{
  "id": 235058,
  "title": "Approach to fine tuning?",
  "url": "/competitions/plant-pathology-2021-fgvc8/discussion/235058",
  "author_name": "",
  "post_date": "2021-04-27T14:32:09.992779600Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi,<br>\nI have a question regarding transfer learning for this competition. How are you fine-tuning your model? </p>\n<ol>\n<li>Are you unfreezing specific layers? If so, how are you selecting which layers to unfreeze? Is it trial and error or is there some hard math behind it?</li>\n<li>Are you making the whole pretrained_model trainable after warming up your model? In this scenario, are you guys using an external resource to train the model because Kaggle's resources run out of memory.</li>\n<li>Are you not fine-tuning at all?</li>\n</ol>",
  "messages": [
    {
      "id": "1286063",
      "postDate": "04/27/2021 14:32:09",
      "content": "<p>Hi,<br>\nI have a question regarding transfer learning for this competition. How are you fine-tuning your model? </p>\n<ol>\n<li>Are you unfreezing specific layers? If so, how are you selecting which layers to unfreeze? Is it trial and error or is there some hard math behind it?</li>\n<li>Are you making the whole pretrained_model trainable after warming up your model? In this scenario, are you guys using an external resource to train the model because Kaggle's resources run out of memory.</li>\n<li>Are you not fine-tuning at all?</li>\n</ol>",
      "rawMarkdown": "Hi,\nI have a question regarding transfer learning for this competition. How are you fine-tuning your model? \n1.  Are you unfreezing specific layers? If so, how are you selecting which layers to unfreeze? Is it trial and error or is there some hard math behind it?\n2. Are you making the whole pretrained_model trainable after warming up your model? In this scenario, are you guys using an external resource to train the model because Kaggle's resources run out of memory.\n3. Are you not fine-tuning at all?",
      "votes": null
    },
    {
      "id": "1286476",
      "postDate": "04/28/2021 03:44:27",
      "content": "<p>you can refer this for better approaches <a href=\"https://keras.io/guides/transfer_learning/\" target=\"_blank\">https://keras.io/guides/transfer_learning/</a></p>",
      "rawMarkdown": "you can refer this for better approaches https://keras.io/guides/transfer_learning/",
      "votes": null
    },
    {
      "id": "1286494",
      "postDate": "04/28/2021 04:21:22",
      "content": "<p>Hi Rajat,<br>\nThanks for the link. I have already gone through the keras documentation however and was looking more for an answer from a Kaggler’s perspective. If you have been using transfer learning in this competition, what is the approach you are taking?</p>",
      "rawMarkdown": "Hi Rajat,\nThanks for the link. I have already gone through the keras documentation however and was looking more for an answer from a Kaggler’s perspective. If you have been using transfer learning in this competition, what is the approach you are taking?",
      "votes": null
    },
    {
      "id": "1288046",
      "postDate": "04/29/2021 16:21:20",
      "content": "<p>I have found some notebooks using the second point but case 2('In this scenario, are you guys using an external resource to train the model because Kaggle's resources run out of memory') it  depends , like i am using external source to speed up training</p>",
      "rawMarkdown": "I have found some notebooks using the second point but case 2('In this scenario, are you guys using an external resource to train the model because Kaggle's resources run out of memory') it  depends , like i am using external source to speed up training",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1286476,
      "author_name": "rajatsanjaypatel",
      "author_url": "",
      "post_date": "04/28/2021 03:44:27",
      "content": "<p>you can refer this for better approaches <a href=\"https://keras.io/guides/transfer_learning/\" target=\"_blank\">https://keras.io/guides/transfer_learning/</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 1286494,
          "author_name": "mreenav",
          "author_url": "",
          "post_date": "04/28/2021 04:21:22",
          "content": "<p>Hi Rajat,<br>\nThanks for the link. I have already gone through the keras documentation however and was looking more for an answer from a Kaggler’s perspective. If you have been using transfer learning in this competition, what is the approach you are taking?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1288046,
      "author_name": "swaralipibose",
      "author_url": "",
      "post_date": "04/29/2021 16:21:20",
      "content": "<p>I have found some notebooks using the second point but case 2('In this scenario, are you guys using an external resource to train the model because Kaggle's resources run out of memory') it  depends , like i am using external source to speed up training</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1286063": "Hi,\nI have a question regarding transfer learning for this competition. How are you fine-tuning your model? \n1.  Are you unfreezing specific layers? If so, how are you selecting which layers to unfreeze? Is it trial and error or is there some hard math behind it?\n2. Are you making the whole pretrained_model trainable after warming up your model? In this scenario, are you guys using an external resource to train the model because Kaggle's resources run out of memory.\n3. Are you not fine-tuning at all?",
    "1286476": "you can refer this for better approaches https://keras.io/guides/transfer_learning/",
    "1286494": "Hi Rajat,\nThanks for the link. I have already gone through the keras documentation however and was looking more for an answer from a Kaggler’s perspective. If you have been using transfer learning in this competition, what is the approach you are taking?",
    "1288046": "I have found some notebooks using the second point but case 2('In this scenario, are you guys using an external resource to train the model because Kaggle's resources run out of memory') it  depends , like i am using external source to speed up training"
  },
  "source": "meta"
}