{
  "id": 93334,
  "title": "Best single model",
  "url": "/competitions/iwildcam-2019-fgvc6/discussion/93334",
  "author_name": "",
  "post_date": "2019-05-25T23:51:34.906920300Z",
  "votes": null,
  "comment_count": 8,
  "views": 0,
  "content": "<p>It seems there has been no discussion of the best single model. I would love to hear what kinds of models you guys are using.</p>\n\n<p>I'll start, though it's probably not very useful given my current score. I am using a densenet121 model in fastai.</p>",
  "messages": [
    {
      "id": "537000",
      "postDate": "05/25/2019 23:51:34",
      "content": "<p>It seems there has been no discussion of the best single model. I would love to hear what kinds of models you guys are using.</p>\n\n<p>I'll start, though it's probably not very useful given my current score. I am using a densenet121 model in fastai.</p>",
      "rawMarkdown": "It seems there has been no discussion of the best single model. I would love to hear what kinds of models you guys are using.\n\nI'll start, though it's probably not very useful given my current score. I am using a densenet121 model in fastai.",
      "votes": null
    },
    {
      "id": "537611",
      "postDate": "05/27/2019 10:22:01",
      "content": "<p>I'm using a resnet34. Trying to keep the model smaller so i can train on larger images without running out of memory. What image sizes are you training with?</p>",
      "rawMarkdown": "I'm using a resnet34. Trying to keep the model smaller so i can train on larger images without running out of memory. What image sizes are you training with?",
      "votes": null
    },
    {
      "id": "537961",
      "postDate": "05/27/2019 23:12:37",
      "content": "<p>I'm using Densenet121. </p>",
      "rawMarkdown": "I'm using Densenet121.",
      "votes": null
    },
    {
      "id": "537962",
      "postDate": "05/27/2019 23:15:24",
      "content": "<p>thanks for sharing. how do you handle large images? \nwhen I tried to rescaling images the kernel died. so I use float16 not float32. \nusing imagegenerator looks useful but it seems very slow.</p>",
      "rawMarkdown": "thanks for sharing. how do you handle large images? \nwhen I tried to rescaling images the kernel died. so I use float16 not float32. \nusing imagegenerator looks useful but it seems very slow.",
      "votes": null
    },
    {
      "id": "537963",
      "postDate": "05/27/2019 23:15:26",
      "content": "<p>Nice... I am using too but having trouble improving score. How long did you run for? Is this with augmentation or k-fold averaging?</p>",
      "rawMarkdown": "Nice... I am using too but having trouble improving score. How long did you run for? Is this with augmentation or k-fold averaging?",
      "votes": null
    },
    {
      "id": "538059",
      "postDate": "05/28/2019 04:14:01",
      "content": "<p>it takes very long time. \n4min for 1 epoch with 64*64 size image and without argument. \n6min for 1 epoch with 32*32 size image and using ImageDataGenerator</p>",
      "rawMarkdown": "it takes very long time. \n4min for 1 epoch with 64*64 size image and without argument. \n6min for 1 epoch with 32*32 size image and using ImageDataGenerator",
      "votes": null
    },
    {
      "id": "538069",
      "postDate": "05/28/2019 04:44:31",
      "content": "<p>what kind of results did you get with 32x32? I currently have 32x32</p>",
      "rawMarkdown": "what kind of results did you get with 32x32? I currently have 32x32",
      "votes": null
    },
    {
      "id": "538201",
      "postDate": "05/28/2019 09:53:43",
      "content": "<p>Yea i'm using mixed precision as well. My batch size is 64.  Epochs takes about ~ 1.5h</p>",
      "rawMarkdown": "Yea i'm using mixed precision as well. My batch size is 64.  Epochs takes about ~ 1.5h",
      "votes": null
    },
    {
      "id": "538271",
      "postDate": "05/28/2019 10:56:01",
      "content": "<p>64x64 size image results a little better score. \nbut it makes learning very slow and many memory problems.</p>",
      "rawMarkdown": "64x64 size image results a little better score. \nbut it makes learning very slow and many memory problems.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 537611,
      "author_name": "shaunspinelli",
      "author_url": "",
      "post_date": "05/27/2019 10:22:01",
      "content": "<p>I'm using a resnet34. Trying to keep the model smaller so i can train on larger images without running out of memory. What image sizes are you training with?</p>",
      "votes": null,
      "replies": [
        {
          "id": 537962,
          "author_name": "jaeson",
          "author_url": "",
          "post_date": "05/27/2019 23:15:24",
          "content": "<p>thanks for sharing. how do you handle large images? \nwhen I tried to rescaling images the kernel died. so I use float16 not float32. \nusing imagegenerator looks useful but it seems very slow.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 538201,
          "author_name": "shaunspinelli",
          "author_url": "",
          "post_date": "05/28/2019 09:53:43",
          "content": "<p>Yea i'm using mixed precision as well. My batch size is 64.  Epochs takes about ~ 1.5h</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 537961,
      "author_name": "jaeson",
      "author_url": "",
      "post_date": "05/27/2019 23:12:37",
      "content": "<p>I'm using Densenet121. </p>",
      "votes": null,
      "replies": [
        {
          "id": 537963,
          "author_name": "tanlikesmath",
          "author_url": "",
          "post_date": "05/27/2019 23:15:26",
          "content": "<p>Nice... I am using too but having trouble improving score. How long did you run for? Is this with augmentation or k-fold averaging?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 538059,
          "author_name": "jaeson",
          "author_url": "",
          "post_date": "05/28/2019 04:14:01",
          "content": "<p>it takes very long time. \n4min for 1 epoch with 64*64 size image and without argument. \n6min for 1 epoch with 32*32 size image and using ImageDataGenerator</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 538069,
          "author_name": "tanlikesmath",
          "author_url": "",
          "post_date": "05/28/2019 04:44:31",
          "content": "<p>what kind of results did you get with 32x32? I currently have 32x32</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 538271,
          "author_name": "jaeson",
          "author_url": "",
          "post_date": "05/28/2019 10:56:01",
          "content": "<p>64x64 size image results a little better score. \nbut it makes learning very slow and many memory problems.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "537000": "It seems there has been no discussion of the best single model. I would love to hear what kinds of models you guys are using.\n\nI'll start, though it's probably not very useful given my current score. I am using a densenet121 model in fastai.",
    "537611": "I'm using a resnet34. Trying to keep the model smaller so i can train on larger images without running out of memory. What image sizes are you training with?",
    "537961": "I'm using Densenet121.",
    "537962": "thanks for sharing. how do you handle large images? \nwhen I tried to rescaling images the kernel died. so I use float16 not float32. \nusing imagegenerator looks useful but it seems very slow.",
    "537963": "Nice... I am using too but having trouble improving score. How long did you run for? Is this with augmentation or k-fold averaging?",
    "538059": "it takes very long time. \n4min for 1 epoch with 64*64 size image and without argument. \n6min for 1 epoch with 32*32 size image and using ImageDataGenerator",
    "538069": "what kind of results did you get with 32x32? I currently have 32x32",
    "538201": "Yea i'm using mixed precision as well. My batch size is 64.  Epochs takes about ~ 1.5h",
    "538271": "64x64 size image results a little better score. \nbut it makes learning very slow and many memory problems."
  },
  "source": "meta"
}