{
  "id": 108048,
  "title": "187th Place Approach + Some Insights",
  "url": "/competitions/aptos2019-blindness-detection/writeups/cykagglers-187th-place-approach-some-insights",
  "author_name": "",
  "post_date": "2019-09-11T03:35:33.990Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Nothing to fancy for us, but I thought I'd share our approach.</p>\n\n<p>For our phase 2 kernels, we selected 2 different ensembles of 2 Effiecientnets: B5 (Image size 456) + B4 (Image size 328) with pretty standard augmentations and some TTA.\nWe also experimented with a few values for the GaussianBlur and ended up going with a value of 20 (empirically got slightly better results).</p>\n\n<p>We actually decided to only use the current competition data (no 2015 data). <strong>I wonder if there are any top solutions who did the same?</strong>\nSince we anticipated there will be a difference between the public and private test data, we also decided to not optimize the QWK coefficients to avoid the risk of overfitting to the public data.\n(we used coef = [0.5, 1.5, 2.5, 3.5])</p>\n\n<p>We relied on local validation QWK combined with validation loss and early stopping. We tried to select models that got low validation loss combined with a relatively low difference between the train and validation loss. This proved to be a good tactic as we jumped almost 100 places between the public and private LB.</p>\n\n<p><strong>What Didn't Work</strong>\nWe have tried using Green CLHE preprocessing but got no apparent improvement. We ended up not using it for any of our ensemble models. <strong>Did anyone have any luck with it?</strong></p>\n\n<p>This was our first Computer Vision contest so we are happy with the results. Thanks for everyone involved with the discussions and awesome kernels! we learned a lot!</p>",
  "messages": [
    {
      "id": "621590",
      "postDate": "09/08/2019 17:25:31",
      "content": "<p>Nothing to fancy for us, but I thought I'd share our approach.</p>\n\n<p>For our phase 2 kernels, we selected 2 different ensembles of 2 Effiecientnets: B5 (Image size 456) + B4 (Image size 328) with pretty standard augmentations and some TTA.\nWe also experimented with a few values for the GaussianBlur and ended up going with a value of 20 (empirically got slightly better results).</p>\n\n<p>We actually decided to only use the current competition data (no 2015 data). <strong>I wonder if there are any top solutions who did the same?</strong>\nSince we anticipated there will be a difference between the public and private test data, we also decided to not optimize the QWK coefficients to avoid the risk of overfitting to the public data.\n(we used coef = [0.5, 1.5, 2.5, 3.5])</p>\n\n<p>We relied on local validation QWK combined with validation loss and early stopping. We tried to select models that got low validation loss combined with a relatively low difference between the train and validation loss. This proved to be a good tactic as we jumped almost 100 places between the public and private LB.</p>\n\n<p><strong>What Didn't Work</strong>\nWe have tried using Green CLHE preprocessing but got no apparent improvement. We ended up not using it for any of our ensemble models. <strong>Did anyone have any luck with it?</strong></p>\n\n<p>This was our first Computer Vision contest so we are happy with the results. Thanks for everyone involved with the discussions and awesome kernels! we learned a lot!</p>",
      "rawMarkdown": "Nothing to fancy for us, but I thought I'd share our approach.\n\nFor our phase 2 kernels, we selected 2 different ensembles of 2 Effiecientnets: B5 (Image size 456) + B4 (Image size 328) with pretty standard augmentations and some TTA.\nWe also experimented with a few values for the GaussianBlur and ended up going with a value of 20 (empirically got slightly better results).\n\nWe actually decided to only use the current competition data (no 2015 data). **I wonder if there are any top solutions who did the same?**\nSince we anticipated there will be a difference between the public and private test data, we also decided to not optimize the QWK coefficients to avoid the risk of overfitting to the public data.\n(we used coef = [0.5, 1.5, 2.5, 3.5])\n\nWe relied on local validation QWK combined with validation loss and early stopping. We tried to select models that got low validation loss combined with a relatively low difference between the train and validation loss. This proved to be a good tactic as we jumped almost 100 places between the public and private LB.\n\n**What Didn't Work**\nWe have tried using Green CLHE preprocessing but got no apparent improvement. We ended up not using it for any of our ensemble models. **Did anyone have any luck with it?**\n\nThis was our first Computer Vision contest so we are happy with the results. Thanks for everyone involved with the discussions and awesome kernels! we learned a lot!",
      "votes": null
    },
    {
      "id": "621900",
      "postDate": "09/09/2019 04:45:40",
      "content": "<p>Congratulations...\nGreat Work...\nThanks for Sharing your Approach &amp; Insights... !! <a href=\"/pyotam\">@pyotam</a> </p>",
      "rawMarkdown": "Congratulations...\nGreat Work...\nThanks for Sharing your Approach &amp; Insights... !! @pyotam",
      "votes": null
    },
    {
      "id": "621987",
      "postDate": "09/09/2019 06:20:40",
      "content": "<p>Thanks <a href=\"/veeralakrishna\">@veeralakrishna</a>!</p>",
      "rawMarkdown": "Thanks @veeralakrishna!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 621900,
      "author_name": "veeralakrishna",
      "author_url": "",
      "post_date": "09/09/2019 04:45:40",
      "content": "<p>Congratulations...\nGreat Work...\nThanks for Sharing your Approach &amp; Insights... !! <a href=\"/pyotam\">@pyotam</a> </p>",
      "votes": null,
      "replies": [
        {
          "id": 621987,
          "author_name": "pyotam",
          "author_url": "",
          "post_date": "09/09/2019 06:20:40",
          "content": "<p>Thanks <a href=\"/veeralakrishna\">@veeralakrishna</a>!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "621590": "Nothing to fancy for us, but I thought I'd share our approach.\n\nFor our phase 2 kernels, we selected 2 different ensembles of 2 Effiecientnets: B5 (Image size 456) + B4 (Image size 328) with pretty standard augmentations and some TTA.\nWe also experimented with a few values for the GaussianBlur and ended up going with a value of 20 (empirically got slightly better results).\n\nWe actually decided to only use the current competition data (no 2015 data). **I wonder if there are any top solutions who did the same?**\nSince we anticipated there will be a difference between the public and private test data, we also decided to not optimize the QWK coefficients to avoid the risk of overfitting to the public data.\n(we used coef = [0.5, 1.5, 2.5, 3.5])\n\nWe relied on local validation QWK combined with validation loss and early stopping. We tried to select models that got low validation loss combined with a relatively low difference between the train and validation loss. This proved to be a good tactic as we jumped almost 100 places between the public and private LB.\n\n**What Didn't Work**\nWe have tried using Green CLHE preprocessing but got no apparent improvement. We ended up not using it for any of our ensemble models. **Did anyone have any luck with it?**\n\nThis was our first Computer Vision contest so we are happy with the results. Thanks for everyone involved with the discussions and awesome kernels! we learned a lot!",
    "621900": "Congratulations...\nGreat Work...\nThanks for Sharing your Approach &amp; Insights... !! @pyotam",
    "621987": "Thanks @veeralakrishna!"
  },
  "source": "meta"
}