{
  "id": 206423,
  "title": "Score is not Crossing 0.87 after many tries",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/206423",
  "author_name": "",
  "post_date": "2020-12-24T14:40:06.988404400Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I have tried  Augmentation, changing the network, epochs, callbacks. Even after so many tries my score is not crossing 0.87. Can anyone help me with this?</p>",
  "messages": [
    {
      "id": "1125252",
      "postDate": "12/24/2020 14:40:06",
      "content": "<p>I have tried  Augmentation, changing the network, epochs, callbacks. Even after so many tries my score is not crossing 0.87. Can anyone help me with this?</p>",
      "rawMarkdown": "I have tried  Augmentation, changing the network, epochs, callbacks. Even after so many tries my score is not crossing 0.87. Can anyone help me with this?",
      "votes": null
    },
    {
      "id": "1125534",
      "postDate": "12/24/2020 19:42:22",
      "content": "<p>Single efficient net b3-b4 (5 fold) with augmentations like transpose, flip, cutoff passed 0.89 in my case. You can find an example <a href=\"https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug\" target=\"_blank\">here</a>. And if you train 1 or 2 network with different backbones like resnet and resnext you can ensemble the results (5 fold again). This will probably push you a little bit forward.</p>",
      "rawMarkdown": "Single efficient net b3-b4 (5 fold) with augmentations like transpose, flip, cutoff passed 0.89 in my case. You can find an example [here](https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug). And if you train 1 or 2 network with different backbones like resnet and resnext you can ensemble the results (5 fold again). This will probably push you a little bit forward.",
      "votes": null
    },
    {
      "id": "1125538",
      "postDate": "12/24/2020 19:44:27",
      "content": "<p>And of course you can add TTA to all of this networks. This will probably give additional points on the leaderboard.</p>",
      "rawMarkdown": "And of course you can add TTA to all of this networks. This will probably give additional points on the leaderboard.",
      "votes": null
    },
    {
      "id": "1125634",
      "postDate": "12/24/2020 22:17:55",
      "content": "<p>Can you tell some things about your training (e.g. used model, loss function etc.)?</p>",
      "rawMarkdown": "Can you tell some things about your training (e.g. used model, loss function etc.)?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1125534,
      "author_name": "snnclsr",
      "author_url": "",
      "post_date": "12/24/2020 19:42:22",
      "content": "<p>Single efficient net b3-b4 (5 fold) with augmentations like transpose, flip, cutoff passed 0.89 in my case. You can find an example <a href=\"https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug\" target=\"_blank\">here</a>. And if you train 1 or 2 network with different backbones like resnet and resnext you can ensemble the results (5 fold again). This will probably push you a little bit forward.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1125538,
          "author_name": "snnclsr",
          "author_url": "",
          "post_date": "12/24/2020 19:44:27",
          "content": "<p>And of course you can add TTA to all of this networks. This will probably give additional points on the leaderboard.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1125634,
      "author_name": "aliabdin1",
      "author_url": "",
      "post_date": "12/24/2020 22:17:55",
      "content": "<p>Can you tell some things about your training (e.g. used model, loss function etc.)?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1125252": "I have tried  Augmentation, changing the network, epochs, callbacks. Even after so many tries my score is not crossing 0.87. Can anyone help me with this?",
    "1125534": "Single efficient net b3-b4 (5 fold) with augmentations like transpose, flip, cutoff passed 0.89 in my case. You can find an example [here](https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug). And if you train 1 or 2 network with different backbones like resnet and resnext you can ensemble the results (5 fold again). This will probably push you a little bit forward.",
    "1125538": "And of course you can add TTA to all of this networks. This will probably give additional points on the leaderboard.",
    "1125634": "Can you tell some things about your training (e.g. used model, loss function etc.)?"
  },
  "source": "meta"
}