{
  "id": 39148,
  "title": "More batch size or bigger input size?",
  "url": "/competitions/carvana-image-masking-challenge/discussion/39148",
  "author_name": "",
  "post_date": "2017-09-08T01:12:00.359299600Z",
  "votes": null,
  "comment_count": 6,
  "views": 0,
  "content": "<ul>\n<li>input size: 1024x1024   batch size : 4   dice: 0.9956  LB score: 0.9962</li>\n<li>input size: 1280x1280   batch size : 4   dice: 0.9963  LB score: 0.9965</li>\n<li>input size: 1280x1920   batch size : 2   dice: 0.9962  LB score: 0.9966</li>\n</ul>\n\n<p>I am using u-net with Keras according to Peter's code.\nIt seemed that more batch size performed better.\nWhat is your input size or batch size?</p>",
  "messages": [
    {
      "id": "219393",
      "postDate": "09/08/2017 01:12:00",
      "content": "<ul>\n<li>input size: 1024x1024   batch size : 4   dice: 0.9956  LB score: 0.9962</li>\n<li>input size: 1280x1280   batch size : 4   dice: 0.9963  LB score: 0.9965</li>\n<li>input size: 1280x1920   batch size : 2   dice: 0.9962  LB score: 0.9966</li>\n</ul>\n\n<p>I am using u-net with Keras according to Peter's code.\nIt seemed that more batch size performed better.\nWhat is your input size or batch size?</p>",
      "rawMarkdown": "input size: 1024x1024   batch size : 4   dice: 0.9956  LB score: 0.9962\n - input size: 1280x1280   batch size : 4   dice: 0.9963  LB score: 0.9965\n - input size: 1280x1920   batch size : 2   dice: 0.9962  LB score: 0.9966\n\nI am using u-net with Keras according to Peter's code.\nIt seemed that more batch size performed better.\nWhat is your input size or batch size?",
      "votes": null
    },
    {
      "id": "220104",
      "postDate": "09/11/2017 07:26:38",
      "content": "<p>Hi, I think bigger input size really matters. Besides , appropriate batch size is important since there are several batch norm operation in U-net. One question, Peter's code does not provide a network for input with 1280x1280, so you code a new network by yourself?</p>",
      "rawMarkdown": "Hi, I think bigger input size really matters. Besides , appropriate batch size is important since there are several batch norm operation in U-net. One question, Peter's code does not provide a network for input with 1280x1280, so you code a new network by yourself?",
      "votes": null
    },
    {
      "id": "220106",
      "postDate": "09/11/2017 07:31:17",
      "content": "<p>Yes, but I just change the input size.</p>",
      "rawMarkdown": "Yes, but I just change the input size.",
      "votes": null
    },
    {
      "id": "220170",
      "postDate": "09/11/2017 11:58:34",
      "content": "<p>how about the memory size of your GPU?</p>",
      "rawMarkdown": "how about the memory size of your GPU?",
      "votes": null
    },
    {
      "id": "220809",
      "postDate": "09/13/2017 06:18:14",
      "content": "<p>11G</p>",
      "rawMarkdown": "11G",
      "votes": null
    },
    {
      "id": "221121",
      "postDate": "09/14/2017 07:03:55",
      "content": "<p>Has anyone tried BR layer instead of BN layer? It seemed that it performed better when using small batch size.</p>",
      "rawMarkdown": "Has anyone tried BR layer instead of BN layer? It seemed that it performed better when using small batch size.",
      "votes": null
    },
    {
      "id": "221150",
      "postDate": "09/14/2017 10:13:17",
      "content": "<p>I tried the BR layer. It trains better than BN but is more computationally intensive. Training time went from 295 s/epoch to 360 s/epoch. Not sure if it gives a better solution though, as I have not trained it till convergence.</p>",
      "rawMarkdown": "I tried the BR layer. It trains better than BN but is more computationally intensive. Training time went from 295 s/epoch to 360 s/epoch. Not sure if it gives a better solution though, as I have not trained it till convergence.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 220104,
      "author_name": "malcolmsun",
      "author_url": "",
      "post_date": "09/11/2017 07:26:38",
      "content": "<p>Hi, I think bigger input size really matters. Besides , appropriate batch size is important since there are several batch norm operation in U-net. One question, Peter's code does not provide a network for input with 1280x1280, so you code a new network by yourself?</p>",
      "votes": null,
      "replies": [
        {
          "id": 220106,
          "author_name": "zhangsongwei",
          "author_url": "",
          "post_date": "09/11/2017 07:31:17",
          "content": "<p>Yes, but I just change the input size.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 220170,
      "author_name": "zhihang",
      "author_url": "",
      "post_date": "09/11/2017 11:58:34",
      "content": "<p>how about the memory size of your GPU?</p>",
      "votes": null,
      "replies": [
        {
          "id": 220809,
          "author_name": "zhangsongwei",
          "author_url": "",
          "post_date": "09/13/2017 06:18:14",
          "content": "<p>11G</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 221121,
      "author_name": "zhangsongwei",
      "author_url": "",
      "post_date": "09/14/2017 07:03:55",
      "content": "<p>Has anyone tried BR layer instead of BN layer? It seemed that it performed better when using small batch size.</p>",
      "votes": null,
      "replies": [
        {
          "id": 221150,
          "author_name": "mitiau",
          "author_url": "",
          "post_date": "09/14/2017 10:13:17",
          "content": "<p>I tried the BR layer. It trains better than BN but is more computationally intensive. Training time went from 295 s/epoch to 360 s/epoch. Not sure if it gives a better solution though, as I have not trained it till convergence.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "219393": "input size: 1024x1024   batch size : 4   dice: 0.9956  LB score: 0.9962\n - input size: 1280x1280   batch size : 4   dice: 0.9963  LB score: 0.9965\n - input size: 1280x1920   batch size : 2   dice: 0.9962  LB score: 0.9966\n\nI am using u-net with Keras according to Peter's code.\nIt seemed that more batch size performed better.\nWhat is your input size or batch size?",
    "220104": "Hi, I think bigger input size really matters. Besides , appropriate batch size is important since there are several batch norm operation in U-net. One question, Peter's code does not provide a network for input with 1280x1280, so you code a new network by yourself?",
    "220106": "Yes, but I just change the input size.",
    "220170": "how about the memory size of your GPU?",
    "220809": "11G",
    "221121": "Has anyone tried BR layer instead of BN layer? It seemed that it performed better when using small batch size.",
    "221150": "I tried the BR layer. It trains better than BN but is more computationally intensive. Training time went from 295 s/epoch to 360 s/epoch. Not sure if it gives a better solution though, as I have not trained it till convergence."
  },
  "source": "meta"
}