{
  "id": 205415,
  "title": "Things that I have tried out with a score of 0.834",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/205415",
  "author_name": "",
  "post_date": "2020-12-20T04:55:08.204961700Z",
  "votes": 1,
  "comment_count": 6,
  "views": 0,
  "content": "<ol>\n<li>Data Augmentations - just random horizontal flips. </li>\n<li>Pretrained VGG16  - which is fine tuned after initial training on top layers added by me</li>\n<li>Using the custom Gambler's loss function as suggested in <a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017\" target=\"_blank\">here</a> by <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> along with 0.2 label smoothing.</li>\n<li>Validation on a test split of 0.2</li>\n<li>Have used tensorflow library.</li>\n<li>Have resized the images to (512,512,3)</li>\n</ol>\n<p>Apart from this I am also thinking of experimenting with other pre-trained models like Xception and Efficientnet<br>\nI have not seen any improvements with the increase of augmentations</p>\n<p>Would love to get know any suggestions that I should work on for improvements</p>\n<p>Thanks,</p>",
  "messages": [
    {
      "id": "1119456",
      "postDate": "12/20/2020 04:55:08",
      "content": "<ol>\n<li>Data Augmentations - just random horizontal flips. </li>\n<li>Pretrained VGG16  - which is fine tuned after initial training on top layers added by me</li>\n<li>Using the custom Gambler's loss function as suggested in <a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017\" target=\"_blank\">here</a> by <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> along with 0.2 label smoothing.</li>\n<li>Validation on a test split of 0.2</li>\n<li>Have used tensorflow library.</li>\n<li>Have resized the images to (512,512,3)</li>\n</ol>\n<p>Apart from this I am also thinking of experimenting with other pre-trained models like Xception and Efficientnet<br>\nI have not seen any improvements with the increase of augmentations</p>\n<p>Would love to get know any suggestions that I should work on for improvements</p>\n<p>Thanks,</p>",
      "rawMarkdown": "1.  Data Augmentations - just random horizontal flips. \n2. Pretrained VGG16  - which is fine tuned after initial training on top layers added by me\n3. Using the custom Gambler's loss function as suggested in [here](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017) by @hengck23 along with 0.2 label smoothing.\n4. Validation on a test split of 0.2\n5. Have used tensorflow library.\n6. Have resized the images to (512,512,3)\n\nApart from this I am also thinking of experimenting with other pre-trained models like Xception and Efficientnet\nI have not seen any improvements with the increase of augmentations\n\nWould love to get know any suggestions that I should work on for improvements\n\nThanks,",
      "votes": null
    },
    {
      "id": "1119614",
      "postDate": "12/20/2020 08:31:12",
      "content": "<ol>\n<li>Try more augmentations</li>\n<li>Try a model from resnet/resnext/efficientnet family</li>\n<li>If you have not used a scheduler use one</li>\n<li>Try K-Fold</li>\n<li>Try TTA </li>\n</ol>\n<p>It should help you to reach 0.9. </p>",
      "rawMarkdown": "1. Try more augmentations\n2. Try a model from resnet/resnext/efficientnet family\n3. If you have not used a scheduler use one\n4. Try K-Fold\n5. Try TTA \n\nIt should help you to reach 0.9.",
      "votes": null
    },
    {
      "id": "1119958",
      "postDate": "12/20/2020 13:39:26",
      "content": "<p>I think you must train the entire network and a much deeper network like DenseNet/EfficientNet, I can confirm the finetuning by freezing some of the layers doesn't works for some reason!</p>",
      "rawMarkdown": "I think you must train the entire network and a much deeper network like DenseNet/EfficientNet, I can confirm the finetuning by freezing some of the layers doesn't works for some reason!",
      "votes": null
    },
    {
      "id": "1120144",
      "postDate": "12/20/2020 15:47:02",
      "content": "<p>This is a revelation!! Thank you</p>",
      "rawMarkdown": "This is a revelation!! Thank you",
      "votes": null
    },
    {
      "id": "1120688",
      "postDate": "12/21/2020 03:54:23",
      "content": "<p>Try Tempered loss function as suggested by the below link-<br>\n<a href=\"https://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html\" target=\"_blank\">https://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html</a></p>",
      "rawMarkdown": "Try Tempered loss function as suggested by the below link-\nhttps://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html",
      "votes": null
    },
    {
      "id": "1120754",
      "postDate": "12/21/2020 05:05:04",
      "content": "<p>I tried it out with the current configuration but got maximum performance in case of gambler's loss. Though will make sure to experiment with it as well going forward</p>",
      "rawMarkdown": "I tried it out with the current configuration but got maximum performance in case of gambler's loss. Though will make sure to experiment with it as well going forward",
      "votes": null
    },
    {
      "id": "1120994",
      "postDate": "12/21/2020 09:22:00",
      "content": "<p>With just a few basic modifications you can reach a 0.898+ score <br>\nWhat I suggest for you and other persons which are struggling under this score is:</p>\n<ol>\n<li>Don't use VGG. Go either with Efficientnet or a high resnext</li>\n<li>You can get very good results with a simple BCE loss</li>\n<li>Use random resize crop. It helps</li>\n<li>Do TTA</li>\n</ol>",
      "rawMarkdown": "With just a few basic modifications you can reach a 0.898+ score \nWhat I suggest for you and other persons which are struggling under this score is:\n1. Don't use VGG. Go either with Efficientnet or a high resnext\n2. You can get very good results with a simple BCE loss\n3. Use random resize crop. It helps\n4. Do TTA",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1119614,
      "author_name": "vishnus",
      "author_url": "",
      "post_date": "12/20/2020 08:31:12",
      "content": "<ol>\n<li>Try more augmentations</li>\n<li>Try a model from resnet/resnext/efficientnet family</li>\n<li>If you have not used a scheduler use one</li>\n<li>Try K-Fold</li>\n<li>Try TTA </li>\n</ol>\n<p>It should help you to reach 0.9. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1119958,
      "author_name": "harveenchadha",
      "author_url": "",
      "post_date": "12/20/2020 13:39:26",
      "content": "<p>I think you must train the entire network and a much deeper network like DenseNet/EfficientNet, I can confirm the finetuning by freezing some of the layers doesn't works for some reason!</p>",
      "votes": null,
      "replies": [
        {
          "id": 1120144,
          "author_name": "kabhinay",
          "author_url": "",
          "post_date": "12/20/2020 15:47:02",
          "content": "<p>This is a revelation!! Thank you</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1120688,
      "author_name": "harininarasimhan",
      "author_url": "",
      "post_date": "12/21/2020 03:54:23",
      "content": "<p>Try Tempered loss function as suggested by the below link-<br>\n<a href=\"https://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html\" target=\"_blank\">https://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 1120754,
          "author_name": "kabhinay",
          "author_url": "",
          "post_date": "12/21/2020 05:05:04",
          "content": "<p>I tried it out with the current configuration but got maximum performance in case of gambler's loss. Though will make sure to experiment with it as well going forward</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1120994,
      "author_name": "vladvdv",
      "author_url": "",
      "post_date": "12/21/2020 09:22:00",
      "content": "<p>With just a few basic modifications you can reach a 0.898+ score <br>\nWhat I suggest for you and other persons which are struggling under this score is:</p>\n<ol>\n<li>Don't use VGG. Go either with Efficientnet or a high resnext</li>\n<li>You can get very good results with a simple BCE loss</li>\n<li>Use random resize crop. It helps</li>\n<li>Do TTA</li>\n</ol>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1119456": "1.  Data Augmentations - just random horizontal flips. \n2. Pretrained VGG16  - which is fine tuned after initial training on top layers added by me\n3. Using the custom Gambler's loss function as suggested in [here](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017) by @hengck23 along with 0.2 label smoothing.\n4. Validation on a test split of 0.2\n5. Have used tensorflow library.\n6. Have resized the images to (512,512,3)\n\nApart from this I am also thinking of experimenting with other pre-trained models like Xception and Efficientnet\nI have not seen any improvements with the increase of augmentations\n\nWould love to get know any suggestions that I should work on for improvements\n\nThanks,",
    "1119614": "1. Try more augmentations\n2. Try a model from resnet/resnext/efficientnet family\n3. If you have not used a scheduler use one\n4. Try K-Fold\n5. Try TTA \n\nIt should help you to reach 0.9.",
    "1119958": "I think you must train the entire network and a much deeper network like DenseNet/EfficientNet, I can confirm the finetuning by freezing some of the layers doesn't works for some reason!",
    "1120144": "This is a revelation!! Thank you",
    "1120688": "Try Tempered loss function as suggested by the below link-\nhttps://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html",
    "1120754": "I tried it out with the current configuration but got maximum performance in case of gambler's loss. Though will make sure to experiment with it as well going forward",
    "1120994": "With just a few basic modifications you can reach a 0.898+ score \nWhat I suggest for you and other persons which are struggling under this score is:\n1. Don't use VGG. Go either with Efficientnet or a high resnext\n2. You can get very good results with a simple BCE loss\n3. Use random resize crop. It helps\n4. Do TTA"
  },
  "source": "meta"
}