{
  "id": 104967,
  "title": "Has anyone achieved a LB score greater than 0.78 using pytorch?",
  "url": "/competitions/aptos2019-blindness-detection/discussion/104967",
  "author_name": "",
  "post_date": "2019-08-20T10:32:27.449871300Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>As title, I checked several kernels, most of them used fastai. I reproduced their data augmentation(flip, ben graham, crop, rotation, color jitter, etc.), used the same network(efficientnet-b3, b5, etc.), try lr scheduler(SGDR, stepLR, etc.), or even trained using old data. My LB score is about only 0.72. The gap between validation score and LB score still large no matter I use old data to train or not(I split data using train_test_split with stratified in sklearn).</p>\n\n<p>I also train a classifier to split train/val set. But it didn't work, either.</p>\n\n<p>Can anyone give me some hints?</p>",
  "messages": [
    {
      "id": "603478",
      "postDate": "08/20/2019 10:32:27",
      "content": "<p>As title, I checked several kernels, most of them used fastai. I reproduced their data augmentation(flip, ben graham, crop, rotation, color jitter, etc.), used the same network(efficientnet-b3, b5, etc.), try lr scheduler(SGDR, stepLR, etc.), or even trained using old data. My LB score is about only 0.72. The gap between validation score and LB score still large no matter I use old data to train or not(I split data using train_test_split with stratified in sklearn).</p>\n\n<p>I also train a classifier to split train/val set. But it didn't work, either.</p>\n\n<p>Can anyone give me some hints?</p>",
      "rawMarkdown": "As title, I checked several kernels, most of them used fastai. I reproduced their data augmentation(flip, ben graham, crop, rotation, color jitter, etc.), used the same network(efficientnet-b3, b5, etc.), try lr scheduler(SGDR, stepLR, etc.), or even trained using old data. My LB score is about only 0.72. The gap between validation score and LB score still large no matter I use old data to train or not(I split data using train_test_split with stratified in sklearn).\n\nI also train a classifier to split train/val set. But it didn't work, either.\n\nCan anyone give me some hints?",
      "votes": null
    },
    {
      "id": "603479",
      "postDate": "08/20/2019 10:34:29",
      "content": "<p>I am using pytorch</p>",
      "rawMarkdown": "I am using pytorch",
      "votes": null
    },
    {
      "id": "603546",
      "postDate": "08/20/2019 12:29:15",
      "content": "<p>It's not pytorch's fault, perhaps your threshold is bad, or you didn't fine tune with 2019 data, or you wrote a bug elsewhere.</p>",
      "rawMarkdown": "It's not pytorch's fault, perhaps your threshold is bad, or you didn't fine tune with 2019 data, or you wrote a bug elsewhere.",
      "votes": null
    },
    {
      "id": "603901",
      "postDate": "08/20/2019 19:40:39",
      "content": "<p>0.820+ with pytorch.  It is possible.  </p>",
      "rawMarkdown": "0.820+ with pytorch.  It is possible.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 603479,
      "author_name": "buaazijian",
      "author_url": "",
      "post_date": "08/20/2019 10:34:29",
      "content": "<p>I am using pytorch</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 603546,
      "author_name": "homoalways",
      "author_url": "",
      "post_date": "08/20/2019 12:29:15",
      "content": "<p>It's not pytorch's fault, perhaps your threshold is bad, or you didn't fine tune with 2019 data, or you wrote a bug elsewhere.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 603901,
      "author_name": "kvigly55",
      "author_url": "",
      "post_date": "08/20/2019 19:40:39",
      "content": "<p>0.820+ with pytorch.  It is possible.  </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "603478": "As title, I checked several kernels, most of them used fastai. I reproduced their data augmentation(flip, ben graham, crop, rotation, color jitter, etc.), used the same network(efficientnet-b3, b5, etc.), try lr scheduler(SGDR, stepLR, etc.), or even trained using old data. My LB score is about only 0.72. The gap between validation score and LB score still large no matter I use old data to train or not(I split data using train_test_split with stratified in sklearn).\n\nI also train a classifier to split train/val set. But it didn't work, either.\n\nCan anyone give me some hints?",
    "603479": "I am using pytorch",
    "603546": "It's not pytorch's fault, perhaps your threshold is bad, or you didn't fine tune with 2019 data, or you wrote a bug elsewhere.",
    "603901": "0.820+ with pytorch.  It is possible."
  },
  "source": "meta"
}