{
  "id": 204842,
  "title": "Accuracy stuck at 61% ",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/204842",
  "author_name": "",
  "post_date": "2020-12-17T05:59:14.969560900Z",
  "votes": 1,
  "comment_count": 10,
  "views": 0,
  "content": "<p>Hello, <br>\nI believe my model is predicting class 3 repeatedly. Can anyone debug my kernel? I am using imagedatagenerator() in keras. </p>\n<p><a href=\"https://www.kaggle.com/zainahmedsharif/cassanava22\" target=\"_blank\">https://www.kaggle.com/zainahmedsharif/cassanava22</a></p>\n<p>I would really appreciate it. </p>",
  "messages": [
    {
      "id": "1116380",
      "postDate": "12/17/2020 05:59:14",
      "content": "<p>Hello, <br>\nI believe my model is predicting class 3 repeatedly. Can anyone debug my kernel? I am using imagedatagenerator() in keras. </p>\n<p><a href=\"https://www.kaggle.com/zainahmedsharif/cassanava22\" target=\"_blank\">https://www.kaggle.com/zainahmedsharif/cassanava22</a></p>\n<p>I would really appreciate it. </p>",
      "rawMarkdown": "Hello, \nI believe my model is predicting class 3 repeatedly. Can anyone debug my kernel? I am using imagedatagenerator() in keras. \n\nhttps://www.kaggle.com/zainahmedsharif/cassanava22\n\nI would really appreciate it.",
      "votes": null
    },
    {
      "id": "1116671",
      "postDate": "12/17/2020 11:38:27",
      "content": "<p>Maybe try to increase input image size?</p>",
      "rawMarkdown": "Maybe try to increase input image size?",
      "votes": null
    },
    {
      "id": "1116683",
      "postDate": "12/17/2020 11:46:02",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/deepdreamx\" target=\"_blank\">@deepdreamx</a> . Model was too simple to extract any meaningful features and image size was too small I think. </p>",
      "rawMarkdown": "Thanks @deepdreamx . Model was too simple to extract any meaningful features and image size was too small I think.",
      "votes": null
    },
    {
      "id": "1116773",
      "postDate": "12/17/2020 13:08:08",
      "content": "<p>Increasing the number of epochs helps too. I get acc=0.8 from 0.6 by increasing epochs.</p>",
      "rawMarkdown": "Increasing the number of epochs helps too. I get acc=0.8 from 0.6 by increasing epochs.",
      "votes": null
    },
    {
      "id": "1116782",
      "postDate": "12/17/2020 13:16:19",
      "content": "<p>Also why didn't you use hyper_model(), you can add more layers in that.</p>",
      "rawMarkdown": "Also why didn't you use hyper_model(), you can add more layers in that.",
      "votes": null
    },
    {
      "id": "1117350",
      "postDate": "12/18/2020 01:03:57",
      "content": "<p>your inference code has some bug, I guess. I've got the same experience before, I fixed this by using tes_df as submission intead of the sample_submission. You can follow this great <a href=\"https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-inference-tta\" target=\"_blank\">kernel</a> to create your inference submission .</p>",
      "rawMarkdown": "your inference code has some bug, I guess. I've got the same experience before, I fixed this by using tes_df as submission intead of the sample_submission. You can follow this great [kernel](https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-inference-tta) to create your inference submission .",
      "votes": null
    },
    {
      "id": "1117433",
      "postDate": "12/18/2020 03:46:21",
      "content": "<p><a href=\"https://www.kaggle.com/wantsu\" target=\"_blank\">@wantsu</a>  yes you are right. I corrected it afterwards using 2 images in the test image folder</p>",
      "rawMarkdown": "wantsu  yes you are right. I corrected it afterwards using 2 images in the test image folder",
      "votes": null
    },
    {
      "id": "1117436",
      "postDate": "12/18/2020 03:53:39",
      "content": "<p>yes you are right. I corrected it afterwards using 2 images in the test image folder</p>",
      "rawMarkdown": "yes you are right. I corrected it afterwards using 2 images in the test image folder",
      "votes": null
    },
    {
      "id": "1117439",
      "postDate": "12/18/2020 03:55:59",
      "content": "<p>I used hyper_model() for keras tuner. But accuracy did not increase past 61%. I have increased image size and model complexity. I think it would help. </p>",
      "rawMarkdown": "I used hyper_model() for keras tuner. But accuracy did not increase past 61%. I have increased image size and model complexity. I think it would help.",
      "votes": null
    },
    {
      "id": "1117440",
      "postDate": "12/18/2020 03:57:02",
      "content": "<p>Thanks for the response. The accuracy stalling actually causes the early stopping to stop the training. </p>",
      "rawMarkdown": "Thanks for the response. The accuracy stalling actually causes the early stopping to stop the training.",
      "votes": null
    },
    {
      "id": "1117441",
      "postDate": "12/18/2020 03:58:03",
      "content": "<p>Yes this has helped. I believe that model was too simple to learn meaningful features and image size was also too small. </p>",
      "rawMarkdown": "Yes this has helped. I believe that model was too simple to learn meaningful features and image size was also too small.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1116671,
      "author_name": "deepdreamx",
      "author_url": "",
      "post_date": "12/17/2020 11:38:27",
      "content": "<p>Maybe try to increase input image size?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1117441,
          "author_name": "zainahmedsharif",
          "author_url": "",
          "post_date": "12/18/2020 03:58:03",
          "content": "<p>Yes this has helped. I believe that model was too simple to learn meaningful features and image size was also too small. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1116683,
      "author_name": "zainahmedsharif",
      "author_url": "",
      "post_date": "12/17/2020 11:46:02",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/deepdreamx\" target=\"_blank\">@deepdreamx</a> . Model was too simple to extract any meaningful features and image size was too small I think. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1116773,
      "author_name": "surayuthpintawong",
      "author_url": "",
      "post_date": "12/17/2020 13:08:08",
      "content": "<p>Increasing the number of epochs helps too. I get acc=0.8 from 0.6 by increasing epochs.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1117440,
          "author_name": "zainahmedsharif",
          "author_url": "",
          "post_date": "12/18/2020 03:57:02",
          "content": "<p>Thanks for the response. The accuracy stalling actually causes the early stopping to stop the training. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1116782,
      "author_name": "harshsdw",
      "author_url": "",
      "post_date": "12/17/2020 13:16:19",
      "content": "<p>Also why didn't you use hyper_model(), you can add more layers in that.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1117439,
          "author_name": "zainahmedsharif",
          "author_url": "",
          "post_date": "12/18/2020 03:55:59",
          "content": "<p>I used hyper_model() for keras tuner. But accuracy did not increase past 61%. I have increased image size and model complexity. I think it would help. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1117350,
      "author_name": "wantsu",
      "author_url": "",
      "post_date": "12/18/2020 01:03:57",
      "content": "<p>your inference code has some bug, I guess. I've got the same experience before, I fixed this by using tes_df as submission intead of the sample_submission. You can follow this great <a href=\"https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-inference-tta\" target=\"_blank\">kernel</a> to create your inference submission .</p>",
      "votes": null,
      "replies": [
        {
          "id": 1117436,
          "author_name": "zainahmedsharif",
          "author_url": "",
          "post_date": "12/18/2020 03:53:39",
          "content": "<p>yes you are right. I corrected it afterwards using 2 images in the test image folder</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1117433,
      "author_name": "zainahmedsharif",
      "author_url": "",
      "post_date": "12/18/2020 03:46:21",
      "content": "<p><a href=\"https://www.kaggle.com/wantsu\" target=\"_blank\">@wantsu</a>  yes you are right. I corrected it afterwards using 2 images in the test image folder</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1116380": "Hello, \nI believe my model is predicting class 3 repeatedly. Can anyone debug my kernel? I am using imagedatagenerator() in keras. \n\nhttps://www.kaggle.com/zainahmedsharif/cassanava22\n\nI would really appreciate it.",
    "1116671": "Maybe try to increase input image size?",
    "1116683": "Thanks @deepdreamx . Model was too simple to extract any meaningful features and image size was too small I think.",
    "1116773": "Increasing the number of epochs helps too. I get acc=0.8 from 0.6 by increasing epochs.",
    "1116782": "Also why didn't you use hyper_model(), you can add more layers in that.",
    "1117350": "your inference code has some bug, I guess. I've got the same experience before, I fixed this by using tes_df as submission intead of the sample_submission. You can follow this great [kernel](https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-inference-tta) to create your inference submission .",
    "1117433": "wantsu  yes you are right. I corrected it afterwards using 2 images in the test image folder",
    "1117436": "yes you are right. I corrected it afterwards using 2 images in the test image folder",
    "1117439": "I used hyper_model() for keras tuner. But accuracy did not increase past 61%. I have increased image size and model complexity. I think it would help.",
    "1117440": "Thanks for the response. The accuracy stalling actually causes the early stopping to stop the training.",
    "1117441": "Yes this has helped. I believe that model was too simple to learn meaningful features and image size was also too small."
  },
  "source": "meta"
}