{
  "id": 90701,
  "title": "Methods for augmentation let's discuss",
  "url": "/competitions/imet-2019-fgvc6/discussion/90701",
  "author_name": "",
  "post_date": "2019-04-26T04:21:41.704403900Z",
  "votes": 18,
  "comment_count": 23,
  "views": 0,
  "content": "<p>I am trying to do simple augmentations like flip, rotate, zoom, wrap and using ResNet34 but wasn't able to get a score &gt; 0.580. What are the augmentation methods that I can use to improve ?</p>",
  "messages": [
    {
      "id": "523355",
      "postDate": "04/26/2019 04:21:41",
      "content": "<p>I am trying to do simple augmentations like flip, rotate, zoom, wrap and using ResNet34 but wasn't able to get a score &gt; 0.580. What are the augmentation methods that I can use to improve ?</p>",
      "rawMarkdown": "I am trying to do simple augmentations like flip, rotate, zoom, wrap and using ResNet34 but wasn't able to get a score &gt; 0.580. What are the augmentation methods that I can use to improve ?",
      "votes": null
    },
    {
      "id": "523433",
      "postDate": "04/26/2019 08:53:37",
      "content": "<p>nice topic to dig !</p>",
      "rawMarkdown": "nice topic to dig !",
      "votes": null
    },
    {
      "id": "523444",
      "postDate": "04/26/2019 09:30:41",
      "content": "<p>Things work for me: </p>\n\n<p>Hflip/Vflip, Transpose, Rot 90</p>\n\n<p>Random rotate, random shift, random scale</p>",
      "rawMarkdown": "Things work for me: \n\nHflip/Vflip, Transpose, Rot 90\n\nRandom rotate, random shift, random scale",
      "votes": null
    },
    {
      "id": "523473",
      "postDate": "04/26/2019 11:16:26",
      "content": "<p>for me works hflip, random crop, soft color jitter.\nand TTA with hflip, random crop </p>",
      "rawMarkdown": "for me works hflip, random crop, soft color jitter.\nand TTA with hflip, random crop",
      "votes": null
    },
    {
      "id": "523578",
      "postDate": "04/26/2019 14:47:57",
      "content": "<p>Interesting I was hesitant in trying jitter, I am trying brightness contrast let's see how it plays out.</p>",
      "rawMarkdown": "Interesting I was hesitant in trying jitter, I am trying brightness contrast let's see how it plays out.",
      "votes": null
    },
    {
      "id": "523583",
      "postDate": "04/26/2019 15:01:05",
      "content": "<p>With 5 fold cv, you can easily break 0.6 with models lower than 0.58.  Still trying to use different loss and augmentation too, can't break 0.585 with single model in fold.</p>",
      "rawMarkdown": "With 5 fold cv, you can easily break 0.6 with models lower than 0.58.  Still trying to use different loss and augmentation too, can't break 0.585 with single model in fold.",
      "votes": null
    },
    {
      "id": "523588",
      "postDate": "04/26/2019 15:12:27",
      "content": "<p>by 5 fold cv you mean  run the same model with 5 different validation splits and average the predictions before deciding on thresholds right?</p>",
      "rawMarkdown": "by 5 fold cv you mean  run the same model with 5 different validation splits and average the predictions before deciding on thresholds right?",
      "votes": null
    },
    {
      "id": "523595",
      "postDate": "04/26/2019 15:27:48",
      "content": "<p>do you use this transforms at tta stage?</p>",
      "rawMarkdown": "do you use this transforms at tta stage?",
      "votes": null
    },
    {
      "id": "523596",
      "postDate": "04/26/2019 15:30:23",
      "content": "<p>In fastai we don't need to explicitly specify to use augmentation right? <code>learn.TTA()</code> handles this right?</p>",
      "rawMarkdown": "In fastai we don't need to explicitly specify to use augmentation right? `learn.TTA()` handles this right?",
      "votes": null
    },
    {
      "id": "523597",
      "postDate": "04/26/2019 15:33:15",
      "content": "<p>Yep. No ensemble right now,  still trying to find how to help model understand the data better.</p>",
      "rawMarkdown": "Yep. No ensemble right now,  still trying to find how to help model understand the data better.",
      "votes": null
    },
    {
      "id": "523600",
      "postDate": "04/26/2019 15:36:47",
      "content": "<p>Nice this helps</p>",
      "rawMarkdown": "Nice this helps",
      "votes": null
    },
    {
      "id": "523695",
      "postDate": "04/26/2019 18:58:23",
      "content": "<p>Transpose is an interesting choice</p>",
      "rawMarkdown": "Transpose is an interesting choice",
      "votes": null
    },
    {
      "id": "523776",
      "postDate": "04/27/2019 01:49:31",
      "content": "<p>For my solution, TTA doesn't help at all.</p>",
      "rawMarkdown": "For my solution, TTA doesn't help at all.",
      "votes": null
    },
    {
      "id": "523813",
      "postDate": "04/27/2019 05:28:58",
      "content": "<p>Random rotate is interesting, since I think for some images the orientation matters, but there seems like there might be enough images where it doesn't matter.</p>",
      "rawMarkdown": "Random rotate is interesting, since I think for some images the orientation matters, but there seems like there might be enough images where it doesn't matter.",
      "votes": null
    },
    {
      "id": "524154",
      "postDate": "04/28/2019 05:03:39",
      "content": "<p>does hflip really work? it seems strange, cause all pictures angles look same.</p>",
      "rawMarkdown": "does hflip really work? it seems strange, cause all pictures angles look same.",
      "votes": null
    },
    {
      "id": "525401",
      "postDate": "04/30/2019 22:22:06",
      "content": "<p>Wondering what augmentation library you are using now? </p>",
      "rawMarkdown": "Wondering what augmentation library you are using now?",
      "votes": null
    },
    {
      "id": "525551",
      "postDate": "05/01/2019 09:13:31",
      "content": "<p>So far I haven't found good augmentation except horizontal flip and random crop. I've tried a few such as vertical flip and rotate but none them has really worked well. </p>",
      "rawMarkdown": "So far I haven't found good augmentation except horizontal flip and random crop. I've tried a few such as vertical flip and rotate but none them has really worked well.",
      "votes": null
    },
    {
      "id": "525585",
      "postDate": "05/01/2019 10:39:33",
      "content": "<p>what if do vflip and rotation for image with suitable labels?</p>",
      "rawMarkdown": "what if do vflip and rotation for image with suitable labels?",
      "votes": null
    },
    {
      "id": "525586",
      "postDate": "05/01/2019 10:42:26",
      "content": "<p>Is your current score on LB from a single model?</p>",
      "rawMarkdown": "Is your current score on LB from a single model?",
      "votes": null
    },
    {
      "id": "525731",
      "postDate": "05/01/2019 16:02:03",
      "content": "<p><a href=\"/allerria\">@allerria</a>\nWhat do you mean by image with suitable labels? </p>\n\n<p><a href=\"/wh1tezzz\">@wh1tezzz</a>\nI averaged 5-fold outputs.</p>",
      "rawMarkdown": "allerria\nWhat do you mean by image with suitable labels? \n\n@wh1tezzz\nI averaged 5-fold outputs.",
      "votes": null
    },
    {
      "id": "525760",
      "postDate": "05/01/2019 17:16:28",
      "content": "<p><a href=\"/appian\">@appian</a> it's just hypothesis that images with specific labels could be transformed more hardly</p>",
      "rawMarkdown": "appian it's just hypothesis that images with specific labels could be transformed more hardly",
      "votes": null
    },
    {
      "id": "525808",
      "postDate": "05/01/2019 18:42:05",
      "content": "<p>Regarding the color jitter, what range of parameters for brightness, contrast and saturation would you suggest? </p>",
      "rawMarkdown": "Regarding the color jitter, what range of parameters for brightness, contrast and saturation would you suggest?",
      "votes": null
    },
    {
      "id": "537735",
      "postDate": "05/27/2019 14:25:05",
      "content": "<p>Hi <a href=\"/appian\">@appian</a> I am wondering right now how many augmentations do you have for you 660 model?</p>",
      "rawMarkdown": "Hi @appian I am wondering right now how many augmentations do you have for you 660 model?",
      "votes": null
    },
    {
      "id": "538299",
      "postDate": "05/28/2019 11:22:23",
      "content": "<p>I added one more augmentation. So it's three of them now.</p>",
      "rawMarkdown": "I added one more augmentation. So it's three of them now.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 523433,
      "author_name": "guru001",
      "author_url": "",
      "post_date": "04/26/2019 08:53:37",
      "content": "<p>nice topic to dig !</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 523444,
      "author_name": "wh1tezzz",
      "author_url": "",
      "post_date": "04/26/2019 09:30:41",
      "content": "<p>Things work for me: </p>\n\n<p>Hflip/Vflip, Transpose, Rot 90</p>\n\n<p>Random rotate, random shift, random scale</p>",
      "votes": null,
      "replies": [
        {
          "id": 523595,
          "author_name": "allerria",
          "author_url": "",
          "post_date": "04/26/2019 15:27:48",
          "content": "<p>do you use this transforms at tta stage?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 523596,
          "author_name": "axel81",
          "author_url": "",
          "post_date": "04/26/2019 15:30:23",
          "content": "<p>In fastai we don't need to explicitly specify to use augmentation right? <code>learn.TTA()</code> handles this right?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 523695,
          "author_name": "kokecacao",
          "author_url": "",
          "post_date": "04/26/2019 18:58:23",
          "content": "<p>Transpose is an interesting choice</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 523776,
          "author_name": "wh1tezzz",
          "author_url": "",
          "post_date": "04/27/2019 01:49:31",
          "content": "<p>For my solution, TTA doesn't help at all.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 523813,
          "author_name": "eseymour",
          "author_url": "",
          "post_date": "04/27/2019 05:28:58",
          "content": "<p>Random rotate is interesting, since I think for some images the orientation matters, but there seems like there might be enough images where it doesn't matter.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 524154,
          "author_name": "allerria",
          "author_url": "",
          "post_date": "04/28/2019 05:03:39",
          "content": "<p>does hflip really work? it seems strange, cause all pictures angles look same.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 523473,
      "author_name": "allerria",
      "author_url": "",
      "post_date": "04/26/2019 11:16:26",
      "content": "<p>for me works hflip, random crop, soft color jitter.\nand TTA with hflip, random crop </p>",
      "votes": null,
      "replies": [
        {
          "id": 523578,
          "author_name": "axel81",
          "author_url": "",
          "post_date": "04/26/2019 14:47:57",
          "content": "<p>Interesting I was hesitant in trying jitter, I am trying brightness contrast let's see how it plays out.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 523583,
          "author_name": "jionie",
          "author_url": "",
          "post_date": "04/26/2019 15:01:05",
          "content": "<p>With 5 fold cv, you can easily break 0.6 with models lower than 0.58.  Still trying to use different loss and augmentation too, can't break 0.585 with single model in fold.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 523588,
          "author_name": "axel81",
          "author_url": "",
          "post_date": "04/26/2019 15:12:27",
          "content": "<p>by 5 fold cv you mean  run the same model with 5 different validation splits and average the predictions before deciding on thresholds right?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 523597,
          "author_name": "jionie",
          "author_url": "",
          "post_date": "04/26/2019 15:33:15",
          "content": "<p>Yep. No ensemble right now,  still trying to find how to help model understand the data better.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 523600,
          "author_name": "axel81",
          "author_url": "",
          "post_date": "04/26/2019 15:36:47",
          "content": "<p>Nice this helps</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 525808,
          "author_name": "",
          "author_url": "",
          "post_date": "05/01/2019 18:42:05",
          "content": "<p>Regarding the color jitter, what range of parameters for brightness, contrast and saturation would you suggest? </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 525401,
      "author_name": "strideradu",
      "author_url": "",
      "post_date": "04/30/2019 22:22:06",
      "content": "<p>Wondering what augmentation library you are using now? </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 525551,
      "author_name": "appian",
      "author_url": "",
      "post_date": "05/01/2019 09:13:31",
      "content": "<p>So far I haven't found good augmentation except horizontal flip and random crop. I've tried a few such as vertical flip and rotate but none them has really worked well. </p>",
      "votes": null,
      "replies": [
        {
          "id": 525585,
          "author_name": "allerria",
          "author_url": "",
          "post_date": "05/01/2019 10:39:33",
          "content": "<p>what if do vflip and rotation for image with suitable labels?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 525586,
          "author_name": "wh1tezzz",
          "author_url": "",
          "post_date": "05/01/2019 10:42:26",
          "content": "<p>Is your current score on LB from a single model?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 525731,
          "author_name": "appian",
          "author_url": "",
          "post_date": "05/01/2019 16:02:03",
          "content": "<p><a href=\"/allerria\">@allerria</a>\nWhat do you mean by image with suitable labels? </p>\n\n<p><a href=\"/wh1tezzz\">@wh1tezzz</a>\nI averaged 5-fold outputs.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 525760,
          "author_name": "allerria",
          "author_url": "",
          "post_date": "05/01/2019 17:16:28",
          "content": "<p><a href=\"/appian\">@appian</a> it's just hypothesis that images with specific labels could be transformed more hardly</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 537735,
          "author_name": "strideradu",
          "author_url": "",
          "post_date": "05/27/2019 14:25:05",
          "content": "<p>Hi <a href=\"/appian\">@appian</a> I am wondering right now how many augmentations do you have for you 660 model?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 538299,
          "author_name": "appian",
          "author_url": "",
          "post_date": "05/28/2019 11:22:23",
          "content": "<p>I added one more augmentation. So it's three of them now.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "523355": "I am trying to do simple augmentations like flip, rotate, zoom, wrap and using ResNet34 but wasn't able to get a score &gt; 0.580. What are the augmentation methods that I can use to improve ?",
    "523433": "nice topic to dig !",
    "523444": "Things work for me: \n\nHflip/Vflip, Transpose, Rot 90\n\nRandom rotate, random shift, random scale",
    "523473": "for me works hflip, random crop, soft color jitter.\nand TTA with hflip, random crop",
    "523578": "Interesting I was hesitant in trying jitter, I am trying brightness contrast let's see how it plays out.",
    "523583": "With 5 fold cv, you can easily break 0.6 with models lower than 0.58.  Still trying to use different loss and augmentation too, can't break 0.585 with single model in fold.",
    "523588": "by 5 fold cv you mean  run the same model with 5 different validation splits and average the predictions before deciding on thresholds right?",
    "523595": "do you use this transforms at tta stage?",
    "523596": "In fastai we don't need to explicitly specify to use augmentation right? `learn.TTA()` handles this right?",
    "523597": "Yep. No ensemble right now,  still trying to find how to help model understand the data better.",
    "523600": "Nice this helps",
    "523695": "Transpose is an interesting choice",
    "523776": "For my solution, TTA doesn't help at all.",
    "523813": "Random rotate is interesting, since I think for some images the orientation matters, but there seems like there might be enough images where it doesn't matter.",
    "524154": "does hflip really work? it seems strange, cause all pictures angles look same.",
    "525401": "Wondering what augmentation library you are using now?",
    "525551": "So far I haven't found good augmentation except horizontal flip and random crop. I've tried a few such as vertical flip and rotate but none them has really worked well.",
    "525585": "what if do vflip and rotation for image with suitable labels?",
    "525586": "Is your current score on LB from a single model?",
    "525731": "allerria\nWhat do you mean by image with suitable labels? \n\n@wh1tezzz\nI averaged 5-fold outputs.",
    "525760": "appian it's just hypothesis that images with specific labels could be transformed more hardly",
    "525808": "Regarding the color jitter, what range of parameters for brightness, contrast and saturation would you suggest?",
    "537735": "Hi @appian I am wondering right now how many augmentations do you have for you 660 model?",
    "538299": "I added one more augmentation. So it's three of them now."
  },
  "source": "meta"
}