{
  "id": 201702,
  "title": "Tips to iterate quickly",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/201702",
  "author_name": "",
  "post_date": "2020-12-06T10:17:04.758248600Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>As most of you probably know we need to iterate as quickly as possible to try different experiments and get the most out of TPU/GPU plus we have 5 subs a day which is pretty good. My way of iterating quickly is using lighter models for example resnet18 instead of resnet50 densenet121 instead of 201,  EfficientNet0 instead of the heavier efficientnets, I break after training for one fold (I'm not really sure if this is a good to evaluate tbh) I also try to use the maximum of regular augs before using the heavier ones (cutmix, mixup ….) let me know what are your tips. </p>",
  "messages": [
    {
      "id": "1103812",
      "postDate": "12/06/2020 10:17:04",
      "content": "<p>As most of you probably know we need to iterate as quickly as possible to try different experiments and get the most out of TPU/GPU plus we have 5 subs a day which is pretty good. My way of iterating quickly is using lighter models for example resnet18 instead of resnet50 densenet121 instead of 201,  EfficientNet0 instead of the heavier efficientnets, I break after training for one fold (I'm not really sure if this is a good to evaluate tbh) I also try to use the maximum of regular augs before using the heavier ones (cutmix, mixup ….) let me know what are your tips. </p>",
      "rawMarkdown": "As most of you probably know we need to iterate as quickly as possible to try different experiments and get the most out of TPU/GPU plus we have 5 subs a day which is pretty good. My way of iterating quickly is using lighter models for example resnet18 instead of resnet50 densenet121 instead of 201,  EfficientNet0 instead of the heavier efficientnets, I break after training for one fold (I'm not really sure if this is a good to evaluate tbh) I also try to use the maximum of regular augs before using the heavier ones (cutmix, mixup ....) let me know what are your tips.",
      "votes": null
    },
    {
      "id": "1107286",
      "postDate": "12/09/2020 15:05:45",
      "content": "<p>I run smaller models, bigger batches, lesser epochs and faster auguments(or regular that would be implemented later). Then if results stay satisfied switch to bigger models.</p>",
      "rawMarkdown": "I run smaller models, bigger batches, lesser epochs and faster auguments(or regular that would be implemented later). Then if results stay satisfied switch to bigger models.",
      "votes": null
    },
    {
      "id": "1116816",
      "postDate": "12/17/2020 13:41:12",
      "content": "<p>What image size do you use and is it affecting the training time (mostly it should na?)</p>",
      "rawMarkdown": "What image size do you use and is it affecting the training time (mostly it should na?)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1107286,
      "author_name": "anku5hk",
      "author_url": "",
      "post_date": "12/09/2020 15:05:45",
      "content": "<p>I run smaller models, bigger batches, lesser epochs and faster auguments(or regular that would be implemented later). Then if results stay satisfied switch to bigger models.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1116816,
      "author_name": "harshsdw",
      "author_url": "",
      "post_date": "12/17/2020 13:41:12",
      "content": "<p>What image size do you use and is it affecting the training time (mostly it should na?)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1103812": "As most of you probably know we need to iterate as quickly as possible to try different experiments and get the most out of TPU/GPU plus we have 5 subs a day which is pretty good. My way of iterating quickly is using lighter models for example resnet18 instead of resnet50 densenet121 instead of 201,  EfficientNet0 instead of the heavier efficientnets, I break after training for one fold (I'm not really sure if this is a good to evaluate tbh) I also try to use the maximum of regular augs before using the heavier ones (cutmix, mixup ....) let me know what are your tips.",
    "1107286": "I run smaller models, bigger batches, lesser epochs and faster auguments(or regular that would be implemented later). Then if results stay satisfied switch to bigger models.",
    "1116816": "What image size do you use and is it affecting the training time (mostly it should na?)"
  },
  "source": "meta"
}