{
  "id": 207568,
  "title": "Why Resnet50 get a higher accuracy than Resnet152?",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/207568",
  "author_name": "",
  "post_date": "2020-12-30T10:02:13.249740500Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I training my first model with Resnet50.  The accuracy is 0.87 <br>\nThen I keep everything the same, just change the model from Resnet50 to Resnet152, and expect to get higher accuracy. But the accuracy of Resnet152 is 0.85. </p>\n<p>In both trainings, I use adam optimizer and changed the learning rate from 0.001 to 0.0001 after 10 epochs.  I also tried to Resnet152 more epochs with a learning rate of 0.00001, but it does no help.</p>\n<p>Could anyone tell me why Resnet152 performs worse than Resnet50?</p>",
  "messages": [
    {
      "id": "1132318",
      "postDate": "12/30/2020 10:02:13",
      "content": "<p>I training my first model with Resnet50.  The accuracy is 0.87 <br>\nThen I keep everything the same, just change the model from Resnet50 to Resnet152, and expect to get higher accuracy. But the accuracy of Resnet152 is 0.85. </p>\n<p>In both trainings, I use adam optimizer and changed the learning rate from 0.001 to 0.0001 after 10 epochs.  I also tried to Resnet152 more epochs with a learning rate of 0.00001, but it does no help.</p>\n<p>Could anyone tell me why Resnet152 performs worse than Resnet50?</p>",
      "rawMarkdown": "I training my first model with Resnet50.  The accuracy is 0.87 \nThen I keep everything the same, just change the model from Resnet50 to Resnet152, and expect to get higher accuracy. But the accuracy of Resnet152 is 0.85. \n\nIn both trainings, I use adam optimizer and changed the learning rate from 0.001 to 0.0001 after 10 epochs.  I also tried to Resnet152 more epochs with a learning rate of 0.00001, but it does no help.\n\nCould anyone tell me why Resnet152 performs worse than Resnet50?",
      "votes": null
    },
    {
      "id": "1132356",
      "postDate": "12/30/2020 10:39:00",
      "content": "<p>Because Resnet152 is bigger model it could overfit your train data and generalize less than smaller models. So you should try some regularization tricks to check if it helps (additional 2019 data, augmentations, regularization itself)</p>",
      "rawMarkdown": "Because Resnet152 is bigger model it could overfit your train data and generalize less than smaller models. So you should try some regularization tricks to check if it helps (additional 2019 data, augmentations, regularization itself)",
      "votes": null
    },
    {
      "id": "1133460",
      "postDate": "12/31/2020 08:28:42",
      "content": "<p>There is also a possibility that resnet152 may be suitable for expressing more complex data structures than this project, and the data structure of this project may not be as complicated as imagined, so some regularization terms need to be added to suppress overfitting.</p>",
      "rawMarkdown": "There is also a possibility that resnet152 may be suitable for expressing more complex data structures than this project, and the data structure of this project may not be as complicated as imagined, so some regularization terms need to be added to suppress overfitting.",
      "votes": null
    },
    {
      "id": "1133490",
      "postDate": "12/31/2020 09:02:02",
      "content": "<p>Here are some links that should help</p>\n<ol>\n<li><a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/199579\" target=\"_blank\">External Data</a></li>\n<li><a href=\"https://www.kaggle.com/vpkprasanna/different-augmentations-on-albumentations\" target=\"_blank\">Augmentations Notebook</a></li>\n<li>Standard methods like Regularization, Dropout, BatchNorm should also help</li>\n</ol>",
      "rawMarkdown": "Here are some links that should help\n1. [External Data](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/199579)\n2. [Augmentations Notebook](https://www.kaggle.com/vpkprasanna/different-augmentations-on-albumentations)\n3. Standard methods like Regularization, Dropout, BatchNorm should also help",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1132356,
      "author_name": "volcanoflash",
      "author_url": "",
      "post_date": "12/30/2020 10:39:00",
      "content": "<p>Because Resnet152 is bigger model it could overfit your train data and generalize less than smaller models. So you should try some regularization tricks to check if it helps (additional 2019 data, augmentations, regularization itself)</p>",
      "votes": null,
      "replies": [
        {
          "id": 1133490,
          "author_name": "yerramvarun",
          "author_url": "",
          "post_date": "12/31/2020 09:02:02",
          "content": "<p>Here are some links that should help</p>\n<ol>\n<li><a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/199579\" target=\"_blank\">External Data</a></li>\n<li><a href=\"https://www.kaggle.com/vpkprasanna/different-augmentations-on-albumentations\" target=\"_blank\">Augmentations Notebook</a></li>\n<li>Standard methods like Regularization, Dropout, BatchNorm should also help</li>\n</ol>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1133460,
      "author_name": "zhangeng",
      "author_url": "",
      "post_date": "12/31/2020 08:28:42",
      "content": "<p>There is also a possibility that resnet152 may be suitable for expressing more complex data structures than this project, and the data structure of this project may not be as complicated as imagined, so some regularization terms need to be added to suppress overfitting.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1132318": "I training my first model with Resnet50.  The accuracy is 0.87 \nThen I keep everything the same, just change the model from Resnet50 to Resnet152, and expect to get higher accuracy. But the accuracy of Resnet152 is 0.85. \n\nIn both trainings, I use adam optimizer and changed the learning rate from 0.001 to 0.0001 after 10 epochs.  I also tried to Resnet152 more epochs with a learning rate of 0.00001, but it does no help.\n\nCould anyone tell me why Resnet152 performs worse than Resnet50?",
    "1132356": "Because Resnet152 is bigger model it could overfit your train data and generalize less than smaller models. So you should try some regularization tricks to check if it helps (additional 2019 data, augmentations, regularization itself)",
    "1133460": "There is also a possibility that resnet152 may be suitable for expressing more complex data structures than this project, and the data structure of this project may not be as complicated as imagined, so some regularization terms need to be added to suppress overfitting.",
    "1133490": "Here are some links that should help\n1. [External Data](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/199579)\n2. [Augmentations Notebook](https://www.kaggle.com/vpkprasanna/different-augmentations-on-albumentations)\n3. Standard methods like Regularization, Dropout, BatchNorm should also help"
  },
  "source": "meta"
}