{
  "id": 329049,
  "title": "1st Place Solution",
  "url": "/competitions/sorghum-id-fgvc-9/writeups/deepblueai-1st-place-solution",
  "author_name": "",
  "post_date": "2022-06-04T13:23:08.329158800Z",
  "votes": 6,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Thanks to Kaggle and the hosting team for an interesting competition.</p>\n<h1>summary</h1>\n<p>My method is very similar to that of the third place(<a href=\"https://www.kaggle.com/competitions/sorghum-id-fgvc-9/discussion/328593\" target=\"_blank\">sorghum-id-fgvc-9/discussion/328593</a>).</p>\n<p>The main differences are as follows:</p>\n<ol>\n<li>I used convnext base network as backbone.</li>\n<li>I trained 5-folds models and ensembled  them.</li>\n<li>I combined this contest data with fgvc8data to generate a new dataset with 223 categories instead of 100.</li>\n</ol>\n<h1>results</h1>\n<table>\n<thead>\n<tr>\n<th>step</th>\n<th>method</th>\n<th>Public Score</th>\n<th>Private Score</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>1</td>\n<td>convb+100categories</td>\n<td>0.953</td>\n<td>0.952</td>\n</tr>\n<tr>\n<td>2</td>\n<td>convb+223categories</td>\n<td>0.955</td>\n<td>0.954</td>\n</tr>\n<tr>\n<td>3</td>\n<td>ensemble(step1+step2)</td>\n<td>0.959</td>\n<td>0.957</td>\n</tr>\n<tr>\n<td>4</td>\n<td>step3+pseudo label</td>\n<td>0.963</td>\n<td>0.962</td>\n</tr>\n<tr>\n<td>5</td>\n<td>step4+5folds</td>\n<td>0.965</td>\n<td>0.965</td>\n</tr>\n</tbody>\n</table>",
  "messages": [
    {
      "id": "1811229",
      "postDate": "06/04/2022 13:23:08",
      "content": "<p>Thanks to Kaggle and the hosting team for an interesting competition.</p>\n<h1>summary</h1>\n<p>My method is very similar to that of the third place(<a href=\"https://www.kaggle.com/competitions/sorghum-id-fgvc-9/discussion/328593\" target=\"_blank\">sorghum-id-fgvc-9/discussion/328593</a>).</p>\n<p>The main differences are as follows:</p>\n<ol>\n<li>I used convnext base network as backbone.</li>\n<li>I trained 5-folds models and ensembled  them.</li>\n<li>I combined this contest data with fgvc8data to generate a new dataset with 223 categories instead of 100.</li>\n</ol>\n<h1>results</h1>\n<table>\n<thead>\n<tr>\n<th>step</th>\n<th>method</th>\n<th>Public Score</th>\n<th>Private Score</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>1</td>\n<td>convb+100categories</td>\n<td>0.953</td>\n<td>0.952</td>\n</tr>\n<tr>\n<td>2</td>\n<td>convb+223categories</td>\n<td>0.955</td>\n<td>0.954</td>\n</tr>\n<tr>\n<td>3</td>\n<td>ensemble(step1+step2)</td>\n<td>0.959</td>\n<td>0.957</td>\n</tr>\n<tr>\n<td>4</td>\n<td>step3+pseudo label</td>\n<td>0.963</td>\n<td>0.962</td>\n</tr>\n<tr>\n<td>5</td>\n<td>step4+5folds</td>\n<td>0.965</td>\n<td>0.965</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "Thanks to Kaggle and the hosting team for an interesting competition.\n# summary\nMy method is very similar to that of the third place([sorghum-id-fgvc-9/discussion/328593](https://www.kaggle.com/competitions/sorghum-id-fgvc-9/discussion/328593)).\n\nThe main differences are as follows:\n1. I used convnext base network as backbone.\n2. I trained 5-folds models and ensembled  them.\n3. I combined this contest data with fgvc8data to generate a new dataset with 223 categories instead of 100.\n# results\n|  step|  method|Public Score|Private Score|\n| ---| --- | --- | --- |\n| 1| convb+100categories | 0.953 |0.952|\n| 2| convb+223categories | 0.955 |0.954|\n| 3| ensemble(step1+step2) | 0.959 |0.957|\n| 4|step3+pseudo label| 0.963 |0.962|\n| 5|step4+5folds| 0.965 |0.965|",
      "votes": null
    },
    {
      "id": "1811255",
      "postDate": "06/04/2022 13:49:12",
      "content": "<p>Did you also use histogram equalization, instance normalisation, arcface and the other tricks of that 3rd place solution?</p>",
      "rawMarkdown": "Did you also use histogram equalization, instance normalisation, arcface and the other tricks of that 3rd place solution?",
      "votes": null
    },
    {
      "id": "1811259",
      "postDate": "06/04/2022 13:58:17",
      "content": "<p>I didn’t use instance normalisation and AWP train.Data enhancement strategies are also slightly different.</p>",
      "rawMarkdown": "I didn’t use instance normalisation and AWP train.Data enhancement strategies are also slightly different.",
      "votes": null
    },
    {
      "id": "1813104",
      "postDate": "06/06/2022 14:47:41",
      "content": "<p>great, thanks for sharing :)</p>",
      "rawMarkdown": "great, thanks for sharing :)",
      "votes": null
    },
    {
      "id": "1821496",
      "postDate": "06/15/2022 15:32:49",
      "content": "<p>This is great! I'm finding that one of the most common elements across lots of different contests (I ran two this year) is the ensembling. I suppose that's not surprising, but it's interesting to me the extent to which the ensembles are the single most common element I see across many high performing approaches.</p>\n<p>I'm also interested in the fact that you combined data with last year's dataset! There was likely some overlap between the sorghum lines in the two contests, which means that some of your 223 categories were overlapping. Given that, I'm not shocked that that gave only a small improvement -- I wonder if you left out step 2 and 3 how things would have worked.</p>\n<p>Great job! (And my apologies for my delay in posting a follow up -- my whole extended family has come down with covid in the last few weeks and I've been basically entirely out of commission.)</p>",
      "rawMarkdown": "This is great! I'm finding that one of the most common elements across lots of different contests (I ran two this year) is the ensembling. I suppose that's not surprising, but it's interesting to me the extent to which the ensembles are the single most common element I see across many high performing approaches.\n\nI'm also interested in the fact that you combined data with last year's dataset! There was likely some overlap between the sorghum lines in the two contests, which means that some of your 223 categories were overlapping. Given that, I'm not shocked that that gave only a small improvement -- I wonder if you left out step 2 and 3 how things would have worked.\n\nGreat job! (And my apologies for my delay in posting a follow up -- my whole extended family has come down with covid in the last few weeks and I've been basically entirely out of commission.)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1811255,
      "author_name": "tfriedel",
      "author_url": "",
      "post_date": "06/04/2022 13:49:12",
      "content": "<p>Did you also use histogram equalization, instance normalisation, arcface and the other tricks of that 3rd place solution?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1811259,
          "author_name": "lonelygeese",
          "author_url": "",
          "post_date": "06/04/2022 13:58:17",
          "content": "<p>I didn’t use instance normalisation and AWP train.Data enhancement strategies are also slightly different.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1813104,
      "author_name": "fajarwibowo",
      "author_url": "",
      "post_date": "06/06/2022 14:47:41",
      "content": "<p>great, thanks for sharing :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1821496,
      "author_name": "abbystylianou",
      "author_url": "",
      "post_date": "06/15/2022 15:32:49",
      "content": "<p>This is great! I'm finding that one of the most common elements across lots of different contests (I ran two this year) is the ensembling. I suppose that's not surprising, but it's interesting to me the extent to which the ensembles are the single most common element I see across many high performing approaches.</p>\n<p>I'm also interested in the fact that you combined data with last year's dataset! There was likely some overlap between the sorghum lines in the two contests, which means that some of your 223 categories were overlapping. Given that, I'm not shocked that that gave only a small improvement -- I wonder if you left out step 2 and 3 how things would have worked.</p>\n<p>Great job! (And my apologies for my delay in posting a follow up -- my whole extended family has come down with covid in the last few weeks and I've been basically entirely out of commission.)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1811229": "Thanks to Kaggle and the hosting team for an interesting competition.\n# summary\nMy method is very similar to that of the third place([sorghum-id-fgvc-9/discussion/328593](https://www.kaggle.com/competitions/sorghum-id-fgvc-9/discussion/328593)).\n\nThe main differences are as follows:\n1. I used convnext base network as backbone.\n2. I trained 5-folds models and ensembled  them.\n3. I combined this contest data with fgvc8data to generate a new dataset with 223 categories instead of 100.\n# results\n|  step|  method|Public Score|Private Score|\n| ---| --- | --- | --- |\n| 1| convb+100categories | 0.953 |0.952|\n| 2| convb+223categories | 0.955 |0.954|\n| 3| ensemble(step1+step2) | 0.959 |0.957|\n| 4|step3+pseudo label| 0.963 |0.962|\n| 5|step4+5folds| 0.965 |0.965|",
    "1811255": "Did you also use histogram equalization, instance normalisation, arcface and the other tricks of that 3rd place solution?",
    "1811259": "I didn’t use instance normalisation and AWP train.Data enhancement strategies are also slightly different.",
    "1813104": "great, thanks for sharing :)",
    "1821496": "This is great! I'm finding that one of the most common elements across lots of different contests (I ran two this year) is the ensembling. I suppose that's not surprising, but it's interesting to me the extent to which the ensembles are the single most common element I see across many high performing approaches.\n\nI'm also interested in the fact that you combined data with last year's dataset! There was likely some overlap between the sorghum lines in the two contests, which means that some of your 223 categories were overlapping. Given that, I'm not shocked that that gave only a small improvement -- I wonder if you left out step 2 and 3 how things would have worked.\n\nGreat job! (And my apologies for my delay in posting a follow up -- my whole extended family has come down with covid in the last few weeks and I've been basically entirely out of commission.)"
  },
  "source": "meta"
}