{
  "id": 176406,
  "title": "What worked in this competition",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/176406",
  "author_name": "",
  "post_date": "2020-08-21T16:01:19.956129500Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Spent quite a time in this competition. Taken this opportunity to learn TF2 and TFP. <br>\nI have tried many things, none of its seems to beat Benchmark established by Chris. I like to list down them<br>\n1) Hair Augmentation using custom TF function using tf.math.acos<br>\n2) Tried to improve CV by metric learning like ArcFace, Triplet loss but no improvements<br>\n3) Custom loss using rank + focal loss<br>\n4) Custom augmentation using different size circles</p>\n<p>Gone though some solution and I think following worked:<br>\n1) SWA<br>\n2) Cross entropy loss instead of BCE<br>\n3) Heavy Augmentation<br>\n4) Bit of luck</p>\n<p>Learned a lot in this competition but expecting some good tricks in the solution shared LB toppers.</p>",
  "messages": [
    {
      "id": "980462",
      "postDate": "08/21/2020 16:01:19",
      "content": "<p>Spent quite a time in this competition. Taken this opportunity to learn TF2 and TFP. <br>\nI have tried many things, none of its seems to beat Benchmark established by Chris. I like to list down them<br>\n1) Hair Augmentation using custom TF function using tf.math.acos<br>\n2) Tried to improve CV by metric learning like ArcFace, Triplet loss but no improvements<br>\n3) Custom loss using rank + focal loss<br>\n4) Custom augmentation using different size circles</p>\n<p>Gone though some solution and I think following worked:<br>\n1) SWA<br>\n2) Cross entropy loss instead of BCE<br>\n3) Heavy Augmentation<br>\n4) Bit of luck</p>\n<p>Learned a lot in this competition but expecting some good tricks in the solution shared LB toppers.</p>",
      "rawMarkdown": "Spent quite a time in this competition. Taken this opportunity to learn TF2 and TFP. \nI have tried many things, none of its seems to beat Benchmark established by Chris. I like to list down them\n1) Hair Augmentation using custom TF function using tf.math.acos\n2) Tried to improve CV by metric learning like ArcFace, Triplet loss but no improvements\n3) Custom loss using rank + focal loss\n4) Custom augmentation using different size circles\n\nGone though some solution and I think following worked:\n1) SWA\n2) Cross entropy loss instead of BCE\n3) Heavy Augmentation\n4) Bit of luck\n\nLearned a lot in this competition but expecting some good tricks in the solution shared LB toppers.",
      "votes": null
    },
    {
      "id": "981165",
      "postDate": "08/22/2020 08:29:34",
      "content": "<p>TTA helps out a lot. <br>\nLR scheduler also helps. <br>\nConcatenating model (parallely) improves the results<br>\nEnsembling good models gets you medal</p>",
      "rawMarkdown": "TTA helps out a lot. \nLR scheduler also helps. \nConcatenating model (parallely) improves the results\nEnsembling good models gets you medal",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 981165,
      "author_name": "dtprksh50",
      "author_url": "",
      "post_date": "08/22/2020 08:29:34",
      "content": "<p>TTA helps out a lot. <br>\nLR scheduler also helps. <br>\nConcatenating model (parallely) improves the results<br>\nEnsembling good models gets you medal</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "980462": "Spent quite a time in this competition. Taken this opportunity to learn TF2 and TFP. \nI have tried many things, none of its seems to beat Benchmark established by Chris. I like to list down them\n1) Hair Augmentation using custom TF function using tf.math.acos\n2) Tried to improve CV by metric learning like ArcFace, Triplet loss but no improvements\n3) Custom loss using rank + focal loss\n4) Custom augmentation using different size circles\n\nGone though some solution and I think following worked:\n1) SWA\n2) Cross entropy loss instead of BCE\n3) Heavy Augmentation\n4) Bit of luck\n\nLearned a lot in this competition but expecting some good tricks in the solution shared LB toppers.",
    "981165": "TTA helps out a lot. \nLR scheduler also helps. \nConcatenating model (parallely) improves the results\nEnsembling good models gets you medal"
  },
  "source": "meta"
}