{
  "id": 166185,
  "title": "How many epochs are good enough or necessary to get stable score?",
  "url": "/competitions/alaska2-image-steganalysis/discussion/166185",
  "author_name": "Urvish",
  "post_date": "2020-07-12T05:21:11.053000",
  "votes": 3,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hello everyone, </p>\n\n<p>I saw posts on the <a href=\"https://www.kaggle.com/c/alaska2-image-steganalysis/discussion/148162\">best single model</a> but did not get any information on how many epochs are most of the competitors doing. So, I would like to ask if others are open to answer, how many epochs are you training to get a stable score or your LB score? </p>\n\n<p>To begin with, I would answer for myself. Based on the public kernel with 0.921 scores, I am doing 30 to 35 epochs and after that my model start overfitting. I am not sure should I do less or more to get a better score. </p>\n\n<p>Thanks for your answers in advance. </p>",
  "messages": [
    {
      "id": 925454,
      "postDate": "2020-07-12T05:21:11.053Z",
      "content": "<p>Hello everyone, </p>\n\n<p>I saw posts on the <a href=\"https://www.kaggle.com/c/alaska2-image-steganalysis/discussion/148162\">best single model</a> but did not get any information on how many epochs are most of the competitors doing. So, I would like to ask if others are open to answer, how many epochs are you training to get a stable score or your LB score? </p>\n\n<p>To begin with, I would answer for myself. Based on the public kernel with 0.921 scores, I am doing 30 to 35 epochs and after that my model start overfitting. I am not sure should I do less or more to get a better score. </p>\n\n<p>Thanks for your answers in advance. </p>",
      "rawMarkdown": "Hello everyone, \n\nI saw posts on the [best single model](https://www.kaggle.com/c/alaska2-image-steganalysis/discussion/148162) but did not get any information on how many epochs are most of the competitors doing. So, I would like to ask if others are open to answer, how many epochs are you training to get a stable score or your LB score? \n\nTo begin with, I would answer for myself. Based on the public kernel with 0.921 scores, I am doing 30 to 35 epochs and after that my model start overfitting. I am not sure should I do less or more to get a better score. \n\nThanks for your answers in advance. ",
      "votes": 3
    },
    {
      "id": 925488,
      "postDate": "2020-07-12T05:54:51.783Z",
      "content": "<p>You should try to combat overfitting. Try more augs, lower the learning rate, maybe even freeze a few layers. You can even start from your best weights instead of spending 40 hours to again train a model</p>",
      "rawMarkdown": "You should try to combat overfitting. Try more augs, lower the learning rate, maybe even freeze a few layers. You can even start from your best weights instead of spending 40 hours to again train a model",
      "votes": 1,
      "replies": [
        {
          "id": 925507,
          "postDate": "2020-07-12T06:11:16.317Z",
          "content": "<p>yep, sounds like a plan. But still, I wanted to know how many epochs are enough 😄 Because I am still confused about how much long training is necessary. </p>",
          "rawMarkdown": "yep, sounds like a plan. But still, I wanted to know how many epochs are enough 😄 Because I am still confused about how much long training is necessary. "
        },
        {
          "id": 925668,
          "postDate": "2020-07-12T08:04:51.570Z",
          "content": "<p>dealing with same issue but in mine case the model is not overfitting but when i resume the training it leads to converge to same point as in previous run could you help <a href=\"/urvishp80\">@urvishp80</a> </p>",
          "rawMarkdown": "dealing with same issue but in mine case the model is not overfitting but when i resume the training it leads to converge to same point as in previous run could you help @urvishp80 ",
          "isDeleted": true
        },
        {
          "id": 925675,
          "postDate": "2020-07-12T08:09:56.750Z",
          "content": "<p>How can you know it converged to the same point/minima?</p>",
          "rawMarkdown": "How can you know it converged to the same point/minima?",
          "votes": 1
        },
        {
          "id": 926392,
          "postDate": "2020-07-12T17:20:57.647Z",
          "content": "<p>So <a href=\"https://www.kaggle.com/shonenkov/train-inference-gpu-baseline\">https://www.kaggle.com/shonenkov/train-inference-gpu-baseline</a> has a record of the training process. It goes through 30 epochs over 2 days or so. </p>\n\n<p>If you copy the methods in that kernel you will get around .921LB </p>",
          "rawMarkdown": "So https://www.kaggle.com/shonenkov/train-inference-gpu-baseline has a record of the training process. It goes through 30 epochs over 2 days or so. \n\nIf you copy the methods in that kernel you will get around .921LB "
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 925488,
      "author_name": "Mighty Rains",
      "author_url": "",
      "post_date": "2020-07-12T05:54:51.783000",
      "content": "<p>You should try to combat overfitting. Try more augs, lower the learning rate, maybe even freeze a few layers. You can even start from your best weights instead of spending 40 hours to again train a model</p>",
      "votes": 1,
      "replies": [
        {
          "id": 925507,
          "author_name": "Urvish",
          "author_url": "",
          "post_date": "2020-07-12T06:11:16.317000",
          "content": "<p>yep, sounds like a plan. But still, I wanted to know how many epochs are enough 😄 Because I am still confused about how much long training is necessary. </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 925668,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-07-12T08:04:51.570000",
          "content": "<p>dealing with same issue but in mine case the model is not overfitting but when i resume the training it leads to converge to same point as in previous run could you help <a href=\"/urvishp80\">@urvishp80</a> </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 925675,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2020-07-12T08:09:56.750000",
          "content": "<p>How can you know it converged to the same point/minima?</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 926392,
          "author_name": "hongy",
          "author_url": "",
          "post_date": "2020-07-12T17:20:57.647000",
          "content": "<p>So <a href=\"https://www.kaggle.com/shonenkov/train-inference-gpu-baseline\">https://www.kaggle.com/shonenkov/train-inference-gpu-baseline</a> has a record of the training process. It goes through 30 epochs over 2 days or so. </p>\n\n<p>If you copy the methods in that kernel you will get around .921LB </p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "925454": "Hello everyone, \n\nI saw posts on the [best single model](https://www.kaggle.com/c/alaska2-image-steganalysis/discussion/148162) but did not get any information on how many epochs are most of the competitors doing. So, I would like to ask if others are open to answer, how many epochs are you training to get a stable score or your LB score? \n\nTo begin with, I would answer for myself. Based on the public kernel with 0.921 scores, I am doing 30 to 35 epochs and after that my model start overfitting. I am not sure should I do less or more to get a better score. \n\nThanks for your answers in advance. ",
    "925488": "You should try to combat overfitting. Try more augs, lower the learning rate, maybe even freeze a few layers. You can even start from your best weights instead of spending 40 hours to again train a model"
  }
}