{
  "id": 325235,
  "title": "***Cannot achieve test accuracy >0.5 for tf learning effnet***",
  "url": "/competitions/sorghum-id-fgvc-9/discussion/325235",
  "author_name": "",
  "post_date": "2022-05-15T13:36:35.308848900Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi all, I did transfer learning and tried converging top layers but was not able to achieve train accuracy&gt;0.4. Then i tried training full model without freezing any layer with lr=0.0001 and ahceived train and val accuracy &gt;0.90 fairly quickly. But now I am not able to achieve test accuracy &gt;0.5 . Why do think this is the case. Also I have few doubts . If willing to clarify pls write in comment box: </p>\n<p>1) What are the major changes in code in FGVC and normal classification?<br>\n2) Why wasn't i able to achieve convergence by just training top layer?<br>\n3) Is it possible that high bias is due to overfitting since i am training whole model without freezing any layer? For reference: I have used efficientnetB2 as my base model.</p>\n<p>Any suggestions are welcome. Pls write your opinions in the comments. </p>\n<p><strong>You guys can check out my notebook on my profile</strong> or <a href=\"https://www.kaggle.com/code/aayushdeswal/95-sorghum-efficientnetb2-by-beginner/notebook\" target=\"_blank\">here</a> is the linkto notebook.</p>",
  "messages": [
    {
      "id": "1790943",
      "postDate": "05/15/2022 13:36:35",
      "content": "<p>Hi all, I did transfer learning and tried converging top layers but was not able to achieve train accuracy&gt;0.4. Then i tried training full model without freezing any layer with lr=0.0001 and ahceived train and val accuracy &gt;0.90 fairly quickly. But now I am not able to achieve test accuracy &gt;0.5 . Why do think this is the case. Also I have few doubts . If willing to clarify pls write in comment box: </p>\n<p>1) What are the major changes in code in FGVC and normal classification?<br>\n2) Why wasn't i able to achieve convergence by just training top layer?<br>\n3) Is it possible that high bias is due to overfitting since i am training whole model without freezing any layer? For reference: I have used efficientnetB2 as my base model.</p>\n<p>Any suggestions are welcome. Pls write your opinions in the comments. </p>\n<p><strong>You guys can check out my notebook on my profile</strong> or <a href=\"https://www.kaggle.com/code/aayushdeswal/95-sorghum-efficientnetb2-by-beginner/notebook\" target=\"_blank\">here</a> is the linkto notebook.</p>",
      "rawMarkdown": "Hi all, I did transfer learning and tried converging top layers but was not able to achieve train accuracy>0.4. Then i tried training full model without freezing any layer with lr=0.0001 and ahceived train and val accuracy >0.90 fairly quickly. But now I am not able to achieve test accuracy >0.5 . Why do think this is the case. Also I have few doubts . If willing to clarify pls write in comment box: \n\n1) What are the major changes in code in FGVC and normal classification?\n2) Why wasn't i able to achieve convergence by just training top layer?\n3) Is it possible that high bias is due to overfitting since i am training whole model without freezing any layer? For reference: I have used efficientnetB2 as my base model.\n\nAny suggestions are welcome. Pls write your opinions in the comments. \n\n\n**You guys can check out my notebook on my profile** or [here](https://www.kaggle.com/code/aayushdeswal/95-sorghum-efficientnetb2-by-beginner/notebook) is the linkto notebook.",
      "votes": null
    },
    {
      "id": "1791537",
      "postDate": "05/16/2022 05:44:42",
      "content": "<p>Try bigger model and bigger image size, Also try adding something anti overfitting methods </p>",
      "rawMarkdown": "Try bigger model and bigger image size, Also try adding something anti overfitting methods",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1791537,
      "author_name": "mithilsalunkhe",
      "author_url": "",
      "post_date": "05/16/2022 05:44:42",
      "content": "<p>Try bigger model and bigger image size, Also try adding something anti overfitting methods </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1790943": "Hi all, I did transfer learning and tried converging top layers but was not able to achieve train accuracy>0.4. Then i tried training full model without freezing any layer with lr=0.0001 and ahceived train and val accuracy >0.90 fairly quickly. But now I am not able to achieve test accuracy >0.5 . Why do think this is the case. Also I have few doubts . If willing to clarify pls write in comment box: \n\n1) What are the major changes in code in FGVC and normal classification?\n2) Why wasn't i able to achieve convergence by just training top layer?\n3) Is it possible that high bias is due to overfitting since i am training whole model without freezing any layer? For reference: I have used efficientnetB2 as my base model.\n\nAny suggestions are welcome. Pls write your opinions in the comments. \n\n\n**You guys can check out my notebook on my profile** or [here](https://www.kaggle.com/code/aayushdeswal/95-sorghum-efficientnetb2-by-beginner/notebook) is the linkto notebook.",
    "1791537": "Try bigger model and bigger image size, Also try adding something anti overfitting methods"
  },
  "source": "meta"
}