{
  "id": 157258,
  "title": "Error analysis",
  "url": "/competitions/global-wheat-detection/discussion/157258",
  "author_name": "",
  "post_date": "2020-06-09T23:54:22.305085Z",
  "votes": 25,
  "comment_count": 3,
  "views": 0,
  "content": "<p>In the previous competitions, I did not spend much time analyzing my model's performance. I did what almost all of the beginners, tried different models, optimizers, losses, augmentations; I tried to implement more advanced techniques like mixup.</p>\n\n<p>I more or less entirely ignored the error analysis. <strong>That was a mistake!</strong></p>\n\n<p>Fortunately, now I had a bit of free time, so I created my first <a href=\"https://www.kaggle.com/pestipeti/error-analysis\">Error analysis notebook</a>. After I finished it, I instantly had lots of ideas on how and where I should improve my model.</p>\n\n<h2>Summary</h2>\n\n<ul>\n<li>Overall results (TP, FP, FN, sources, sizes, brightness, contrast)</li>\n<li>Results by sources (overall, FN/FP by sources, normal/border by sources, etc.)</li>\n<li>Results by size (small, medium, large; you can configure the sizes)</li>\n<li>Normal and \"border\" results</li>\n<li>Noisy targets</li>\n</ul>\n\n<h2>Model's error</h2>\n\n<p>Our models (or at least my models) have to handle these issues:\n- Different size of wheat-heads (too small, too large)\n- Group of wheat-heads\n- Partially invisible heads (covered by another head or a leaf)\n- Wheat-heads at the borders\n- Noisy boxes</p>\n\n<hr>\n\n<p><a href=\"https://www.kaggle.com/pestipeti/error-analysis\">You can find the notebook here</a>.</p>",
  "messages": [
    {
      "id": "880026",
      "postDate": "06/09/2020 23:54:22",
      "content": "<p>In the previous competitions, I did not spend much time analyzing my model's performance. I did what almost all of the beginners, tried different models, optimizers, losses, augmentations; I tried to implement more advanced techniques like mixup.</p>\n\n<p>I more or less entirely ignored the error analysis. <strong>That was a mistake!</strong></p>\n\n<p>Fortunately, now I had a bit of free time, so I created my first <a href=\"https://www.kaggle.com/pestipeti/error-analysis\">Error analysis notebook</a>. After I finished it, I instantly had lots of ideas on how and where I should improve my model.</p>\n\n<h2>Summary</h2>\n\n<ul>\n<li>Overall results (TP, FP, FN, sources, sizes, brightness, contrast)</li>\n<li>Results by sources (overall, FN/FP by sources, normal/border by sources, etc.)</li>\n<li>Results by size (small, medium, large; you can configure the sizes)</li>\n<li>Normal and \"border\" results</li>\n<li>Noisy targets</li>\n</ul>\n\n<h2>Model's error</h2>\n\n<p>Our models (or at least my models) have to handle these issues:\n- Different size of wheat-heads (too small, too large)\n- Group of wheat-heads\n- Partially invisible heads (covered by another head or a leaf)\n- Wheat-heads at the borders\n- Noisy boxes</p>\n\n<hr>\n\n<p><a href=\"https://www.kaggle.com/pestipeti/error-analysis\">You can find the notebook here</a>.</p>",
      "rawMarkdown": "In the previous competitions, I did not spend much time analyzing my model's performance. I did what almost all of the beginners, tried different models, optimizers, losses, augmentations; I tried to implement more advanced techniques like mixup.\n\nI more or less entirely ignored the error analysis. **That was a mistake!**\n\nFortunately, now I had a bit of free time, so I created my first [Error analysis notebook](https://www.kaggle.com/pestipeti/error-analysis). After I finished it, I instantly had lots of ideas on how and where I should improve my model.\n\n## Summary\n- Overall results (TP, FP, FN, sources, sizes, brightness, contrast)\n- Results by sources (overall, FN/FP by sources, normal/border by sources, etc.)\n- Results by size (small, medium, large; you can configure the sizes)\n- Normal and \"border\" results\n- Noisy targets\n\n## Model's error\nOur models (or at least my models) have to handle these issues:\n- Different size of wheat-heads (too small, too large)\n- Group of wheat-heads\n- Partially invisible heads (covered by another head or a leaf)\n- Wheat-heads at the borders\n- Noisy boxes\n\n\n------\n[You can find the notebook here](https://www.kaggle.com/pestipeti/error-analysis).",
      "votes": null
    },
    {
      "id": "880028",
      "postDate": "06/09/2020 23:56:32",
      "content": "<p>Good work Peter and thanks for your contributions! I agree, looking at examples of predictions vs. ground truth is a good method to find ways to improve</p>",
      "rawMarkdown": "Good work Peter and thanks for your contributions! I agree, looking at examples of predictions vs. ground truth is a good method to find ways to improve",
      "votes": null
    },
    {
      "id": "880524",
      "postDate": "06/10/2020 11:19:04",
      "content": "<p>Great!</p>",
      "rawMarkdown": "Great!",
      "votes": null
    },
    {
      "id": "892405",
      "postDate": "06/18/2020 21:11:41",
      "content": "<p>\"I did what almost all of the beginners, tried different models, optimizers, losses, augmentations\" I am one of those beginners😂 \nThanks a lot for this error analysis. I would use this approach for my experiments.</p>",
      "rawMarkdown": "\"I did what almost all of the beginners, tried different models, optimizers, losses, augmentations\" I am one of those beginners😂 \nThanks a lot for this error analysis. I would use this approach for my experiments.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 880028,
      "author_name": "matthewmasters",
      "author_url": "",
      "post_date": "06/09/2020 23:56:32",
      "content": "<p>Good work Peter and thanks for your contributions! I agree, looking at examples of predictions vs. ground truth is a good method to find ways to improve</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 880524,
      "author_name": "pratikbarua",
      "author_url": "",
      "post_date": "06/10/2020 11:19:04",
      "content": "<p>Great!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 892405,
      "author_name": "yashchoudhary",
      "author_url": "",
      "post_date": "06/18/2020 21:11:41",
      "content": "<p>\"I did what almost all of the beginners, tried different models, optimizers, losses, augmentations\" I am one of those beginners😂 \nThanks a lot for this error analysis. I would use this approach for my experiments.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "880026": "In the previous competitions, I did not spend much time analyzing my model's performance. I did what almost all of the beginners, tried different models, optimizers, losses, augmentations; I tried to implement more advanced techniques like mixup.\n\nI more or less entirely ignored the error analysis. **That was a mistake!**\n\nFortunately, now I had a bit of free time, so I created my first [Error analysis notebook](https://www.kaggle.com/pestipeti/error-analysis). After I finished it, I instantly had lots of ideas on how and where I should improve my model.\n\n## Summary\n- Overall results (TP, FP, FN, sources, sizes, brightness, contrast)\n- Results by sources (overall, FN/FP by sources, normal/border by sources, etc.)\n- Results by size (small, medium, large; you can configure the sizes)\n- Normal and \"border\" results\n- Noisy targets\n\n## Model's error\nOur models (or at least my models) have to handle these issues:\n- Different size of wheat-heads (too small, too large)\n- Group of wheat-heads\n- Partially invisible heads (covered by another head or a leaf)\n- Wheat-heads at the borders\n- Noisy boxes\n\n\n------\n[You can find the notebook here](https://www.kaggle.com/pestipeti/error-analysis).",
    "880028": "Good work Peter and thanks for your contributions! I agree, looking at examples of predictions vs. ground truth is a good method to find ways to improve",
    "880524": "Great!",
    "892405": "\"I did what almost all of the beginners, tried different models, optimizers, losses, augmentations\" I am one of those beginners😂 \nThanks a lot for this error analysis. I would use this approach for my experiments."
  },
  "source": "meta"
}