient prediction, CV performance evaluated based on competition metrics</li>\n<li>Quantile Regression with modified loss function</li>\n</ul>\n<pre><code>def new_asy_qloss(y_true,y_pred):\n    qs = [0.2, 0.50, 0.8]\n    q = tf.constant(np.array([qs]), dtype=tf.float32)\n    e = y_true - y_pred\n    epsilon = 0.8\n\n    v = tf.maximum( -(1-q)*(e+q*epsilon), q*(e-(1-q)*epsilon))\n    v1 = tf.maximum(v,0.0)\n    return K.mean(v1)\n</code></pre>\n<ul>\n<li>ElasticNet with 10folds on GroupKFold </li>\n</ul>\n<p>And finally an ensemble of all the models.</p>\n<p><strong>Just an ElasticNet on 10folds with GroupKfold performed better compared to our ensemble submission.</strong></p>\n<p><a href=\"https://www.kaggle.com/jagadish13/osic-baseline-elasticnet-eda-testing?scriptVersionId=43936071\" target=\"_blank\">Check: ElasticNet Final Submission</a></p>\n<p><strong>Congrats to all the winners!</strong></p>",
  "messages": [
    {
      "id": 1040250,
      "postDate": "2020-10-07T03:55:25.970Z",
      "content": "<p>Thanks to OSIC and Kaggle for hosting this competition. It was lot of fun exploring and learning from public kernels, discussion forums and testing new methods</p>\n<p><strong>Brief Summary</strong><br>\nOur Solution is an ensemble of EffNet B5 with competition metrics for evaluating CV performance. Quantile Regression with modified loss functions and ElasticNet</p>\n<ul>\n<li>EffNet B5 -&gt; (520x520) -&gt; Linear Decay based on EffNet for coefficient prediction, CV performance evaluated based on competition metrics</li>\n<li>Quantile Regression with modified loss function</li>\n</ul>\n<pre><code>def new_asy_qloss(y_true,y_pred):\n    qs = [0.2, 0.50, 0.8]\n    q = tf.constant(np.array([qs]), dtype=tf.float32)\n    e = y_true - y_pred\n    epsilon = 0.8\n\n    v = tf.maximum( -(1-q)*(e+q*epsilon), q*(e-(1-q)*epsilon))\n    v1 = tf.maximum(v,0.0)\n    return K.mean(v1)\n</code></pre>\n<ul>\n<li>ElasticNet with 10folds on GroupKFold </li>\n</ul>\n<p>And finally an ensemble of all the models.</p>\n<p><strong>Just an ElasticNet on 10folds with GroupKfold performed better compared to our ensemble submission.</strong></p>\n<p><a href=\"https://www.kaggle.com/jagadish13/osic-baseline-elasticnet-eda-testing?scriptVersionId=43936071\" target=\"_blank\">Check: ElasticNet Final Submission</a></p>\n<p><strong>Congrats to all the winners!</strong></p>",
      "rawMarkdown": "Thanks to OSIC and Kaggle for hosting this competition. It was lot of fun exploring and learning from public kernels, discussion forums and testing new methods\n\n**Brief Summary**\nOur Solution is an ensemble of EffNet B5 with competition metrics for evaluating CV performance. Quantile Regression with modified loss functions and ElasticNet\n\n- EffNet B5 -> (520x520) -> Linear Decay based on EffNet for coefficient prediction, CV performance evaluated based on competition metrics\n- Quantile Regression with modified loss function\n\n```\ndef new_asy_qloss(y_true,y_pred):\n    qs = [0.2, 0.50, 0.8]\n    q = tf.constant(np.array([qs]), dtype=tf.float32)\n    e = y_true - y_pred\n    epsilon = 0.8\n  \n    v = tf.maximum( -(1-q)*(e+q*epsilon), q*(e-(1-q)*epsilon))\n    v1 = tf.maximum(v,0.0)\n    return K.mean(v1)\n```\n- ElasticNet with 10folds on GroupKFold \n\nAnd finally an ensemble of all the models.\n\n**Just an ElasticNet on 10folds with GroupKfold performed better compared to our ensemble submission.**\n\n[Check: ElasticNet Final Submission](https://www.kaggle.com/jagadish13/osic-baseline-elasticnet-eda-testing?scriptVersionId=43936071)\n\n**Congrats to all the winners!**",
      "votes": 9
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1040250": "Thanks to OSIC and Kaggle for hosting this competition. It was lot of fun exploring and learning from public kernels, discussion forums and testing new methods\n\n**Brief Summary**\nOur Solution is an ensemble of EffNet B5 with competition metrics for evaluating CV performance. Quantile Regression with modified loss functions and ElasticNet\n\n- EffNet B5 -> (520x520) -> Linear Decay based on EffNet for coefficient prediction, CV performance evaluated based on competition metrics\n- Quantile Regression with modified loss function\n\n```\ndef new_asy_qloss(y_true,y_pred):\n    qs = [0.2, 0.50, 0.8]\n    q = tf.constant(np.array([qs]), dtype=tf.float32)\n    e = y_true - y_pred\n    epsilon = 0.8\n  \n    v = tf.maximum( -(1-q)*(e+q*epsilon), q*(e-(1-q)*epsilon))\n    v1 = tf.maximum(v,0.0)\n    return K.mean(v1)\n```\n- ElasticNet with 10folds on GroupKFold \n\nAnd finally an ensemble of all the models.\n\n**Just an ElasticNet on 10folds with GroupKfold performed better compared to our ensemble submission.**\n\n[Check: ElasticNet Final Submission](https://www.kaggle.com/jagadish13/osic-baseline-elasticnet-eda-testing?scriptVersionId=43936071)\n\n**Congrats to all the winners!**"
  }
}