{
  "id": 159140,
  "title": "Higher number of epochs = better score?",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/159140",
  "author_name": "",
  "post_date": "2020-06-16T14:42:58.658193800Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I just let one of my fastest high scoring models (0.905 in 10 epochs in 15 minutes on TPU) run for 50 epochs. What should I expect? </p>\n\n<p>Also, what regularization steps do you guys suggest other than image augmentation and dropout for the comparatively higher number of epochs?</p>",
  "messages": [
    {
      "id": "888754",
      "postDate": "06/16/2020 14:42:58",
      "content": "<p>I just let one of my fastest high scoring models (0.905 in 10 epochs in 15 minutes on TPU) run for 50 epochs. What should I expect? </p>\n\n<p>Also, what regularization steps do you guys suggest other than image augmentation and dropout for the comparatively higher number of epochs?</p>",
      "rawMarkdown": "I just let one of my fastest high scoring models (0.905 in 10 epochs in 15 minutes on TPU) run for 50 epochs. What should I expect? \n\nAlso, what regularization steps do you guys suggest other than image augmentation and dropout for the comparatively higher number of epochs?",
      "votes": null
    },
    {
      "id": "888777",
      "postDate": "06/16/2020 14:59:32",
      "content": "<p>sometimes it does. sometimes it doesnt. you cant know for sure unless you try :) \nunless you use a big models, usually 10 epoches is not enough to get best performance</p>",
      "rawMarkdown": "sometimes it does. sometimes it doesnt. you cant know for sure unless you try :) \nunless you use a big models, usually 10 epoches is not enough to get best performance",
      "votes": null
    },
    {
      "id": "888789",
      "postDate": "06/16/2020 15:04:39",
      "content": "<p>It really depends. After some epochs it goes to overfit your model and you need to find that threshold. 👍 👍 💪 </p>",
      "rawMarkdown": "It really depends. After some epochs it goes to overfit your model and you need to find that threshold. 👍 👍 💪",
      "votes": null
    },
    {
      "id": "888870",
      "postDate": "06/16/2020 15:58:01",
      "content": "<p>There are libraries for a more intuitive <strong>Early Stopping</strong>, using that the model will stop when the model doesn't improve with the epochs. </p>",
      "rawMarkdown": "There are libraries for a more intuitive **Early Stopping**, using that the model will stop when the model doesn't improve with the epochs.",
      "votes": null
    },
    {
      "id": "891552",
      "postDate": "06/18/2020 08:53:06",
      "content": "<p>Try Batch Normalization and downsampling the images using convolution strides. \nBatch Norm almost always helps.</p>\n\n<p>Moreover, you can also try to fiddle with learning rate. This also helps in jumping the accuracy by at least 10% in all my past experience.</p>",
      "rawMarkdown": "Try Batch Normalization and downsampling the images using convolution strides. \nBatch Norm almost always helps.\n\nMoreover, you can also try to fiddle with learning rate. This also helps in jumping the accuracy by at least 10% in all my past experience.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 888777,
      "author_name": "moewie94",
      "author_url": "",
      "post_date": "06/16/2020 14:59:32",
      "content": "<p>sometimes it does. sometimes it doesnt. you cant know for sure unless you try :) \nunless you use a big models, usually 10 epoches is not enough to get best performance</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 888789,
      "author_name": "",
      "author_url": "",
      "post_date": "06/16/2020 15:04:39",
      "content": "<p>It really depends. After some epochs it goes to overfit your model and you need to find that threshold. 👍 👍 💪 </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 888870,
      "author_name": "hiramcho",
      "author_url": "",
      "post_date": "06/16/2020 15:58:01",
      "content": "<p>There are libraries for a more intuitive <strong>Early Stopping</strong>, using that the model will stop when the model doesn't improve with the epochs. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 891552,
      "author_name": "fireheart7",
      "author_url": "",
      "post_date": "06/18/2020 08:53:06",
      "content": "<p>Try Batch Normalization and downsampling the images using convolution strides. \nBatch Norm almost always helps.</p>\n\n<p>Moreover, you can also try to fiddle with learning rate. This also helps in jumping the accuracy by at least 10% in all my past experience.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "888754": "I just let one of my fastest high scoring models (0.905 in 10 epochs in 15 minutes on TPU) run for 50 epochs. What should I expect? \n\nAlso, what regularization steps do you guys suggest other than image augmentation and dropout for the comparatively higher number of epochs?",
    "888777": "sometimes it does. sometimes it doesnt. you cant know for sure unless you try :) \nunless you use a big models, usually 10 epoches is not enough to get best performance",
    "888789": "It really depends. After some epochs it goes to overfit your model and you need to find that threshold. 👍 👍 💪",
    "888870": "There are libraries for a more intuitive **Early Stopping**, using that the model will stop when the model doesn't improve with the epochs.",
    "891552": "Try Batch Normalization and downsampling the images using convolution strides. \nBatch Norm almost always helps.\n\nMoreover, you can also try to fiddle with learning rate. This also helps in jumping the accuracy by at least 10% in all my past experience."
  },
  "source": "meta"
}