{
  "id": 252997,
  "title": "batch size or resize?",
  "url": "/competitions/g2net-gravitational-wave-detection/discussion/252997",
  "author_name": "",
  "post_date": "2021-07-14T15:36:58.865108900Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Which one is better, decrease batch size or increase size of images? Let your CV, LB, Batch, Size, Channel, fold and model?<br>\nCV= 0.8648 <br>\nLB= 0.864<br>\nBatch=32<br>\nSize=640<br>\nChannel=3 <br>\nfold=1<br>\nmodel_name='tf_efficientnetv2_b1'</p>",
  "messages": [
    {
      "id": "1388015",
      "postDate": "07/14/2021 15:36:58",
      "content": "<p>Which one is better, decrease batch size or increase size of images? Let your CV, LB, Batch, Size, Channel, fold and model?<br>\nCV= 0.8648 <br>\nLB= 0.864<br>\nBatch=32<br>\nSize=640<br>\nChannel=3 <br>\nfold=1<br>\nmodel_name='tf_efficientnetv2_b1'</p>",
      "rawMarkdown": "Which one is better, decrease batch size or increase size of images? Let your CV, LB, Batch, Size, Channel, fold and model?\nCV= 0.8648 \nLB= 0.864\nBatch=32\nSize=640\nChannel=3 \nfold=1\nmodel_name='tf_efficientnetv2_b1'",
      "votes": null
    },
    {
      "id": "1390653",
      "postDate": "07/16/2021 23:47:43",
      "content": "<p>higher batches = better generalization<br>\nless quality data = worse results</p>",
      "rawMarkdown": "higher batches = better generalization\nless quality data = worse results",
      "votes": null
    },
    {
      "id": "1395064",
      "postDate": "07/20/2021 21:24:33",
      "content": "<p>Generally spoken a higher image size will mostly result into better scores this is what the EfficientNet paper states and on what their model is build upon. </p>\n<p>When it comes to batchsizes there is no real trend, some scientist claim that higher is better due to better computed gradients but others have proven in papers e.g. <a href=\"https://arxiv.org/pdf/1804.07612.pdf\" target=\"_blank\">https://arxiv.org/pdf/1804.07612.pdf</a> that lower batchsize e.g. 32 is good enough for a lot of tasks. For sure a very low batchsize of e.g. 1 updates the learning rate with worse gradients everytime you process an image. You will have to find it out by experimenting a little.</p>",
      "rawMarkdown": "Generally spoken a higher image size will mostly result into better scores this is what the EfficientNet paper states and on what their model is build upon. \n\nWhen it comes to batchsizes there is no real trend, some scientist claim that higher is better due to better computed gradients but others have proven in papers e.g. https://arxiv.org/pdf/1804.07612.pdf that lower batchsize e.g. 32 is good enough for a lot of tasks. For sure a very low batchsize of e.g. 1 updates the learning rate with worse gradients everytime you process an image. You will have to find it out by experimenting a little.",
      "votes": null
    },
    {
      "id": "1561276",
      "postDate": "10/27/2021 13:11:44",
      "content": "<p>Hey All,</p>\n<p>Thank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey <a href=\"https://forms.gle/QP9L16niPexozyhu5\" target=\"_blank\">https://forms.gle/QP9L16niPexozyhu5</a>.</p>\n<p>Thank you all,</p>\n<p>Regards,<br>\nChris</p>",
      "rawMarkdown": "Hey All,\n\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.\n\nThank you all,\n\nRegards,\nChris",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1390653,
      "author_name": "eladwar",
      "author_url": "",
      "post_date": "07/16/2021 23:47:43",
      "content": "<p>higher batches = better generalization<br>\nless quality data = worse results</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1395064,
      "author_name": "aliabdin1",
      "author_url": "",
      "post_date": "07/20/2021 21:24:33",
      "content": "<p>Generally spoken a higher image size will mostly result into better scores this is what the EfficientNet paper states and on what their model is build upon. </p>\n<p>When it comes to batchsizes there is no real trend, some scientist claim that higher is better due to better computed gradients but others have proven in papers e.g. <a href=\"https://arxiv.org/pdf/1804.07612.pdf\" target=\"_blank\">https://arxiv.org/pdf/1804.07612.pdf</a> that lower batchsize e.g. 32 is good enough for a lot of tasks. For sure a very low batchsize of e.g. 1 updates the learning rate with worse gradients everytime you process an image. You will have to find it out by experimenting a little.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1561276,
      "author_name": "zerafachris",
      "author_url": "",
      "post_date": "10/27/2021 13:11:44",
      "content": "<p>Hey All,</p>\n<p>Thank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey <a href=\"https://forms.gle/QP9L16niPexozyhu5\" target=\"_blank\">https://forms.gle/QP9L16niPexozyhu5</a>.</p>\n<p>Thank you all,</p>\n<p>Regards,<br>\nChris</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1388015": "Which one is better, decrease batch size or increase size of images? Let your CV, LB, Batch, Size, Channel, fold and model?\nCV= 0.8648 \nLB= 0.864\nBatch=32\nSize=640\nChannel=3 \nfold=1\nmodel_name='tf_efficientnetv2_b1'",
    "1390653": "higher batches = better generalization\nless quality data = worse results",
    "1395064": "Generally spoken a higher image size will mostly result into better scores this is what the EfficientNet paper states and on what their model is build upon. \n\nWhen it comes to batchsizes there is no real trend, some scientist claim that higher is better due to better computed gradients but others have proven in papers e.g. https://arxiv.org/pdf/1804.07612.pdf that lower batchsize e.g. 32 is good enough for a lot of tasks. For sure a very low batchsize of e.g. 1 updates the learning rate with worse gradients everytime you process an image. You will have to find it out by experimenting a little.",
    "1561276": "Hey All,\n\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.\n\nThank you all,\n\nRegards,\nChris"
  },
  "source": "meta"
}