{
  "id": 149204,
  "title": "A repo with better structured version of my kernel",
  "url": "/competitions/global-wheat-detection/discussion/149204",
  "author_name": "",
  "post_date": "2020-05-07T09:49:16.751666300Z",
  "votes": 6,
  "comment_count": 3,
  "views": 0,
  "content": "<p>As you may have seen, I have published a kernel with training for this competition: <a href=\"https://www.kaggle.com/artgor/object-detection-with-pytorch-lightning\">https://www.kaggle.com/artgor/object-detection-with-pytorch-lightning</a></p>\n\n<p>But I think it is better to split code in separate scripts and configs, than to keep everything in one kernel. You can see a different version of the code in my repo here:\n<a href=\"https://github.com/Erlemar/wheat\">https://github.com/Erlemar/wheat</a></p>",
  "messages": [
    {
      "id": "836849",
      "postDate": "05/07/2020 09:49:16",
      "content": "<p>As you may have seen, I have published a kernel with training for this competition: <a href=\"https://www.kaggle.com/artgor/object-detection-with-pytorch-lightning\">https://www.kaggle.com/artgor/object-detection-with-pytorch-lightning</a></p>\n\n<p>But I think it is better to split code in separate scripts and configs, than to keep everything in one kernel. You can see a different version of the code in my repo here:\n<a href=\"https://github.com/Erlemar/wheat\">https://github.com/Erlemar/wheat</a></p>",
      "rawMarkdown": "As you may have seen, I have published a kernel with training for this competition: https://www.kaggle.com/artgor/object-detection-with-pytorch-lightning\n\nBut I think it is better to split code in separate scripts and configs, than to keep everything in one kernel. You can see a different version of the code in my repo here:\nhttps://github.com/Erlemar/wheat",
      "votes": null
    },
    {
      "id": "837061",
      "postDate": "05/07/2020 13:44:41",
      "content": "<p>Great Kernel. I have been hearing a lot lately about Pytorch-Lightning, was wondering if you could share your thoughts on Pytorch Lightning? </p>",
      "rawMarkdown": "Great Kernel. I have been hearing a lot lately about Pytorch-Lightning, was wondering if you could share your thoughts on Pytorch Lightning?",
      "votes": null
    },
    {
      "id": "837124",
      "postDate": "05/07/2020 14:38:05",
      "content": "<p>I have recently started using pytorch-lightning and l like it.</p>\n\n<p>I have used pure pytorch for a long time, but it is difficult to write all the code for myself. I have tried a couple of popular frameworks, but they usually require you to write a code using some special API, which is sometimes difficult to understand.</p>\n\n<p>I like pytorch-lightning because it lets you to write pytorch code.</p>\n\n<p>You can use a similar structure of main class for different tasks. As a result it is quite easy to understand code written in pytorch-lightning.</p>\n\n<p>It has most of the commonly used features - distributed training, gradient accumulation, various logging and so on.</p>\n\n<p>Adding hydra configs also allows to easily switch architectures, datasets, losses and hyperparameters.</p>\n\n<p>In fact, I had written a code in pytorch-lightning and hydra for M5 competition and it took me ~1 day to modify it for this wheat competition. And most of the time was spent on understanding how coco api works.</p>",
      "rawMarkdown": "I have recently started using pytorch-lightning and l like it.\n\nI have used pure pytorch for a long time, but it is difficult to write all the code for myself. I have tried a couple of popular frameworks, but they usually require you to write a code using some special API, which is sometimes difficult to understand.\n\nI like pytorch-lightning because it lets you to write pytorch code.\n\nYou can use a similar structure of main class for different tasks. As a result it is quite easy to understand code written in pytorch-lightning.\n\nIt has most of the commonly used features - distributed training, gradient accumulation, various logging and so on.\n\nAdding hydra configs also allows to easily switch architectures, datasets, losses and hyperparameters.\n\nIn fact, I had written a code in pytorch-lightning and hydra for M5 competition and it took me ~1 day to modify it for this wheat competition. And most of the time was spent on understanding how coco api works.",
      "votes": null
    },
    {
      "id": "837127",
      "postDate": "05/07/2020 14:40:49",
      "content": "<p>Thanks for your replay! I guess its time to try! =) </p>",
      "rawMarkdown": "Thanks for your replay! I guess its time to try! =)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 837061,
      "author_name": "drhabib",
      "author_url": "",
      "post_date": "05/07/2020 13:44:41",
      "content": "<p>Great Kernel. I have been hearing a lot lately about Pytorch-Lightning, was wondering if you could share your thoughts on Pytorch Lightning? </p>",
      "votes": null,
      "replies": [
        {
          "id": 837124,
          "author_name": "artgor",
          "author_url": "",
          "post_date": "05/07/2020 14:38:05",
          "content": "<p>I have recently started using pytorch-lightning and l like it.</p>\n\n<p>I have used pure pytorch for a long time, but it is difficult to write all the code for myself. I have tried a couple of popular frameworks, but they usually require you to write a code using some special API, which is sometimes difficult to understand.</p>\n\n<p>I like pytorch-lightning because it lets you to write pytorch code.</p>\n\n<p>You can use a similar structure of main class for different tasks. As a result it is quite easy to understand code written in pytorch-lightning.</p>\n\n<p>It has most of the commonly used features - distributed training, gradient accumulation, various logging and so on.</p>\n\n<p>Adding hydra configs also allows to easily switch architectures, datasets, losses and hyperparameters.</p>\n\n<p>In fact, I had written a code in pytorch-lightning and hydra for M5 competition and it took me ~1 day to modify it for this wheat competition. And most of the time was spent on understanding how coco api works.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 837127,
          "author_name": "drhabib",
          "author_url": "",
          "post_date": "05/07/2020 14:40:49",
          "content": "<p>Thanks for your replay! I guess its time to try! =) </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "836849": "As you may have seen, I have published a kernel with training for this competition: https://www.kaggle.com/artgor/object-detection-with-pytorch-lightning\n\nBut I think it is better to split code in separate scripts and configs, than to keep everything in one kernel. You can see a different version of the code in my repo here:\nhttps://github.com/Erlemar/wheat",
    "837061": "Great Kernel. I have been hearing a lot lately about Pytorch-Lightning, was wondering if you could share your thoughts on Pytorch Lightning?",
    "837124": "I have recently started using pytorch-lightning and l like it.\n\nI have used pure pytorch for a long time, but it is difficult to write all the code for myself. I have tried a couple of popular frameworks, but they usually require you to write a code using some special API, which is sometimes difficult to understand.\n\nI like pytorch-lightning because it lets you to write pytorch code.\n\nYou can use a similar structure of main class for different tasks. As a result it is quite easy to understand code written in pytorch-lightning.\n\nIt has most of the commonly used features - distributed training, gradient accumulation, various logging and so on.\n\nAdding hydra configs also allows to easily switch architectures, datasets, losses and hyperparameters.\n\nIn fact, I had written a code in pytorch-lightning and hydra for M5 competition and it took me ~1 day to modify it for this wheat competition. And most of the time was spent on understanding how coco api works.",
    "837127": "Thanks for your replay! I guess its time to try! =)"
  },
  "source": "meta"
}