{
  "id": 76867,
  "title": "How do we decide step_size while including Cylic Learning Rate(CLR) in callbacks?",
  "url": "/competitions/quora-insincere-questions-classification/discussion/76867",
  "author_name": "Rajesh Shreedhar",
  "post_date": "2019-01-07T12:48:59.814000",
  "votes": 5,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I can see below documentation in the CLR implementation:\n<strong>step_size: number of training iterations perhalf cycle. Authors suggest setting step_size 2-8 x training iterations in epoch.</strong></p>\n\n<p>Any example would help a lot. </p>",
  "messages": [
    {
      "id": 451653,
      "postDate": "2019-01-07T12:48:59.813Z",
      "content": "<p>I can see below documentation in the CLR implementation:\n<strong>step_size: number of training iterations perhalf cycle. Authors suggest setting step_size 2-8 x training iterations in epoch.</strong></p>\n\n<p>Any example would help a lot. </p>",
      "rawMarkdown": "I can see below documentation in the CLR implementation:\n**step_size: number of training iterations perhalf cycle. Authors suggest setting step_size 2-8 x training iterations in epoch.**\n\nAny example would help a lot. ",
      "votes": 5
    },
    {
      "id": 451854,
      "postDate": "2019-01-07T19:18:30.837Z",
      "content": "<p>Take total training data size (~1.3m) divided by whatever batch size you are using. That is your number of training iterations. </p>",
      "rawMarkdown": "Take total training data size (~1.3m) divided by whatever batch size you are using. That is your number of training iterations. ",
      "votes": 3
    },
    {
      "id": 451802,
      "postDate": "2019-01-07T17:18:44.263Z",
      "content": "<p>You already posted the explanation. Check number of iterations in one epoch and set the step size accordingly.</p>",
      "rawMarkdown": "You already posted the explanation. Check number of iterations in one epoch and set the step size accordingly.",
      "replies": [
        {
          "id": 452839,
          "postDate": "2019-01-09T08:32:02.293Z",
          "content": "<p>dear 大哥...  i want to ask about the batch size without CLR (actually i dont know how to implement CLR in keras 😂)  my batch size is 512 in keras, under this setting, i could only finish 4 fold and 3 epoch in per fold,  but when i increase the batch_size, my local cv will go down, so should i increase the batch size and ignore the cv  in order to train more fold and epoch??</p>\n\n<p>thanks in advance</p>",
          "rawMarkdown": "dear 大哥...  i want to ask about the batch size without CLR (actually i dont know how to implement CLR in keras 😂)  my batch size is 512 in keras, under this setting, i could only finish 4 fold and 3 epoch in per fold,  but when i increase the batch_size, my local cv will go down, so should i increase the batch size and ignore the cv  in order to train more fold and epoch??\n\nthanks in advance\n"
        },
        {
          "id": 453057,
          "postDate": "2019-01-09T15:44:47.027Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 451854,
      "author_name": "Julius Hochmuth",
      "author_url": "",
      "post_date": "2019-01-07T19:18:30.837000",
      "content": "<p>Take total training data size (~1.3m) divided by whatever batch size you are using. That is your number of training iterations. </p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 451802,
      "author_name": "Psi",
      "author_url": "",
      "post_date": "2019-01-07T17:18:44.263000",
      "content": "<p>You already posted the explanation. Check number of iterations in one epoch and set the step size accordingly.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 452839,
          "author_name": "young_boy",
          "author_url": "",
          "post_date": "2019-01-09T08:32:02.293000",
          "content": "<p>dear 大哥...  i want to ask about the batch size without CLR (actually i dont know how to implement CLR in keras 😂)  my batch size is 512 in keras, under this setting, i could only finish 4 fold and 3 epoch in per fold,  but when i increase the batch_size, my local cv will go down, so should i increase the batch size and ignore the cv  in order to train more fold and epoch??</p>\n\n<p>thanks in advance</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 453057,
          "author_name": "",
          "author_url": "",
          "post_date": "2019-01-09T15:44:47.027000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "451653": "I can see below documentation in the CLR implementation:\n**step_size: number of training iterations perhalf cycle. Authors suggest setting step_size 2-8 x training iterations in epoch.**\n\nAny example would help a lot. ",
    "451854": "Take total training data size (~1.3m) divided by whatever batch size you are using. That is your number of training iterations. ",
    "451802": "You already posted the explanation. Check number of iterations in one epoch and set the step size accordingly."
  }
}