{
  "id": 92271,
  "title": "Does anyone use fastai?",
  "url": "/competitions/imet-2019-fgvc6/discussion/92271",
  "author_name": "",
  "post_date": "2019-05-15T01:27:13.486589Z",
  "votes": 1,
  "comment_count": 11,
  "views": 0,
  "content": "<p>How can I use fastai to achieve 6.0 or above, can I use se_resnet in fastai?</p>",
  "messages": [
    {
      "id": "531472",
      "postDate": "05/15/2019 01:27:13",
      "content": "<p>How can I use fastai to achieve 6.0 or above, can I use se_resnet in fastai?</p>",
      "rawMarkdown": "How can I use fastai to achieve 6.0 or above, can I use se_resnet in fastai?",
      "votes": null
    },
    {
      "id": "531535",
      "postDate": "05/15/2019 04:52:32",
      "content": "<p><a href=\"/guitar123\">@guitar123</a> yes you can use se_resnet, se_resnext, nasnet in fastai. FastAI has cadene_models here <code>https://github.com/fastai/fastai/blob/master/fastai/vision/models/cadene_models.py</code> you can use these. And for your other question I acheived 0.601 using a single model in fastai.</p>",
      "rawMarkdown": "guitar123 yes you can use se_resnet, se_resnext, nasnet in fastai. FastAI has cadene_models here `https://github.com/fastai/fastai/blob/master/fastai/vision/models/cadene_models.py` you can use these. And for your other question I acheived 0.601 using a single model in fastai.",
      "votes": null
    },
    {
      "id": "531551",
      "postDate": "05/15/2019 05:53:01",
      "content": "<p>I've used se_resnet, but I've found it takes a long time to train with it. Do you encounter this problem when you use Se_resnet in fastai?</p>",
      "rawMarkdown": "I've used se_resnet, but I've found it takes a long time to train with it. Do you encounter this problem when you use Se_resnet in fastai?",
      "votes": null
    },
    {
      "id": "531554",
      "postDate": "05/15/2019 05:59:01",
      "content": "<p>0.601 - is this 5 fold result or 1 fold? Thanks</p>",
      "rawMarkdown": "0.601 - is this 5 fold result or 1 fold? Thanks",
      "votes": null
    },
    {
      "id": "531580",
      "postDate": "05/15/2019 06:45:42",
      "content": "<p>Yes, \nWith Resnet50, i  get .604 with single model with 1 fold.\n[update] now .606 :)</p>",
      "rawMarkdown": "Yes, \nWith Resnet50, i  get .604 with single model with 1 fold.\n[update] now .606 :)",
      "votes": null
    },
    {
      "id": "531634",
      "postDate": "05/15/2019 08:46:34",
      "content": "<p>May I ask, what data augmentations did you use？ Can I use a different folder in fastai?</p>",
      "rawMarkdown": "May I ask, what data augmentations did you use？ Can I use a different folder in fastai?",
      "votes": null
    },
    {
      "id": "531690",
      "postDate": "05/15/2019 10:46:05",
      "content": "<p><a href=\"/demonplus\">@demonplus</a> It's a single fold</p>\n\n<p><a href=\"/guitar123\">@guitar123</a> yes it takes a longer time to train</p>",
      "rawMarkdown": "demonplus It's a single fold\n\n@guitar123 yes it takes a longer time to train",
      "votes": null
    },
    {
      "id": "531711",
      "postDate": "05/15/2019 11:40:41",
      "content": "<p>How did you finish running in 9 hours with se_resnet? May I ask，what loss function do you use?</p>",
      "rawMarkdown": "How did you finish running in 9 hours with se_resnet? May I ask，what loss function do you use?",
      "votes": null
    },
    {
      "id": "531822",
      "postDate": "05/15/2019 15:50:32",
      "content": "<p>It seems it is a great result for ResNet50. Did you use any tricks comparing to public kernels?</p>",
      "rawMarkdown": "It seems it is a great result for ResNet50. Did you use any tricks comparing to public kernels?",
      "votes": null
    },
    {
      "id": "531837",
      "postDate": "05/15/2019 16:10:54",
      "content": "<p>resnet50 single fold 0.607</p>",
      "rawMarkdown": "resnet50 single fold 0.607",
      "votes": null
    },
    {
      "id": "532291",
      "postDate": "05/16/2019 15:20:34",
      "content": "<p>I use H flip, max_rotate, max_zoom, max_warp, max_lighting parameters and approach proposed in the FastAI course v3.</p>",
      "rawMarkdown": "I use H flip, max_rotate, max_zoom, max_warp, max_lighting parameters and approach proposed in the FastAI course v3.",
      "votes": null
    },
    {
      "id": "532294",
      "postDate": "05/16/2019 15:25:20",
      "content": "<p>May i ask you the number of epoch you use ?</p>",
      "rawMarkdown": "May i ask you the number of epoch you use ?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 531535,
      "author_name": "axel81",
      "author_url": "",
      "post_date": "05/15/2019 04:52:32",
      "content": "<p><a href=\"/guitar123\">@guitar123</a> yes you can use se_resnet, se_resnext, nasnet in fastai. FastAI has cadene_models here <code>https://github.com/fastai/fastai/blob/master/fastai/vision/models/cadene_models.py</code> you can use these. And for your other question I acheived 0.601 using a single model in fastai.</p>",
      "votes": null,
      "replies": [
        {
          "id": 531551,
          "author_name": "guitar123",
          "author_url": "",
          "post_date": "05/15/2019 05:53:01",
          "content": "<p>I've used se_resnet, but I've found it takes a long time to train with it. Do you encounter this problem when you use Se_resnet in fastai?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 531554,
          "author_name": "demonplus",
          "author_url": "",
          "post_date": "05/15/2019 05:59:01",
          "content": "<p>0.601 - is this 5 fold result or 1 fold? Thanks</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 531690,
          "author_name": "axel81",
          "author_url": "",
          "post_date": "05/15/2019 10:46:05",
          "content": "<p><a href=\"/demonplus\">@demonplus</a> It's a single fold</p>\n\n<p><a href=\"/guitar123\">@guitar123</a> yes it takes a longer time to train</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 531711,
          "author_name": "guitar123",
          "author_url": "",
          "post_date": "05/15/2019 11:40:41",
          "content": "<p>How did you finish running in 9 hours with se_resnet? May I ask，what loss function do you use?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 531580,
      "author_name": "a45632",
      "author_url": "",
      "post_date": "05/15/2019 06:45:42",
      "content": "<p>Yes, \nWith Resnet50, i  get .604 with single model with 1 fold.\n[update] now .606 :)</p>",
      "votes": null,
      "replies": [
        {
          "id": 531634,
          "author_name": "guitar123",
          "author_url": "",
          "post_date": "05/15/2019 08:46:34",
          "content": "<p>May I ask, what data augmentations did you use？ Can I use a different folder in fastai?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 531822,
          "author_name": "demonplus",
          "author_url": "",
          "post_date": "05/15/2019 15:50:32",
          "content": "<p>It seems it is a great result for ResNet50. Did you use any tricks comparing to public kernels?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 532291,
          "author_name": "a45632",
          "author_url": "",
          "post_date": "05/16/2019 15:20:34",
          "content": "<p>I use H flip, max_rotate, max_zoom, max_warp, max_lighting parameters and approach proposed in the FastAI course v3.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 531837,
      "author_name": "projdev",
      "author_url": "",
      "post_date": "05/15/2019 16:10:54",
      "content": "<p>resnet50 single fold 0.607</p>",
      "votes": null,
      "replies": [
        {
          "id": 532294,
          "author_name": "a45632",
          "author_url": "",
          "post_date": "05/16/2019 15:25:20",
          "content": "<p>May i ask you the number of epoch you use ?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "531472": "How can I use fastai to achieve 6.0 or above, can I use se_resnet in fastai?",
    "531535": "guitar123 yes you can use se_resnet, se_resnext, nasnet in fastai. FastAI has cadene_models here `https://github.com/fastai/fastai/blob/master/fastai/vision/models/cadene_models.py` you can use these. And for your other question I acheived 0.601 using a single model in fastai.",
    "531551": "I've used se_resnet, but I've found it takes a long time to train with it. Do you encounter this problem when you use Se_resnet in fastai?",
    "531554": "0.601 - is this 5 fold result or 1 fold? Thanks",
    "531580": "Yes, \nWith Resnet50, i  get .604 with single model with 1 fold.\n[update] now .606 :)",
    "531634": "May I ask, what data augmentations did you use？ Can I use a different folder in fastai?",
    "531690": "demonplus It's a single fold\n\n@guitar123 yes it takes a longer time to train",
    "531711": "How did you finish running in 9 hours with se_resnet? May I ask，what loss function do you use?",
    "531822": "It seems it is a great result for ResNet50. Did you use any tricks comparing to public kernels?",
    "531837": "resnet50 single fold 0.607",
    "532291": "I use H flip, max_rotate, max_zoom, max_warp, max_lighting parameters and approach proposed in the FastAI course v3.",
    "532294": "May i ask you the number of epoch you use ?"
  },
  "source": "meta"
}