{
  "id": 129456,
  "title": "Strong correlation between val_grapheme_recall and LB",
  "url": "/competitions/bengaliai-cv19/discussion/129456",
  "author_name": "",
  "post_date": "2020-02-08T04:33:00.859230500Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>During experiments, I've found that LB correlates well with val grapheme recall.\nModel: Densenet121 (128x128x1)\nAugmentation: Random Rotate (20 Degrees) and Cutout\nVal Split: 0.15 with fixed seeds\nVal Recall: 0.9795\nVal Grapheme Recall: 0.968\nLB: 0.9636\nThis is just one of my experiments, other experiments also show similar results (same setup but different models), high val grapheme recall -&gt; high LB, and the difference is very close (roughly ~0.004 to 0.006). </p>",
  "messages": [
    {
      "id": "739623",
      "postDate": "02/08/2020 04:33:00",
      "content": "<p>During experiments, I've found that LB correlates well with val grapheme recall.\nModel: Densenet121 (128x128x1)\nAugmentation: Random Rotate (20 Degrees) and Cutout\nVal Split: 0.15 with fixed seeds\nVal Recall: 0.9795\nVal Grapheme Recall: 0.968\nLB: 0.9636\nThis is just one of my experiments, other experiments also show similar results (same setup but different models), high val grapheme recall -&gt; high LB, and the difference is very close (roughly ~0.004 to 0.006). </p>",
      "rawMarkdown": "During experiments, I've found that LB correlates well with val grapheme recall.\nModel: Densenet121 (128x128x1)\nAugmentation: Random Rotate (20 Degrees) and Cutout\nVal Split: 0.15 with fixed seeds\nVal Recall: 0.9795\nVal Grapheme Recall: 0.968\nLB: 0.9636\nThis is just one of my experiments, other experiments also show similar results (same setup but different models), high val grapheme recall -&gt; high LB, and the difference is very close (roughly ~0.004 to 0.006).",
      "votes": null
    },
    {
      "id": "739720",
      "postDate": "02/08/2020 08:55:46",
      "content": "<p>The grapheme root has a double weight in the evaluation metric compared with the other 2 components. This is why the correlation is strong, it's impact on the public leaderboard score is the strongest (it is also the one with the lower recall, at lest in my tests so far)</p>",
      "rawMarkdown": "The grapheme root has a double weight in the evaluation metric compared with the other 2 components. This is why the correlation is strong, it's impact on the public leaderboard score is the strongest (it is also the one with the lower recall, at lest in my tests so far)",
      "votes": null
    },
    {
      "id": "740609",
      "postDate": "02/09/2020 15:29:54",
      "content": "<p>Yes, seems like the key to breaking into the gold is to be able to perfect the grapheme recall - the other two components are fairly trivial in comparison.</p>",
      "rawMarkdown": "Yes, seems like the key to breaking into the gold is to be able to perfect the grapheme recall - the other two components are fairly trivial in comparison.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 739720,
      "author_name": "vladvdv",
      "author_url": "",
      "post_date": "02/08/2020 08:55:46",
      "content": "<p>The grapheme root has a double weight in the evaluation metric compared with the other 2 components. This is why the correlation is strong, it's impact on the public leaderboard score is the strongest (it is also the one with the lower recall, at lest in my tests so far)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 740609,
      "author_name": "rohitagarwal",
      "author_url": "",
      "post_date": "02/09/2020 15:29:54",
      "content": "<p>Yes, seems like the key to breaking into the gold is to be able to perfect the grapheme recall - the other two components are fairly trivial in comparison.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "739623": "During experiments, I've found that LB correlates well with val grapheme recall.\nModel: Densenet121 (128x128x1)\nAugmentation: Random Rotate (20 Degrees) and Cutout\nVal Split: 0.15 with fixed seeds\nVal Recall: 0.9795\nVal Grapheme Recall: 0.968\nLB: 0.9636\nThis is just one of my experiments, other experiments also show similar results (same setup but different models), high val grapheme recall -&gt; high LB, and the difference is very close (roughly ~0.004 to 0.006).",
    "739720": "The grapheme root has a double weight in the evaluation metric compared with the other 2 components. This is why the correlation is strong, it's impact on the public leaderboard score is the strongest (it is also the one with the lower recall, at lest in my tests so far)",
    "740609": "Yes, seems like the key to breaking into the gold is to be able to perfect the grapheme recall - the other two components are fairly trivial in comparison."
  },
  "source": "meta"
}