{
  "id": 204793,
  "title": "Best Single Model",
  "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/204793",
  "author_name": "Tahsin Mostafiz",
  "post_date": "2020-12-16T21:27:28.934000",
  "votes": 35,
  "comment_count": 34,
  "views": 0,
  "content": "<p>Drop your best single model scores here. I'll start:</p>\n<p>Model: Fast-Resnest50<br>\nIMG_DIM: 320<br>\nEPOCH: 15<br>\nCV: 0.9216<br>\nLB: 0.9230</p>",
  "messages": [
    {
      "id": 1116118,
      "postDate": "2020-12-16T21:27:28.933Z",
      "content": "<p>Drop your best single model scores here. I'll start:</p>\n<p>Model: Fast-Resnest50<br>\nIMG_DIM: 320<br>\nEPOCH: 15<br>\nCV: 0.9216<br>\nLB: 0.9230</p>",
      "rawMarkdown": "Drop your best single model scores here. I'll start:\n\nModel: Fast-Resnest50\nIMG_DIM: 320\nEPOCH: 15\nCV: 0.9216\nLB: 0.9230",
      "votes": 35
    },
    {
      "id": 1130768,
      "postDate": "2020-12-29T09:50:46.003Z",
      "content": "<p>Model: resnet200d_320<br>\nSize: 512x512<br>\nCV: 0.9564 [0.974  0.9629 0.9918 0.9601 0.9615 0.9835 0.9858 0.9253 0.8633 0.9146  0.998 ]<br>\nLB: 0.965 (TTA: hflip)</p>",
      "rawMarkdown": "Model: resnet200d_320\nSize: 512x512\nCV: 0.9564 [0.974  0.9629 0.9918 0.9601 0.9615 0.9835 0.9858 0.9253 0.8633 0.9146  0.998 ]\nLB: 0.965 (TTA: hflip)\n",
      "votes": 17,
      "replies": [
        {
          "id": 1130878,
          "postDate": "2020-12-29T11:18:10.210Z",
          "content": "<p>Hi,how do you use resnet200d?I am not find the pretrained weight in the 'timm' package.And,do you use kfold?Thanks!</p>",
          "rawMarkdown": "Hi,how do you use resnet200d?I am not find the pretrained weight in the 'timm' package.And,do you use kfold?Thanks!"
        },
        {
          "id": 1130889,
          "postDate": "2020-12-29T11:40:17.063Z",
          "content": "<p>You can install latest version like below.</p>\n<blockquote>\n  <p>pip install git+<a href=\"https://github.com/rwightman/pytorch-image-models@392595c7eb02c3f6353a7806aa9d1e3f569d47e7\" target=\"_blank\">https://github.com/rwightman/pytorch-image-models@392595c7eb02c3f6353a7806aa9d1e3f569d47e7</a></p>\n</blockquote>\n<p>Yes, it's 5fold CV.</p>",
          "rawMarkdown": "You can install latest version like below.\n\n> pip install git+https://github.com/rwightman/pytorch-image-models@392595c7eb02c3f6353a7806aa9d1e3f569d47e7\n\nYes, it's 5fold CV."
        },
        {
          "id": 1130898,
          "postDate": "2020-12-29T11:53:02.387Z",
          "content": "<p>Thank you very much!</p>",
          "rawMarkdown": "Thank you very much!"
        },
        {
          "id": 1130899,
          "postDate": "2020-12-29T11:53:47.063Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 1131160,
          "postDate": "2020-12-29T15:13:14.200Z",
          "content": "<p>i tried all the \"big models\" available, efficientnetb7, resnest, TResNet …etc.<br>\nsuprising, only resnet200d works.</p>",
          "rawMarkdown": "i tried all the \"big models\" available, efficientnetb7, resnest, TResNet ...etc.\nsuprising, only resnet200d works.",
          "votes": 15
        },
        {
          "id": 1131932,
          "postDate": "2020-12-30T04:52:46.123Z",
          "content": "<p>Hi,what do you use lr_scheduler?And how many epochs do you train in one fold,it can convergence？Thanks！</p>",
          "rawMarkdown": "Hi,what do you use lr_scheduler?And how many epochs do you train in one fold,it can convergence？Thanks！"
        },
        {
          "id": 1132395,
          "postDate": "2020-12-30T11:33:03.843Z",
          "content": "<p>I also test the resnet200d,but I find in fold0 cv is 0.9347,compared to you score is so lower,I think I have some error in training.</p>",
          "rawMarkdown": "I also test the resnet200d,but I find in fold0 cv is 0.9347,compared to you score is so lower,I think I have some error in training."
        },
        {
          "id": 1132400,
          "postDate": "2020-12-30T11:35:51.947Z",
          "content": "<p>I shared my current approach here <a href=\"https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/207577\" target=\"_blank\">https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/207577</a></p>",
          "rawMarkdown": "I shared my current approach here https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/207577",
          "votes": 2
        },
        {
          "id": 1212478,
          "postDate": "2021-02-21T08:54:58.253Z",
          "content": "<p>any idea why?</p>",
          "rawMarkdown": "any idea why?"
        }
      ]
    },
    {
      "id": 1156973,
      "postDate": "2021-01-17T14:53:12.493Z",
      "content": "<p>efficient net b7 5 fold cv: 0.94943    0.94844 0.9443  0.9498223719    0.9531905262 LB 0.962</p>",
      "rawMarkdown": "efficient net b7 5 fold cv: 0.94943\t0.94844\t0.9443\t0.9498223719\t0.9531905262 LB 0.962",
      "votes": 6
    },
    {
      "id": 1119574,
      "postDate": "2020-12-20T07:33:16.727Z",
      "content": "<p>Model: Fast-Resnest50<br>\nCV Strategy: Multi-Label Stratified Group K-Fold(K=5)<br>\nSize: 384x384<br>\nEpoch: 20<br>\n[worst fold] CV: 0.918, LB: 0.933<br>\n[best fold] CV: 0.931, LB: 0.939<br>\n[5-Fold Avg] OOF: 0.925, LB: 0.943</p>\n<p>I'll try to increase image sizes.</p>\n<p>------------ Update 1 ------------ </p>\n<p>Size: 512x512<br>\n[5-Fold Avg] OOF: 0.931, LB: 0.949</p>\n<p>------------ Update 2 ------------ <br>\nfix some bugs</p>\n<p>Size: 512x512<br>\n[5-Fold Avg] OOF: 0.939, LB: 0.956</p>\n<p>------------ Update 3 ------------ </p>\n<p>Model: Resnest101<br>\nSize: 512x512<br>\n[5-Fold Avg] OOF: 0.944, LB: 0.959</p>",
      "rawMarkdown": "Model: Fast-Resnest50\nCV Strategy: Multi-Label Stratified Group K-Fold(K=5)\nSize: 384x384\nEpoch: 20\n[worst fold] CV: 0.918, LB: 0.933\n[best fold] CV: 0.931, LB: 0.939\n[5-Fold Avg] OOF: 0.925, LB: 0.943\n\nI'll try to increase image sizes.\n\n------------ Update 1 ------------ \n\nSize: 512x512\n[5-Fold Avg] OOF: 0.931, LB: 0.949\n\n------------ Update 2 ------------ \nfix some bugs\n\nSize: 512x512\n[5-Fold Avg] OOF: 0.939, LB: 0.956\n\n------------ Update 3 ------------ \n\nModel: Resnest101\nSize: 512x512\n[5-Fold Avg] OOF: 0.944, LB: 0.959",
      "votes": 6,
      "replies": [
        {
          "id": 1131128,
          "postDate": "2020-12-29T14:58:16.333Z",
          "content": "<p>Hi,can you share how do you use scheduler?And how many epochs do you train?Thanks!</p>",
          "rawMarkdown": "Hi,can you share how do you use scheduler?And how many epochs do you train?Thanks!",
          "votes": 1
        },
        {
          "id": 1131702,
          "postDate": "2020-12-29T22:36:56.830Z",
          "content": "<p>epoch: 16 - 20<br>\nscheduler: CosineAnnealingWarmRestarts (as well as <a href=\"https://www.kaggle.com/ttahara/ranzcr-multi-head-resnext50-32x4d-training\" target=\"_blank\">my notebook</a> )</p>",
          "rawMarkdown": "epoch: 16 - 20\nscheduler: CosineAnnealingWarmRestarts (as well as [my notebook](https://www.kaggle.com/ttahara/ranzcr-multi-head-resnext50-32x4d-training) )"
        },
        {
          "id": 1131786,
          "postDate": "2020-12-30T00:53:16.813Z",
          "content": "<p>Thanks for you reply!Do you use only one cycle to trian 16-20 epochs?Thanks!</p>",
          "rawMarkdown": "Thanks for you reply!Do you use only one cycle to trian 16-20 epochs?Thanks!"
        },
        {
          "id": 1131836,
          "postDate": "2020-12-30T02:57:44.387Z",
          "content": "<p>Yes. I use only one cycle.</p>",
          "rawMarkdown": "Yes. I use only one cycle."
        },
        {
          "id": 1132576,
          "postDate": "2020-12-30T14:00:50.170Z",
          "content": "<p>oh,you get a new score!how many init lr do you use?1e-4?</p>",
          "rawMarkdown": "oh,you get a new score!how many init lr do you use?1e-4?",
          "votes": -1
        }
      ]
    },
    {
      "id": 1226633,
      "postDate": "2021-03-04T17:49:01.683Z",
      "content": "<p>Resnet200d<br>\n640x640<br>\nlight augs<br>\n8 epochs<br>\ncustom loss<br>\nCV 0.9673, LB 0.966</p>",
      "rawMarkdown": "Resnet200d\n640x640\nlight augs\n8 epochs\ncustom loss\nCV 0.9673, LB 0.966",
      "votes": 3,
      "replies": [
        {
          "id": 1233386,
          "postDate": "2021-03-10T11:03:57.107Z",
          "content": "<p><a href=\"https://www.kaggle.com/yannmajewski\" target=\"_blank\">@yannmajewski</a> I tried 3 stage training strategy and got a maximum up to CV 96.39 with:  <br>\nresnet200d_320<br>\n640x640<br>\nheavy augs<br>\n9 epochs<br>\nfocal loss<br>\nCan you tell how you managed to get a 96.73 CV?</p>",
          "rawMarkdown": "@yannmajewski I tried 3 stage training strategy and got a maximum up to CV 96.39 with:  \nresnet200d_320\n640x640\nheavy augs\n9 epochs\nfocal loss\nCan you tell how you managed to get a 96.73 CV?"
        }
      ]
    },
    {
      "id": 1223158,
      "postDate": "2021-03-02T13:06:10.060Z",
      "content": "<p>Model: resnet200d (5fold avg without TTA)<br>\nimage size: 684 x 684</p>\n<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.9617</td>\n<td>0.969</td>\n</tr>\n<tr>\n<td>0.9598</td>\n<td>0.967</td>\n</tr>\n<tr>\n<td>0.9582</td>\n<td>0.966</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "Model: resnet200d (5fold avg without TTA)\nimage size: 684 x 684\n| CV | LB |\n| --- | --- |\n| 0.9617 | 0.969 | \n0.9598 | 0.967 | \n0.9582 | 0.966 |\n",
      "votes": 4,
      "replies": [
        {
          "id": 1230792,
          "postDate": "2021-03-08T13:09:00.787Z",
          "content": "<p>How this possible, Impressive 👍</p>",
          "rawMarkdown": "How this possible, Impressive 👍",
          "votes": 1
        }
      ]
    },
    {
      "id": 1140928,
      "postDate": "2021-01-06T11:24:42.097Z",
      "content": "<p>Model: resnet200d (Single Model)<br>\nIMG_DIM: 384<br>\nEpoch: 384<br>\nTTA: HFlip, Rotate, Crop etc. <br>\nCV: 0.946<br>\nLB: 0.943</p>",
      "rawMarkdown": "Model: resnet200d (Single Model)\nIMG_DIM: 384\nEpoch: 384\nTTA: HFlip, Rotate, Crop etc. \nCV: 0.946\nLB: 0.943",
      "votes": 1
    },
    {
      "id": 1117877,
      "postDate": "2020-12-18T14:47:08.067Z",
      "content": "<p>I use efficientnet_b4_ns,512 size,5 folds,cv is 0.9078,and lb is 0.935</p>",
      "rawMarkdown": "I use efficientnet_b4_ns,512 size,5 folds,cv is 0.9078,and lb is 0.935",
      "votes": 1,
      "replies": [
        {
          "id": 1118150,
          "postDate": "2020-12-18T19:19:30.953Z",
          "content": "<p>What type of augmentations are you using? </p>",
          "rawMarkdown": "What type of augmentations are you using? "
        },
        {
          "id": 1119610,
          "postDate": "2020-12-20T08:27:14.923Z",
          "content": "<p>training I use RandomCrop,hflip</p>",
          "rawMarkdown": "training I use RandomCrop,hflip"
        }
      ]
    },
    {
      "id": 1240895,
      "postDate": "2021-03-16T17:54:52.200Z",
      "content": "<p>Single Eff B1-&gt;B4 local_val 95.9-96.3 LB 96.3-&gt; and with ensemble LB 96.5</p>",
      "rawMarkdown": "Single Eff B1->B4 local_val 95.9-96.3 LB 96.3-> and with ensemble LB 96.5"
    },
    {
      "id": 1235325,
      "postDate": "2021-03-12T03:53:09.107Z",
      "content": "<p>InceptionResNetV2 (tensorflow, TPU)<br>\n5 kfold <br>\nCV: 0.953<br>\nLB: 0.949</p>",
      "rawMarkdown": "InceptionResNetV2 (tensorflow, TPU)\n5 kfold \nCV: 0.953\nLB: 0.949\n\n"
    },
    {
      "id": 1234153,
      "postDate": "2021-03-11T02:29:43.293Z",
      "content": "<p>So many people using 640x640. How long does it take you to run 1 epoch at 640x640?</p>",
      "rawMarkdown": "So many people using 640x640. How long does it take you to run 1 epoch at 640x640?",
      "replies": [
        {
          "id": 1234281,
          "postDate": "2021-03-11T05:25:23.957Z",
          "content": "<p>For me is: <br>\nResnet200d<br>\n35MIN - V100<br>\n65MIN - P100</p>",
          "rawMarkdown": "For me is: \nResnet200d\n35MIN - V100\n65MIN - P100"
        },
        {
          "id": 1234905,
          "postDate": "2021-03-11T17:15:23.917Z",
          "content": "<p>by my local machine (a single Titan RTX):</p>\n<p>ResNet200d: 20min<br>\nResNeSt200e: 35min</p>",
          "rawMarkdown": "by my local machine (a single Titan RTX):\n\nResNet200d: 20min\nResNeSt200e: 35min"
        }
      ]
    },
    {
      "id": 1225385,
      "postDate": "2021-03-03T15:25:49.903Z",
      "content": "<p>EfficientNet5 (Tensorflow)<br>\nImage Size : 600<br>\n5 Folds without TTA<br>\nCV - 94.8 <br>\nLB - 95.9</p>",
      "rawMarkdown": "EfficientNet5 (Tensorflow)\nImage Size : 600\n5 Folds without TTA\nCV - 94.8 \nLB - 95.9"
    },
    {
      "id": 1166297,
      "postDate": "2021-01-23T14:20:49.470Z",
      "content": "<p>is it possible to achieve decent LB score with efficientnet-b4 rather than big models?</p>",
      "rawMarkdown": "is it possible to achieve decent LB score with efficientnet-b4 rather than big models?"
    },
    {
      "id": 1120676,
      "postDate": "2020-12-21T03:40:26.013Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 1156456,
      "postDate": "2021-01-17T07:03:12.073Z",
      "content": "<p>Thanks for sharing</p>",
      "rawMarkdown": "Thanks for sharing"
    }
  ],
  "comments": [
    {
      "id": 1130768,
      "author_name": "Y.Nakama",
      "author_url": "",
      "post_date": "2020-12-29T09:50:46.003000",
      "content": "<p>Model: resnet200d_320<br>\nSize: 512x512<br>\nCV: 0.9564 [0.974  0.9629 0.9918 0.9601 0.9615 0.9835 0.9858 0.9253 0.8633 0.9146  0.998 ]<br>\nLB: 0.965 (TTA: hflip)</p>",
      "votes": 17,
      "replies": [
        {
          "id": 1130878,
          "author_name": "Bcw93",
          "author_url": "",
          "post_date": "2020-12-29T11:18:10.210000",
          "content": "<p>Hi,how do you use resnet200d?I am not find the pretrained weight in the 'timm' package.And,do you use kfold?Thanks!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1130889,
          "author_name": "Y.Nakama",
          "author_url": "",
          "post_date": "2020-12-29T11:40:17.063000",
          "content": "<p>You can install latest version like below.</p>\n<blockquote>\n  <p>pip install git+<a href=\"https://github.com/rwightman/pytorch-image-models@392595c7eb02c3f6353a7806aa9d1e3f569d47e7\" target=\"_blank\">https://github.com/rwightman/pytorch-image-models@392595c7eb02c3f6353a7806aa9d1e3f569d47e7</a></p>\n</blockquote>\n<p>Yes, it's 5fold CV.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1130898,
          "author_name": "Bcw93",
          "author_url": "",
          "post_date": "2020-12-29T11:53:02.387000",
          "content": "<p>Thank you very much!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1130899,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-12-29T11:53:47.063000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1131160,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2020-12-29T15:13:14.200000",
          "content": "<p>i tried all the \"big models\" available, efficientnetb7, resnest, TResNet …etc.<br>\nsuprising, only resnet200d works.</p>",
          "votes": 15,
          "replies": []
        },
        {
          "id": 1131932,
          "author_name": "Bcw93",
          "author_url": "",
          "post_date": "2020-12-30T04:52:46.123000",
          "content": "<p>Hi,what do you use lr_scheduler?And how many epochs do you train in one fold,it can convergence？Thanks！</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1132395,
          "author_name": "Bcw93",
          "author_url": "",
          "post_date": "2020-12-30T11:33:03.843000",
          "content": "<p>I also test the resnet200d,but I find in fold0 cv is 0.9347,compared to you score is so lower,I think I have some error in training.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1132400,
          "author_name": "Y.Nakama",
          "author_url": "",
          "post_date": "2020-12-30T11:35:51.947000",
          "content": "<p>I shared my current approach here <a href=\"https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/207577\" target=\"_blank\">https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/207577</a></p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1212478,
          "author_name": "Yi Wu",
          "author_url": "",
          "post_date": "2021-02-21T08:54:58.253000",
          "content": "<p>any idea why?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1156973,
      "author_name": "YL",
      "author_url": "",
      "post_date": "2021-01-17T14:53:12.493000",
      "content": "<p>efficient net b7 5 fold cv: 0.94943    0.94844 0.9443  0.9498223719    0.9531905262 LB 0.962</p>",
      "votes": 6,
      "replies": []
    },
    {
      "id": 1119574,
      "author_name": "Tawara",
      "author_url": "",
      "post_date": "2020-12-20T07:33:16.727000",
      "content": "<p>Model: Fast-Resnest50<br>\nCV Strategy: Multi-Label Stratified Group K-Fold(K=5)<br>\nSize: 384x384<br>\nEpoch: 20<br>\n[worst fold] CV: 0.918, LB: 0.933<br>\n[best fold] CV: 0.931, LB: 0.939<br>\n[5-Fold Avg] OOF: 0.925, LB: 0.943</p>\n<p>I'll try to increase image sizes.</p>\n<p>------------ Update 1 ------------ </p>\n<p>Size: 512x512<br>\n[5-Fold Avg] OOF: 0.931, LB: 0.949</p>\n<p>------------ Update 2 ------------ <br>\nfix some bugs</p>\n<p>Size: 512x512<br>\n[5-Fold Avg] OOF: 0.939, LB: 0.956</p>\n<p>------------ Update 3 ------------ </p>\n<p>Model: Resnest101<br>\nSize: 512x512<br>\n[5-Fold Avg] OOF: 0.944, LB: 0.959</p>",
      "votes": 6,
      "replies": [
        {
          "id": 1131128,
          "author_name": "Bcw93",
          "author_url": "",
          "post_date": "2020-12-29T14:58:16.333000",
          "content": "<p>Hi,can you share how do you use scheduler?And how many epochs do you train?Thanks!</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1131702,
          "author_name": "Tawara",
          "author_url": "",
          "post_date": "2020-12-29T22:36:56.830000",
          "content": "<p>epoch: 16 - 20<br>\nscheduler: CosineAnnealingWarmRestarts (as well as <a href=\"https://www.kaggle.com/ttahara/ranzcr-multi-head-resnext50-32x4d-training\" target=\"_blank\">my notebook</a> )</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1131786,
          "author_name": "Bcw93",
          "author_url": "",
          "post_date": "2020-12-30T00:53:16.813000",
          "content": "<p>Thanks for you reply!Do you use only one cycle to trian 16-20 epochs?Thanks!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1131836,
          "author_name": "Tawara",
          "author_url": "",
          "post_date": "2020-12-30T02:57:44.387000",
          "content": "<p>Yes. I use only one cycle.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1132576,
          "author_name": "Bcw93",
          "author_url": "",
          "post_date": "2020-12-30T14:00:50.170000",
          "content": "<p>oh,you get a new score!how many init lr do you use?1e-4?</p>",
          "votes": -1,
          "replies": []
        }
      ]
    },
    {
      "id": 1226633,
      "author_name": "Yann Majewski",
      "author_url": "",
      "post_date": "2021-03-04T17:49:01.683000",
      "content": "<p>Resnet200d<br>\n640x640<br>\nlight augs<br>\n8 epochs<br>\ncustom loss<br>\nCV 0.9673, LB 0.966</p>",
      "votes": 3,
      "replies": [
        {
          "id": 1233386,
          "author_name": "Sanchit Vijay",
          "author_url": "",
          "post_date": "2021-03-10T11:03:57.107000",
          "content": "<p><a href=\"https://www.kaggle.com/yannmajewski\" target=\"_blank\">@yannmajewski</a> I tried 3 stage training strategy and got a maximum up to CV 96.39 with:  <br>\nresnet200d_320<br>\n640x640<br>\nheavy augs<br>\n9 epochs<br>\nfocal loss<br>\nCan you tell how you managed to get a 96.73 CV?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1223158,
      "author_name": "Yiemon773",
      "author_url": "",
      "post_date": "2021-03-02T13:06:10.060000",
      "content": "<p>Model: resnet200d (5fold avg without TTA)<br>\nimage size: 684 x 684</p>\n<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.9617</td>\n<td>0.969</td>\n</tr>\n<tr>\n<td>0.9598</td>\n<td>0.967</td>\n</tr>\n<tr>\n<td>0.9582</td>\n<td>0.966</td>\n</tr>\n</tbody>\n</table>",
      "votes": 4,
      "replies": [
        {
          "id": 1230792,
          "author_name": "Potential",
          "author_url": "",
          "post_date": "2021-03-08T13:09:00.787000",
          "content": "<p>How this possible, Impressive 👍</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1140928,
      "author_name": "Tahsin Mostafiz",
      "author_url": "",
      "post_date": "2021-01-06T11:24:42.097000",
      "content": "<p>Model: resnet200d (Single Model)<br>\nIMG_DIM: 384<br>\nEpoch: 384<br>\nTTA: HFlip, Rotate, Crop etc. <br>\nCV: 0.946<br>\nLB: 0.943</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1117877,
      "author_name": "Bcw93",
      "author_url": "",
      "post_date": "2020-12-18T14:47:08.067000",
      "content": "<p>I use efficientnet_b4_ns,512 size,5 folds,cv is 0.9078,and lb is 0.935</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1118150,
          "author_name": "Tahsin Mostafiz",
          "author_url": "",
          "post_date": "2020-12-18T19:19:30.953000",
          "content": "<p>What type of augmentations are you using? </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1119610,
          "author_name": "Bcw93",
          "author_url": "",
          "post_date": "2020-12-20T08:27:14.923000",
          "content": "<p>training I use RandomCrop,hflip</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1240895,
      "author_name": "Kirderf",
      "author_url": "",
      "post_date": "2021-03-16T17:54:52.200000",
      "content": "<p>Single Eff B1-&gt;B4 local_val 95.9-96.3 LB 96.3-&gt; and with ensemble LB 96.5</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1235325,
      "author_name": "lhagiimn",
      "author_url": "",
      "post_date": "2021-03-12T03:53:09.107000",
      "content": "<p>InceptionResNetV2 (tensorflow, TPU)<br>\n5 kfold <br>\nCV: 0.953<br>\nLB: 0.949</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1234153,
      "author_name": "Gianluca Rossi",
      "author_url": "",
      "post_date": "2021-03-11T02:29:43.293000",
      "content": "<p>So many people using 640x640. How long does it take you to run 1 epoch at 640x640?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1234281,
          "author_name": "zeyu",
          "author_url": "",
          "post_date": "2021-03-11T05:25:23.957000",
          "content": "<p>For me is: <br>\nResnet200d<br>\n35MIN - V100<br>\n65MIN - P100</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1234905,
          "author_name": "Tawara",
          "author_url": "",
          "post_date": "2021-03-11T17:15:23.917000",
          "content": "<p>by my local machine (a single Titan RTX):</p>\n<p>ResNet200d: 20min<br>\nResNeSt200e: 35min</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1225385,
      "author_name": "Vicky Goyal",
      "author_url": "",
      "post_date": "2021-03-03T15:25:49.903000",
      "content": "<p>EfficientNet5 (Tensorflow)<br>\nImage Size : 600<br>\n5 Folds without TTA<br>\nCV - 94.8 <br>\nLB - 95.9</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1166297,
      "author_name": "Yi Wu",
      "author_url": "",
      "post_date": "2021-01-23T14:20:49.470000",
      "content": "<p>is it possible to achieve decent LB score with efficientnet-b4 rather than big models?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1120676,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-12-21T03:40:26.013000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1156456,
      "author_name": "Prediction Girl",
      "author_url": "",
      "post_date": "2021-01-17T07:03:12.073000",
      "content": "<p>Thanks for sharing</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1116118": "Drop your best single model scores here. I'll start:\n\nModel: Fast-Resnest50\nIMG_DIM: 320\nEPOCH: 15\nCV: 0.9216\nLB: 0.9230",
    "1130768": "Model: resnet200d_320\nSize: 512x512\nCV: 0.9564 [0.974  0.9629 0.9918 0.9601 0.9615 0.9835 0.9858 0.9253 0.8633 0.9146  0.998 ]\nLB: 0.965 (TTA: hflip)\n",
    "1156973": "efficient net b7 5 fold cv: 0.94943\t0.94844\t0.9443\t0.9498223719\t0.9531905262 LB 0.962",
    "1119574": "Model: Fast-Resnest50\nCV Strategy: Multi-Label Stratified Group K-Fold(K=5)\nSize: 384x384\nEpoch: 20\n[worst fold] CV: 0.918, LB: 0.933\n[best fold] CV: 0.931, LB: 0.939\n[5-Fold Avg] OOF: 0.925, LB: 0.943\n\nI'll try to increase image sizes.\n\n------------ Update 1 ------------ \n\nSize: 512x512\n[5-Fold Avg] OOF: 0.931, LB: 0.949\n\n------------ Update 2 ------------ \nfix some bugs\n\nSize: 512x512\n[5-Fold Avg] OOF: 0.939, LB: 0.956\n\n------------ Update 3 ------------ \n\nModel: Resnest101\nSize: 512x512\n[5-Fold Avg] OOF: 0.944, LB: 0.959",
    "1226633": "Resnet200d\n640x640\nlight augs\n8 epochs\ncustom loss\nCV 0.9673, LB 0.966",
    "1223158": "Model: resnet200d (5fold avg without TTA)\nimage size: 684 x 684\n| CV | LB |\n| --- | --- |\n| 0.9617 | 0.969 | \n0.9598 | 0.967 | \n0.9582 | 0.966 |\n",
    "1140928": "Model: resnet200d (Single Model)\nIMG_DIM: 384\nEpoch: 384\nTTA: HFlip, Rotate, Crop etc. \nCV: 0.946\nLB: 0.943",
    "1117877": "I use efficientnet_b4_ns,512 size,5 folds,cv is 0.9078,and lb is 0.935",
    "1240895": "Single Eff B1->B4 local_val 95.9-96.3 LB 96.3-> and with ensemble LB 96.5",
    "1235325": "InceptionResNetV2 (tensorflow, TPU)\n5 kfold \nCV: 0.953\nLB: 0.949\n\n",
    "1234153": "So many people using 640x640. How long does it take you to run 1 epoch at 640x640?",
    "1225385": "EfficientNet5 (Tensorflow)\nImage Size : 600\n5 Folds without TTA\nCV - 94.8 \nLB - 95.9",
    "1166297": "is it possible to achieve decent LB score with efficientnet-b4 rather than big models?",
    "1120676": "",
    "1156456": "Thanks for sharing"
  }
}