{
  "id": 157812,
  "title": "3rd place solution",
  "url": "/competitions/herbarium-2020-fgvc7/discussion/157812",
  "author_name": "Ballins",
  "post_date": "2020-06-12T06:17:58.803000",
  "votes": 6,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Sorry for the delay, I'll briefly describe my solution. </p>\n\n<h1><strong>Methods</strong></h1>\n\n<p>I used lip-r50 and seresnext101 as the backbones, and 448 * 448 resolution for training. To save resources, I used 352 * 352 resolution to train seresnext101 and fintune with 448 *  448 resolution in last few epochs. For both of two backbones, I trained for 120 epoch in cosine decay, using labelsmooth and mixup(balanced sampling). I chose the standard cross entropy loss as the loss function.\nDuring testing, 576 * 576 resolution and tencrop were used. Finally, I ensembled the two models above to get the final softmax score(seresnext101_softmaxscore * 0.6 + lip-r50_softmax_score * 0.4).Just like other teams, rescore was used to refine the  prediction, becase each category has no more than 10 samples in the test set, so I sorted the samples in test set by conficence score, for those samples whose top-1 category already has 10 samples, I changed them to top-2 category instead, and resort the score. And we got the final score.</p>\n\n<p>I also spent a lot time training some offline classifiers(such as knn and svm) using cnn features, but got little improvement.</p>",
  "messages": [
    {
      "id": 882790,
      "postDate": "2020-06-12T06:17:58.803Z",
      "content": "<p>Sorry for the delay, I'll briefly describe my solution. </p>\n\n<h1><strong>Methods</strong></h1>\n\n<p>I used lip-r50 and seresnext101 as the backbones, and 448 * 448 resolution for training. To save resources, I used 352 * 352 resolution to train seresnext101 and fintune with 448 *  448 resolution in last few epochs. For both of two backbones, I trained for 120 epoch in cosine decay, using labelsmooth and mixup(balanced sampling). I chose the standard cross entropy loss as the loss function.\nDuring testing, 576 * 576 resolution and tencrop were used. Finally, I ensembled the two models above to get the final softmax score(seresnext101_softmaxscore * 0.6 + lip-r50_softmax_score * 0.4).Just like other teams, rescore was used to refine the  prediction, becase each category has no more than 10 samples in the test set, so I sorted the samples in test set by conficence score, for those samples whose top-1 category already has 10 samples, I changed them to top-2 category instead, and resort the score. And we got the final score.</p>\n\n<p>I also spent a lot time training some offline classifiers(such as knn and svm) using cnn features, but got little improvement.</p>",
      "rawMarkdown": "Sorry for the delay, I'll briefly describe my solution. \n# **Methods**\nI used lip-r50 and seresnext101 as the backbones, and 448 * 448 resolution for training. To save resources, I used 352 * 352 resolution to train seresnext101 and fintune with 448 *  448 resolution in last few epochs. For both of two backbones, I trained for 120 epoch in cosine decay, using labelsmooth and mixup(balanced sampling). I chose the standard cross entropy loss as the loss function.\nDuring testing, 576 * 576 resolution and tencrop were used. Finally, I ensembled the two models above to get the final softmax score(seresnext101_softmaxscore * 0.6 + lip-r50_softmax_score * 0.4).Just like other teams, rescore was used to refine the  prediction, becase each category has no more than 10 samples in the test set, so I sorted the samples in test set by conficence score, for those samples whose top-1 category already has 10 samples, I changed them to top-2 category instead, and resort the score. And we got the final score.\n\nI also spent a lot time training some offline classifiers(such as knn and svm) using cnn features, but got little improvement.",
      "votes": 6
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "882790": "Sorry for the delay, I'll briefly describe my solution. \n# **Methods**\nI used lip-r50 and seresnext101 as the backbones, and 448 * 448 resolution for training. To save resources, I used 352 * 352 resolution to train seresnext101 and fintune with 448 *  448 resolution in last few epochs. For both of two backbones, I trained for 120 epoch in cosine decay, using labelsmooth and mixup(balanced sampling). I chose the standard cross entropy loss as the loss function.\nDuring testing, 576 * 576 resolution and tencrop were used. Finally, I ensembled the two models above to get the final softmax score(seresnext101_softmaxscore * 0.6 + lip-r50_softmax_score * 0.4).Just like other teams, rescore was used to refine the  prediction, becase each category has no more than 10 samples in the test set, so I sorted the samples in test set by conficence score, for those samples whose top-1 category already has 10 samples, I changed them to top-2 category instead, and resort the score. And we got the final score.\n\nI also spent a lot time training some offline classifiers(such as knn and svm) using cnn features, but got little improvement."
  }
}