{
  "id": 226709,
  "title": "What's the best way to experiment with architectures and different hyperparameters?",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/226709",
  "author_name": "",
  "post_date": "2021-03-17T11:47:32.478298700Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hello everyone!</p>\n<p>I am planning to use the <a href=\"https://github.com/qubvel/segmentation_models.pytorch\" target=\"_blank\">segmentation-models-pytorch</a> library to create good single models. As you all know that this has a variety of architectures with different encoder backbones.</p>\n<p>So, I wanted to know from the experienced folks here - </p>\n<ol>\n<li>How should I organize my experiments?</li>\n<li>How to play around with the hyperparameters like learning rate, optimizer, scheduler etc?</li>\n<li>How to get the intuition that a particular architecture with a particular backbone will never work irrespective of the values of hyperparameters? </li>\n</ol>",
  "messages": [
    {
      "id": "1242083",
      "postDate": "03/17/2021 11:47:32",
      "content": "<p>Hello everyone!</p>\n<p>I am planning to use the <a href=\"https://github.com/qubvel/segmentation_models.pytorch\" target=\"_blank\">segmentation-models-pytorch</a> library to create good single models. As you all know that this has a variety of architectures with different encoder backbones.</p>\n<p>So, I wanted to know from the experienced folks here - </p>\n<ol>\n<li>How should I organize my experiments?</li>\n<li>How to play around with the hyperparameters like learning rate, optimizer, scheduler etc?</li>\n<li>How to get the intuition that a particular architecture with a particular backbone will never work irrespective of the values of hyperparameters? </li>\n</ol>",
      "rawMarkdown": "Hello everyone!\n\nI am planning to use the [segmentation-models-pytorch](https://github.com/qubvel/segmentation_models.pytorch) library to create good single models. As you all know that this has a variety of architectures with different encoder backbones.\n\nSo, I wanted to know from the experienced folks here - \n1.  How should I organize my experiments?\n2. How to play around with the hyperparameters like learning rate, optimizer, scheduler etc?\n3. How to get the intuition that a particular architecture with a particular backbone will never work irrespective of the values of hyperparameters?",
      "votes": null
    },
    {
      "id": "1259260",
      "postDate": "04/01/2021 09:15:13",
      "content": "<p>Hi there! Have you looked into <a href=\"https://keras-team.github.io/keras-tuner/\" target=\"_blank\">keras-tuner</a>? You can test with a range of hyperparameters, node and layer sizes and it will give the best models out of your choices. </p>\n<p>You can look how to get started in this <a href=\"https://www.kaggle.com/codeaesthete/house-price-prediction-with-deep-neural-networks\" target=\"_blank\">notebook</a> if you dont want to go through the docs. The final model is created after tuning the paramters in keras-tuner.</p>",
      "rawMarkdown": "Hi there! Have you looked into [keras-tuner](https://keras-team.github.io/keras-tuner/)? You can test with a range of hyperparameters, node and layer sizes and it will give the best models out of your choices. \n\nYou can look how to get started in this [notebook](https://www.kaggle.com/codeaesthete/house-price-prediction-with-deep-neural-networks) if you dont want to go through the docs. The final model is created after tuning the paramters in keras-tuner.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1259260,
      "author_name": "codeaesthete",
      "author_url": "",
      "post_date": "04/01/2021 09:15:13",
      "content": "<p>Hi there! Have you looked into <a href=\"https://keras-team.github.io/keras-tuner/\" target=\"_blank\">keras-tuner</a>? You can test with a range of hyperparameters, node and layer sizes and it will give the best models out of your choices. </p>\n<p>You can look how to get started in this <a href=\"https://www.kaggle.com/codeaesthete/house-price-prediction-with-deep-neural-networks\" target=\"_blank\">notebook</a> if you dont want to go through the docs. The final model is created after tuning the paramters in keras-tuner.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1242083": "Hello everyone!\n\nI am planning to use the [segmentation-models-pytorch](https://github.com/qubvel/segmentation_models.pytorch) library to create good single models. As you all know that this has a variety of architectures with different encoder backbones.\n\nSo, I wanted to know from the experienced folks here - \n1.  How should I organize my experiments?\n2. How to play around with the hyperparameters like learning rate, optimizer, scheduler etc?\n3. How to get the intuition that a particular architecture with a particular backbone will never work irrespective of the values of hyperparameters?",
    "1259260": "Hi there! Have you looked into [keras-tuner](https://keras-team.github.io/keras-tuner/)? You can test with a range of hyperparameters, node and layer sizes and it will give the best models out of your choices. \n\nYou can look how to get started in this [notebook](https://www.kaggle.com/codeaesthete/house-price-prediction-with-deep-neural-networks) if you dont want to go through the docs. The final model is created after tuning the paramters in keras-tuner."
  },
  "source": "meta"
}