{
  "id": 102621,
  "title": "What is the highest LB score could you achieve using only data from this competition?",
  "url": "/competitions/aptos2019-blindness-detection/discussion/102621",
  "author_name": "Kirill Talalaev",
  "post_date": "2019-08-03T09:32:55.351000",
  "votes": 9,
  "comment_count": 34,
  "views": 0,
  "content": "",
  "messages": [
    {
      "id": 591187,
      "postDate": "2019-08-03T09:32:55.350Z",
      "rawMarkdown": "",
      "votes": 8
    },
    {
      "id": 591211,
      "postDate": "2019-08-03T10:29:56.707Z",
      "content": "<p>My highest LB score is 0.805 .</p>",
      "rawMarkdown": "My highest LB score is 0.805 .\n",
      "votes": 6,
      "replies": [
        {
          "id": 591231,
          "postDate": "2019-08-03T11:06:38.550Z",
          "content": "<p>thank you!</p>",
          "rawMarkdown": "thank you!"
        },
        {
          "id": 591586,
          "postDate": "2019-08-04T01:02:56.010Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 591624,
          "postDate": "2019-08-04T03:40:11.970Z",
          "content": "<p>Wow from not using data from previous competition?</p>",
          "rawMarkdown": "Wow from not using data from previous competition?"
        },
        {
          "id": 591733,
          "postDate": "2019-08-04T07:02:21.460Z",
          "content": "<p>Yes it is pretty doable .</p>",
          "rawMarkdown": "Yes it is pretty doable .",
          "votes": 1
        },
        {
          "id": 591736,
          "postDate": "2019-08-04T07:05:06.680Z",
          "content": "<p>Yes you are right i am using a lot of augmentations like flips,zoom,working on zoom can improve the model a lot and one thing dont use TTA.</p>",
          "rawMarkdown": "Yes you are right i am using a lot of augmentations like flips,zoom,working on zoom can improve the model a lot and one thing dont use TTA.\n",
          "votes": 4
        },
        {
          "id": 591737,
          "postDate": "2019-08-04T07:10:13.843Z",
          "content": "<p>what kind of augmentation tools are you use ? albumentations, torchvision or fastai.vision?</p>",
          "rawMarkdown": "what kind of augmentation tools are you use ? albumentations, torchvision or fastai.vision?",
          "votes": 1
        },
        {
          "id": 591745,
          "postDate": "2019-08-04T07:41:52.227Z",
          "content": "<p>fastai.vision</p>",
          "rawMarkdown": "fastai.vision\n",
          "votes": 2
        },
        {
          "id": 591782,
          "postDate": "2019-08-04T08:37:16.130Z",
          "rawMarkdown": ""
        },
        {
          "id": 591854,
          "postDate": "2019-08-04T11:30:11.317Z",
          "content": "<p>thanks a lot!</p>",
          "rawMarkdown": "thanks a lot!"
        },
        {
          "id": 592597,
          "postDate": "2019-08-05T14:23:12.893Z",
          "content": "<p><a href=\"/amardeepganguly\">@amardeepganguly</a> Your score is great. May I ask you a question, I am also working at fast.ai augmentations and trying different max_zoom values like 1.1-1.3 but it doesn't seem to make difference. Do you use larger values?</p>",
          "rawMarkdown": "@amardeepganguly Your score is great. May I ask you a question, I am also working at fast.ai augmentations and trying different max_zoom values like 1.1-1.3 but it doesn't seem to make difference. Do you use larger values?",
          "votes": 1
        },
        {
          "id": 593930,
          "postDate": "2019-08-07T09:56:55.490Z",
          "content": "<p>The max  zoom which I have used is 1.35</p>",
          "rawMarkdown": "The max  zoom which I have used is 1.35",
          "votes": 3
        },
        {
          "id": 593931,
          "postDate": "2019-08-07T09:57:38.683Z",
          "content": "<p>Although I noticed improvements from 1.25 for my model</p>",
          "rawMarkdown": "Although I noticed improvements from 1.25 for my model",
          "votes": 2
        },
        {
          "id": 594344,
          "postDate": "2019-08-07T21:39:39.127Z",
          "content": "<p><a href=\"/amardeepganguly\">@amardeepganguly</a> That's really impressive. Which model did you use for training? </p>",
          "rawMarkdown": "@amardeepganguly That's really impressive. Which model did you use for training? "
        },
        {
          "id": 596083,
          "postDate": "2019-08-10T06:29:57.973Z",
          "content": "<p><a href=\"/amardeepganguly\">@amardeepganguly</a>  For this score did you train model from scratch or did you use a pretrained model ?</p>",
          "rawMarkdown": "@amardeepganguly  For this score did you train model from scratch or did you use a pretrained model ?"
        },
        {
          "id": 596460,
          "postDate": "2019-08-10T16:57:36.350Z",
          "content": "<p>I used  a pretrained model.</p>",
          "rawMarkdown": "I used  a pretrained model."
        },
        {
          "id": 599710,
          "postDate": "2019-08-15T09:27:48.497Z",
          "content": "<p>I am using efficientnet now.</p>",
          "rawMarkdown": "I am using efficientnet now."
        },
        {
          "id": 600198,
          "postDate": "2019-08-15T19:25:05.583Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 600636,
          "postDate": "2019-08-16T11:50:00.403Z",
          "content": "<p>The Bens Preprocessing  and using really heavy augmentations like lighting,in the Bens Processing a value of 14 works better in my case.</p>",
          "rawMarkdown": "The Bens Preprocessing  and using really heavy augmentations like lighting,in the Bens Processing a value of 14 works better in my case.\n"
        },
        {
          "id": 600719,
          "postDate": "2019-08-16T13:47:05.933Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 598177,
      "postDate": "2019-08-13T08:31:44.957Z",
      "content": "<p>Now it is 0.811</p>",
      "rawMarkdown": "Now it is 0.811",
      "votes": 2,
      "replies": [
        {
          "id": 598180,
          "postDate": "2019-08-13T08:34:28.820Z",
          "content": "<p>Still using new competition data, increased the no of epochs .</p>",
          "rawMarkdown": "Still using new competition data, increased the no of epochs .",
          "votes": 1
        },
        {
          "id": 598755,
          "postDate": "2019-08-14T02:00:44.810Z",
          "content": "<p>How do you split train/valid?</p>",
          "rawMarkdown": "How do you split train/valid?"
        },
        {
          "id": 599499,
          "postDate": "2019-08-15T03:45:48.090Z",
          "content": "<p>Are you treating the problem as multilabel classification or regression or the normal multiclass classification?  I am getting very bad LB with regression but achieving 0.794 only on new data with multilabel classification.</p>",
          "rawMarkdown": "Are you treating the problem as multilabel classification or regression or the normal multiclass classification?  I am getting very bad LB with regression but achieving 0.794 only on new data with multilabel classification."
        },
        {
          "id": 599708,
          "postDate": "2019-08-15T09:26:47.240Z",
          "content": "<p>The simple spllit_by_rand thing in fastai DataLoaders</p>",
          "rawMarkdown": "The simple spllit_by_rand thing in fastai DataLoaders"
        },
        {
          "id": 614584,
          "postDate": "2019-08-31T17:58:23.577Z",
          "content": "<p><a href=\"/amardeepganguly\">@amardeepganguly</a> that's quite an impressive score for a single model trained on new competition data only. I am struggling to go beyond 0.75 on single model with new competition data only. Using ben's pre-processing, a lot of augmentations (zoom, rotation, horizontal flip), balanced data set using upsampling (tried with class_weights too), normalisation of inputs, short training, long training, batch size of 32, effiientnet-b3, but nothing seems to get me beyond 0.75.</p>\n\n<p>Any tips that may help hit 0.8? <a href=\"/ilovescience\">@ilovescience</a> </p>",
          "rawMarkdown": "@amardeepganguly that's quite an impressive score for a single model trained on new competition data only. I am struggling to go beyond 0.75 on single model with new competition data only. Using ben's pre-processing, a lot of augmentations (zoom, rotation, horizontal flip), balanced data set using upsampling (tried with class_weights too), normalisation of inputs, short training, long training, batch size of 32, effiientnet-b3, but nothing seems to get me beyond 0.75.\n\nAny tips that may help hit 0.8? @ilovescience "
        },
        {
          "id": 614686,
          "postDate": "2019-08-31T22:29:38.947Z",
          "content": "<p>Basit as per my experience the batch size didnt create much of a difference ,I would prefer using efficient b5 with a batch size of 16 rather than using b3 with a batch size of 32 ,using optimised thresholding you can increase your your lb score but i am sure this approach will definitely kick me out of the top 10% in the private leaderboard😄 .Hope this helps.</p>",
          "rawMarkdown": "Basit as per my experience the batch size didnt create much of a difference ,I would prefer using efficient b5 with a batch size of 16 rather than using b3 with a batch size of 32 ,using optimised thresholding you can increase your your lb score but i am sure this approach will definitely kick me out of the top 10% in the private leaderboard😄 .Hope this helps.",
          "votes": 1
        }
      ]
    },
    {
      "id": 592024,
      "postDate": "2019-08-04T16:47:53.820Z",
      "content": "<p>I personally managed to achieve 0.78 LB while testing different concepts. Probably was able push it more - but doesn't make sense - cuz model trained on more and diverse data will always be better. So no point concentrating too much on this competition dataset (personal opinion).</p>",
      "rawMarkdown": "I personally managed to achieve 0.78 LB while testing different concepts. Probably was able push it more - but doesn't make sense - cuz model trained on more and diverse data will always be better. So no point concentrating too much on this competition dataset (personal opinion).",
      "votes": 2,
      "replies": [
        {
          "id": 596085,
          "postDate": "2019-08-10T06:30:53.373Z",
          "content": "<p><a href=\"/gpamoukoff\">@gpamoukoff</a> Were you able to achieve this using pretrained model or from scratch?Thanks</p>",
          "rawMarkdown": "@gpamoukoff Were you able to achieve this using pretrained model or from scratch?Thanks"
        },
        {
          "id": 596153,
          "postDate": "2019-08-10T08:17:09.567Z",
          "content": "<p><a href=\"/decentmakeover\">@decentmakeover</a> finetuned from ImageNet using only this competition data.</p>",
          "rawMarkdown": "@decentmakeover finetuned from ImageNet using only this competition data."
        },
        {
          "id": 596234,
          "postDate": "2019-08-10T10:53:54.090Z",
          "content": "<p>Amazing Thanks, Any Tips for me?</p>",
          "rawMarkdown": "Amazing Thanks, Any Tips for me?"
        },
        {
          "id": 596263,
          "postDate": "2019-08-10T11:44:36.190Z",
          "content": "<p>Keras, proper data augmentation, play a bit with LR schedulers - that will easily get you around 0.78 LB. If you include previous competition data on top - 0.8 is easily achievable. Above that - you'll need to think on how to get out of the box :) </p>",
          "rawMarkdown": "Keras, proper data augmentation, play a bit with LR schedulers - that will easily get you around 0.78 LB. If you include previous competition data on top - 0.8 is easily achievable. Above that - you'll need to think on how to get out of the box :) "
        },
        {
          "id": 596348,
          "postDate": "2019-08-10T13:57:47.150Z",
          "content": "<p>using pytorch reaching 62, ill work harder</p>",
          "rawMarkdown": "using pytorch reaching 62, ill work harder"
        }
      ]
    },
    {
      "id": 599483,
      "postDate": "2019-08-15T03:09:21.327Z",
      "content": "<p>0.794 ... trying harder to increase  further</p>",
      "rawMarkdown": "0.794 ... trying harder to increase  further"
    }
  ],
  "comments": [
    {
      "id": 591211,
      "author_name": "AmardeepGanguly",
      "author_url": "",
      "post_date": "2019-08-03T10:29:56.707000",
      "content": "<p>My highest LB score is 0.805 .</p>",
      "votes": 6,
      "replies": [
        {
          "id": 591231,
          "author_name": "Kirill Talalaev",
          "author_url": "",
          "post_date": "2019-08-03T11:06:38.550000",
          "content": "<p>thank you!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 591586,
          "author_name": "",
          "author_url": "",
          "post_date": "2019-08-04T01:02:56.010000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 591624,
          "author_name": "Quan",
          "author_url": "",
          "post_date": "2019-08-04T03:40:11.970000",
          "content": "<p>Wow from not using data from previous competition?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 591733,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-04T07:02:21.460000",
          "content": "<p>Yes it is pretty doable .</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 591736,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-04T07:05:06.680000",
          "content": "<p>Yes you are right i am using a lot of augmentations like flips,zoom,working on zoom can improve the model a lot and one thing dont use TTA.</p>",
          "votes": 4,
          "replies": []
        },
        {
          "id": 591737,
          "author_name": "leixiang@AInnovation",
          "author_url": "",
          "post_date": "2019-08-04T07:10:13.843000",
          "content": "<p>what kind of augmentation tools are you use ? albumentations, torchvision or fastai.vision?</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 591745,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-04T07:41:52.227000",
          "content": "<p>fastai.vision</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 591782,
          "author_name": "Quan",
          "author_url": "",
          "post_date": "2019-08-04T08:37:16.130000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 591854,
          "author_name": "leixiang@AInnovation",
          "author_url": "",
          "post_date": "2019-08-04T11:30:11.317000",
          "content": "<p>thanks a lot!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 592597,
          "author_name": "Anna Novikova",
          "author_url": "",
          "post_date": "2019-08-05T14:23:12.893000",
          "content": "<p><a href=\"/amardeepganguly\">@amardeepganguly</a> Your score is great. May I ask you a question, I am also working at fast.ai augmentations and trying different max_zoom values like 1.1-1.3 but it doesn't seem to make difference. Do you use larger values?</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 593930,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-07T09:56:55.490000",
          "content": "<p>The max  zoom which I have used is 1.35</p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 593931,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-07T09:57:38.683000",
          "content": "<p>Although I noticed improvements from 1.25 for my model</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 594344,
          "author_name": "Tahsin Mostafiz",
          "author_url": "",
          "post_date": "2019-08-07T21:39:39.127000",
          "content": "<p><a href=\"/amardeepganguly\">@amardeepganguly</a> That's really impressive. Which model did you use for training? </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 596083,
          "author_name": "DecentMakeover",
          "author_url": "",
          "post_date": "2019-08-10T06:29:57.973000",
          "content": "<p><a href=\"/amardeepganguly\">@amardeepganguly</a>  For this score did you train model from scratch or did you use a pretrained model ?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 596460,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-10T16:57:36.350000",
          "content": "<p>I used  a pretrained model.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 599710,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-15T09:27:48.497000",
          "content": "<p>I am using efficientnet now.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 600198,
          "author_name": "",
          "author_url": "",
          "post_date": "2019-08-15T19:25:05.583000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 600636,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-16T11:50:00.403000",
          "content": "<p>The Bens Preprocessing  and using really heavy augmentations like lighting,in the Bens Processing a value of 14 works better in my case.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 600719,
          "author_name": "",
          "author_url": "",
          "post_date": "2019-08-16T13:47:05.933000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 598177,
      "author_name": "AmardeepGanguly",
      "author_url": "",
      "post_date": "2019-08-13T08:31:44.957000",
      "content": "<p>Now it is 0.811</p>",
      "votes": 2,
      "replies": [
        {
          "id": 598180,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-13T08:34:28.820000",
          "content": "<p>Still using new competition data, increased the no of epochs .</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 598755,
          "author_name": "Quan",
          "author_url": "",
          "post_date": "2019-08-14T02:00:44.810000",
          "content": "<p>How do you split train/valid?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 599499,
          "author_name": "Viraj Bagal",
          "author_url": "",
          "post_date": "2019-08-15T03:45:48.090000",
          "content": "<p>Are you treating the problem as multilabel classification or regression or the normal multiclass classification?  I am getting very bad LB with regression but achieving 0.794 only on new data with multilabel classification.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 599708,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-15T09:26:47.240000",
          "content": "<p>The simple spllit_by_rand thing in fastai DataLoaders</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 614584,
          "author_name": "Basit Riaz Sheikh",
          "author_url": "",
          "post_date": "2019-08-31T17:58:23.577000",
          "content": "<p><a href=\"/amardeepganguly\">@amardeepganguly</a> that's quite an impressive score for a single model trained on new competition data only. I am struggling to go beyond 0.75 on single model with new competition data only. Using ben's pre-processing, a lot of augmentations (zoom, rotation, horizontal flip), balanced data set using upsampling (tried with class_weights too), normalisation of inputs, short training, long training, batch size of 32, effiientnet-b3, but nothing seems to get me beyond 0.75.</p>\n\n<p>Any tips that may help hit 0.8? <a href=\"/ilovescience\">@ilovescience</a> </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 614686,
          "author_name": "AmardeepGanguly",
          "author_url": "",
          "post_date": "2019-08-31T22:29:38.947000",
          "content": "<p>Basit as per my experience the batch size didnt create much of a difference ,I would prefer using efficient b5 with a batch size of 16 rather than using b3 with a batch size of 32 ,using optimised thresholding you can increase your your lb score but i am sure this approach will definitely kick me out of the top 10% in the private leaderboard😄 .Hope this helps.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 592024,
      "author_name": "Georgi Pamukov",
      "author_url": "",
      "post_date": "2019-08-04T16:47:53.820000",
      "content": "<p>I personally managed to achieve 0.78 LB while testing different concepts. Probably was able push it more - but doesn't make sense - cuz model trained on more and diverse data will always be better. So no point concentrating too much on this competition dataset (personal opinion).</p>",
      "votes": 2,
      "replies": [
        {
          "id": 596085,
          "author_name": "DecentMakeover",
          "author_url": "",
          "post_date": "2019-08-10T06:30:53.373000",
          "content": "<p><a href=\"/gpamoukoff\">@gpamoukoff</a> Were you able to achieve this using pretrained model or from scratch?Thanks</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 596153,
          "author_name": "Georgi Pamukov",
          "author_url": "",
          "post_date": "2019-08-10T08:17:09.567000",
          "content": "<p><a href=\"/decentmakeover\">@decentmakeover</a> finetuned from ImageNet using only this competition data.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 596234,
          "author_name": "DecentMakeover",
          "author_url": "",
          "post_date": "2019-08-10T10:53:54.090000",
          "content": "<p>Amazing Thanks, Any Tips for me?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 596263,
          "author_name": "Georgi Pamukov",
          "author_url": "",
          "post_date": "2019-08-10T11:44:36.190000",
          "content": "<p>Keras, proper data augmentation, play a bit with LR schedulers - that will easily get you around 0.78 LB. If you include previous competition data on top - 0.8 is easily achievable. Above that - you'll need to think on how to get out of the box :) </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 596348,
          "author_name": "DecentMakeover",
          "author_url": "",
          "post_date": "2019-08-10T13:57:47.150000",
          "content": "<p>using pytorch reaching 62, ill work harder</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 599483,
      "author_name": "Viraj Bagal",
      "author_url": "",
      "post_date": "2019-08-15T03:09:21.327000",
      "content": "<p>0.794 ... trying harder to increase  further</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "591187": "",
    "591211": "My highest LB score is 0.805 .\n",
    "598177": "Now it is 0.811",
    "592024": "I personally managed to achieve 0.78 LB while testing different concepts. Probably was able push it more - but doesn't make sense - cuz model trained on more and diverse data will always be better. So no point concentrating too much on this competition dataset (personal opinion).",
    "599483": "0.794 ... trying harder to increase  further"
  }
}