{
  "id": 106043,
  "title": "Models other than Efficientnets that are working for this competition",
  "url": "/competitions/aptos2019-blindness-detection/discussion/106043",
  "author_name": "",
  "post_date": "2019-08-27T21:18:57.665382900Z",
  "votes": 3,
  "comment_count": 7,
  "views": 0,
  "content": "<p>We are using Efficientnets so far and having a hard time producing the same kind of results with other models like resnet, densenet. Any suggestions about which models to work with would be really helpful.</p>",
  "messages": [
    {
      "id": "609562",
      "postDate": "08/27/2019 21:18:57",
      "content": "<p>We are using Efficientnets so far and having a hard time producing the same kind of results with other models like resnet, densenet. Any suggestions about which models to work with would be really helpful.</p>",
      "rawMarkdown": "We are using Efficientnets so far and having a hard time producing the same kind of results with other models like resnet, densenet. Any suggestions about which models to work with would be really helpful.",
      "votes": null
    },
    {
      "id": "609631",
      "postDate": "08/28/2019 00:01:37",
      "content": "<p>in my case ig resnext101 works slightly better than efficientnet however harder to train</p>",
      "rawMarkdown": "in my case ig resnext101 works slightly better than efficientnet however harder to train",
      "votes": null
    },
    {
      "id": "609690",
      "postDate": "08/28/2019 02:19:37",
      "content": "<p>I use efficientnet B5,  I have no idea to impove my current score </p>",
      "rawMarkdown": "I use efficientnet B5,  I have no idea to impove my current score",
      "votes": null
    },
    {
      "id": "609704",
      "postDate": "08/28/2019 02:45:20",
      "content": "<p>Up until recently I haven't been able to get any model &gt; 0.8 except my trusty DenseNet201. I found that label smoothing was quite helpful as well setting different learning rates for different layers. </p>\n\n<p>Recently EfficientNet started working well for me (my team mate already has a solid EfficientNet ensemble). Seems to be mainly a case of getting the learning rate right.</p>",
      "rawMarkdown": "Up until recently I haven't been able to get any model &gt; 0.8 except my trusty DenseNet201. I found that label smoothing was quite helpful as well setting different learning rates for different layers. \n\nRecently EfficientNet started working well for me (my team mate already has a solid EfficientNet ensemble). Seems to be mainly a case of getting the learning rate right.",
      "votes": null
    },
    {
      "id": "609859",
      "postDate": "08/28/2019 07:14:00",
      "content": "<p>I am also having a hard time getting the learning rate right. </p>",
      "rawMarkdown": "I am also having a hard time getting the learning rate right.",
      "votes": null
    },
    {
      "id": "609860",
      "postDate": "08/28/2019 07:15:05",
      "content": "<p>Is it harder in terms of number of epochs to train or getting the correct learning rate??</p>",
      "rawMarkdown": "Is it harder in terms of number of epochs to train or getting the correct learning rate??",
      "votes": null
    },
    {
      "id": "611457",
      "postDate": "08/29/2019 09:16:28",
      "content": "<p>Try smaller models (like EffNet b2, b3). It can help you</p>",
      "rawMarkdown": "Try smaller models (like EffNet b2, b3). It can help you",
      "votes": null
    },
    {
      "id": "613145",
      "postDate": "08/30/2019 09:00:31",
      "content": "<p>I'm running out of CUDA memory all the time on Kaggle kernels while training DenseNet201. Lowering the batch size doesn't show any improvements over my previous models.</p>",
      "rawMarkdown": "I'm running out of CUDA memory all the time on Kaggle kernels while training DenseNet201. Lowering the batch size doesn't show any improvements over my previous models.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 609631,
      "author_name": "leixiang",
      "author_url": "",
      "post_date": "08/28/2019 00:01:37",
      "content": "<p>in my case ig resnext101 works slightly better than efficientnet however harder to train</p>",
      "votes": null,
      "replies": [
        {
          "id": 609860,
          "author_name": "sabbiracoustic1006",
          "author_url": "",
          "post_date": "08/28/2019 07:15:05",
          "content": "<p>Is it harder in terms of number of epochs to train or getting the correct learning rate??</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 609690,
      "author_name": "jiangkun2",
      "author_url": "",
      "post_date": "08/28/2019 02:19:37",
      "content": "<p>I use efficientnet B5,  I have no idea to impove my current score </p>",
      "votes": null,
      "replies": [
        {
          "id": 611457,
          "author_name": "mnikita",
          "author_url": "",
          "post_date": "08/29/2019 09:16:28",
          "content": "<p>Try smaller models (like EffNet b2, b3). It can help you</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 609704,
      "author_name": "lextoumbourou",
      "author_url": "",
      "post_date": "08/28/2019 02:45:20",
      "content": "<p>Up until recently I haven't been able to get any model &gt; 0.8 except my trusty DenseNet201. I found that label smoothing was quite helpful as well setting different learning rates for different layers. </p>\n\n<p>Recently EfficientNet started working well for me (my team mate already has a solid EfficientNet ensemble). Seems to be mainly a case of getting the learning rate right.</p>",
      "votes": null,
      "replies": [
        {
          "id": 609859,
          "author_name": "sabbiracoustic1006",
          "author_url": "",
          "post_date": "08/28/2019 07:14:00",
          "content": "<p>I am also having a hard time getting the learning rate right. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 613145,
          "author_name": "abhinand05",
          "author_url": "",
          "post_date": "08/30/2019 09:00:31",
          "content": "<p>I'm running out of CUDA memory all the time on Kaggle kernels while training DenseNet201. Lowering the batch size doesn't show any improvements over my previous models.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "609562": "We are using Efficientnets so far and having a hard time producing the same kind of results with other models like resnet, densenet. Any suggestions about which models to work with would be really helpful.",
    "609631": "in my case ig resnext101 works slightly better than efficientnet however harder to train",
    "609690": "I use efficientnet B5,  I have no idea to impove my current score",
    "609704": "Up until recently I haven't been able to get any model &gt; 0.8 except my trusty DenseNet201. I found that label smoothing was quite helpful as well setting different learning rates for different layers. \n\nRecently EfficientNet started working well for me (my team mate already has a solid EfficientNet ensemble). Seems to be mainly a case of getting the learning rate right.",
    "609859": "I am also having a hard time getting the learning rate right.",
    "609860": "Is it harder in terms of number of epochs to train or getting the correct learning rate??",
    "611457": "Try smaller models (like EffNet b2, b3). It can help you",
    "613145": "I'm running out of CUDA memory all the time on Kaggle kernels while training DenseNet201. Lowering the batch size doesn't show any improvements over my previous models."
  },
  "source": "meta"
}