{
  "id": 307393,
  "title": "gpu train the competition model ，when the image size is bigger than 512，the gpu is out of memory  ",
  "url": "/competitions/happy-whale-and-dolphin/discussion/307393",
  "author_name": "",
  "post_date": "2022-02-14T03:37:50.157488600Z",
  "votes": 3,
  "comment_count": 6,
  "views": 0,
  "content": "<p>gpu train the competition model ，when the image size is bigger than 512 ，the gpu is out of memory ；but if the image size is too small，the score is very lower；<br>\nif use tpu to train，but the kaggle tpu hours is not enough to train </p>",
  "messages": [
    {
      "id": "1689110",
      "postDate": "02/14/2022 03:37:50",
      "content": "<p>gpu train the competition model ，when the image size is bigger than 512 ，the gpu is out of memory ；but if the image size is too small，the score is very lower；<br>\nif use tpu to train，but the kaggle tpu hours is not enough to train </p>",
      "rawMarkdown": "gpu train the competition model ，when the image size is bigger than 512 ，the gpu is out of memory ；but if the image size is too small，the score is very lower；\nif use tpu to train，but the kaggle tpu hours is not enough to train",
      "votes": null
    },
    {
      "id": "1689240",
      "postDate": "02/14/2022 05:06:57",
      "content": "<p>how you get current model then?</p>",
      "rawMarkdown": "how you get current model then?",
      "votes": null
    },
    {
      "id": "1689338",
      "postDate": "02/14/2022 06:54:51",
      "content": "<p>use kaggle free tpu only train a single model of single fold</p>",
      "rawMarkdown": "use kaggle free tpu only train a single model of single fold",
      "votes": null
    },
    {
      "id": "1689447",
      "postDate": "02/14/2022 08:36:13",
      "content": "<p>kaggle TPUv3-8.   My colab pro only support batch size=4 for B6 model.</p>\n<p>free colab TPU, you can try.</p>",
      "rawMarkdown": "kaggle TPUv3-8.   My colab pro only support batch size=4 for B6 model.\n\nfree colab TPU, you can try.",
      "votes": null
    },
    {
      "id": "1689456",
      "postDate": "02/14/2022 08:45:20",
      "content": "<p><a href=\"https://www.kaggle.com/dragonzhang\" target=\"_blank\">@dragonzhang</a> Many people are finding B5 with higher resolutions, have you tried that?</p>",
      "rawMarkdown": "dragonzhang Many people are finding B5 with higher resolutions, have you tried that?",
      "votes": null
    },
    {
      "id": "1689536",
      "postDate": "02/14/2022 09:59:24",
      "content": "<p>I just enter today.  may try later.  small model with bigger batch size, perhaps performance better than big model with limited hardware resource. </p>",
      "rawMarkdown": "I just enter today.  may try later.  small model with bigger batch size, perhaps performance better than big model with limited hardware resource.",
      "votes": null
    },
    {
      "id": "1696203",
      "postDate": "02/18/2022 16:50:32",
      "content": "<p><a href=\"https://www.kaggle.com/dragonzhang\" target=\"_blank\">@dragonzhang</a> Do you use colab pro? May I ask you what kind of TPU does colab provide for you? I am planning to upgrade my colab, but I am not sure if it will fit my needs. Unfortunately, I can't google any meaningful TPU specs for colab…</p>",
      "rawMarkdown": "dragonzhang Do you use colab pro? May I ask you what kind of TPU does colab provide for you? I am planning to upgrade my colab, but I am not sure if it will fit my needs. Unfortunately, I can't google any meaningful TPU specs for colab...",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1689240,
      "author_name": "dragonzhang",
      "author_url": "",
      "post_date": "02/14/2022 05:06:57",
      "content": "<p>how you get current model then?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1689338,
          "author_name": "guochangzhang",
          "author_url": "",
          "post_date": "02/14/2022 06:54:51",
          "content": "<p>use kaggle free tpu only train a single model of single fold</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1689447,
          "author_name": "dragonzhang",
          "author_url": "",
          "post_date": "02/14/2022 08:36:13",
          "content": "<p>kaggle TPUv3-8.   My colab pro only support batch size=4 for B6 model.</p>\n<p>free colab TPU, you can try.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1689456,
          "author_name": "init27",
          "author_url": "",
          "post_date": "02/14/2022 08:45:20",
          "content": "<p><a href=\"https://www.kaggle.com/dragonzhang\" target=\"_blank\">@dragonzhang</a> Many people are finding B5 with higher resolutions, have you tried that?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1689536,
          "author_name": "dragonzhang",
          "author_url": "",
          "post_date": "02/14/2022 09:59:24",
          "content": "<p>I just enter today.  may try later.  small model with bigger batch size, perhaps performance better than big model with limited hardware resource. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1696203,
          "author_name": "meowmeowmeowmeowmeow",
          "author_url": "",
          "post_date": "02/18/2022 16:50:32",
          "content": "<p><a href=\"https://www.kaggle.com/dragonzhang\" target=\"_blank\">@dragonzhang</a> Do you use colab pro? May I ask you what kind of TPU does colab provide for you? I am planning to upgrade my colab, but I am not sure if it will fit my needs. Unfortunately, I can't google any meaningful TPU specs for colab…</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1689110": "gpu train the competition model ，when the image size is bigger than 512 ，the gpu is out of memory ；but if the image size is too small，the score is very lower；\nif use tpu to train，but the kaggle tpu hours is not enough to train",
    "1689240": "how you get current model then?",
    "1689338": "use kaggle free tpu only train a single model of single fold",
    "1689447": "kaggle TPUv3-8.   My colab pro only support batch size=4 for B6 model.\n\nfree colab TPU, you can try.",
    "1689456": "dragonzhang Many people are finding B5 with higher resolutions, have you tried that?",
    "1689536": "I just enter today.  may try later.  small model with bigger batch size, perhaps performance better than big model with limited hardware resource.",
    "1696203": "dragonzhang Do you use colab pro? May I ask you what kind of TPU does colab provide for you? I am planning to upgrade my colab, but I am not sure if it will fit my needs. Unfortunately, I can't google any meaningful TPU specs for colab..."
  },
  "source": "meta"
}