{
  "id": 173404,
  "title": "Please help me with some suggestions, Pytorch users",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/173404",
  "author_name": "",
  "post_date": "2020-08-09T06:05:03.385883800Z",
  "votes": 3,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I have implemented image+meta data on efficient-net b4 with 5 folds. Used OOF predictions and got LB score of 0.89 and CV 0.95 ! Sighs ! I am struggling to fine tune the model parameters and hyper-parameters . Also can't find anything to try out. If you guys have any suggestion or advice for me in this regards that would be much helpful. Thanks in advance.</p>",
  "messages": [
    {
      "id": "963563",
      "postDate": "08/09/2020 06:05:03",
      "content": "<p>I have implemented image+meta data on efficient-net b4 with 5 folds. Used OOF predictions and got LB score of 0.89 and CV 0.95 ! Sighs ! I am struggling to fine tune the model parameters and hyper-parameters . Also can't find anything to try out. If you guys have any suggestion or advice for me in this regards that would be much helpful. Thanks in advance.</p>",
      "rawMarkdown": "I have implemented image+meta data on efficient-net b4 with 5 folds. Used OOF predictions and got LB score of 0.89 and CV 0.95 ! Sighs ! I am struggling to fine tune the model parameters and hyper-parameters . Also can't find anything to try out. If you guys have any suggestion or advice for me in this regards that would be much helpful. Thanks in advance.",
      "votes": null
    },
    {
      "id": "963601",
      "postDate": "08/09/2020 06:47:42",
      "content": "<p>One Advice: If you are not already, Use data provided by Chris, it is triple stratified, it will at least give you some stability!</p>",
      "rawMarkdown": "One Advice: If you are not already, Use data provided by Chris, it is triple stratified, it will at least give you some stability!",
      "votes": null
    },
    {
      "id": "963795",
      "postDate": "08/09/2020 09:40:00",
      "content": "<p>A senior brother shared his dataset that was extracted from Chris's tfrecord dataset. Working on it. Thank you for your suggestion. Any suggestion on how can I fine tune my models hyperparameters/parameters? I am just using a learning rate scheduler.</p>",
      "rawMarkdown": "A senior brother shared his dataset that was extracted from Chris's tfrecord dataset. Working on it. Thank you for your suggestion. Any suggestion on how can I fine tune my models hyperparameters/parameters? I am just using a learning rate scheduler.",
      "votes": null
    },
    {
      "id": "963799",
      "postDate": "08/09/2020 09:46:44",
      "content": "<p>Fine tuning needs a lot of experimentation, you might want to try different hyper parameters but since the competition is about to end, there is not much time to do that, what you can do is shorten up many changes, like as most of the people here are doing, you can use:</p>\n<ul>\n<li>Optimizer: Adam</li>\n<li>Loss: BCE or Focal Loss</li>\n<li>Scheduler: ReduceLROnPlateau</li>\n</ul>\n<p>Now, keeping them same, you can try different augmentations and learning rate!</p>\n<p>Other than that, I think you should read notebooks shared by people and study their approach, that will help.</p>",
      "rawMarkdown": "Fine tuning needs a lot of experimentation, you might want to try different hyper parameters but since the competition is about to end, there is not much time to do that, what you can do is shorten up many changes, like as most of the people here are doing, you can use:\n\n- Optimizer: Adam\n- Loss: BCE or Focal Loss\n- Scheduler: ReduceLROnPlateau\n\nNow, keeping them same, you can try different augmentations and learning rate!\n\nOther than that, I think you should read notebooks shared by people and study their approach, that will help.",
      "votes": null
    },
    {
      "id": "964205",
      "postDate": "08/09/2020 16:55:43",
      "content": "<p>Thank you for your invaluable suggestion. I am working with it. </p>",
      "rawMarkdown": "Thank you for your invaluable suggestion. I am working with it.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 963601,
      "author_name": "sarques",
      "author_url": "",
      "post_date": "08/09/2020 06:47:42",
      "content": "<p>One Advice: If you are not already, Use data provided by Chris, it is triple stratified, it will at least give you some stability!</p>",
      "votes": null,
      "replies": [
        {
          "id": 963795,
          "author_name": "tawheedrony",
          "author_url": "",
          "post_date": "08/09/2020 09:40:00",
          "content": "<p>A senior brother shared his dataset that was extracted from Chris's tfrecord dataset. Working on it. Thank you for your suggestion. Any suggestion on how can I fine tune my models hyperparameters/parameters? I am just using a learning rate scheduler.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 963799,
          "author_name": "sarques",
          "author_url": "",
          "post_date": "08/09/2020 09:46:44",
          "content": "<p>Fine tuning needs a lot of experimentation, you might want to try different hyper parameters but since the competition is about to end, there is not much time to do that, what you can do is shorten up many changes, like as most of the people here are doing, you can use:</p>\n<ul>\n<li>Optimizer: Adam</li>\n<li>Loss: BCE or Focal Loss</li>\n<li>Scheduler: ReduceLROnPlateau</li>\n</ul>\n<p>Now, keeping them same, you can try different augmentations and learning rate!</p>\n<p>Other than that, I think you should read notebooks shared by people and study their approach, that will help.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 964205,
          "author_name": "tawheedrony",
          "author_url": "",
          "post_date": "08/09/2020 16:55:43",
          "content": "<p>Thank you for your invaluable suggestion. I am working with it. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "963563": "I have implemented image+meta data on efficient-net b4 with 5 folds. Used OOF predictions and got LB score of 0.89 and CV 0.95 ! Sighs ! I am struggling to fine tune the model parameters and hyper-parameters . Also can't find anything to try out. If you guys have any suggestion or advice for me in this regards that would be much helpful. Thanks in advance.",
    "963601": "One Advice: If you are not already, Use data provided by Chris, it is triple stratified, it will at least give you some stability!",
    "963795": "A senior brother shared his dataset that was extracted from Chris's tfrecord dataset. Working on it. Thank you for your suggestion. Any suggestion on how can I fine tune my models hyperparameters/parameters? I am just using a learning rate scheduler.",
    "963799": "Fine tuning needs a lot of experimentation, you might want to try different hyper parameters but since the competition is about to end, there is not much time to do that, what you can do is shorten up many changes, like as most of the people here are doing, you can use:\n\n- Optimizer: Adam\n- Loss: BCE or Focal Loss\n- Scheduler: ReduceLROnPlateau\n\nNow, keeping them same, you can try different augmentations and learning rate!\n\nOther than that, I think you should read notebooks shared by people and study their approach, that will help.",
    "964205": "Thank you for your invaluable suggestion. I am working with it."
  },
  "source": "meta"
}