{
  "id": 310056,
  "title": "Why do most open source code use tf in this competition?",
  "url": "/competitions/happy-whale-and-dolphin/discussion/310056",
  "author_name": "",
  "post_date": "2022-02-27T11:07:17.324454300Z",
  "votes": 3,
  "comment_count": 7,
  "views": 0,
  "content": "<p>To use tpu?</p>",
  "messages": [
    {
      "id": "1706289",
      "postDate": "02/27/2022 11:07:17",
      "content": "<p>To use tpu?</p>",
      "rawMarkdown": "To use tpu?",
      "votes": null
    },
    {
      "id": "1706358",
      "postDate": "02/27/2022 12:26:41",
      "content": "<p>I think that It depends on hardware.  </p>\n<p>In general,  TF/TPU seems much faster than  pytorch.</p>\n<p>However Pytorch users do leading top on LB.</p>",
      "rawMarkdown": "I think that It depends on hardware.  \n\nIn general,  TF/TPU seems much faster than  pytorch.\n\nHowever Pytorch users do leading top on LB.",
      "votes": null
    },
    {
      "id": "1706392",
      "postDate": "02/27/2022 13:14:23",
      "content": "<p>I would say that TensorFlow is better to fit big models on TPU and use bigger images. However, I believe that with PyTorch you are much more in control of what you're doing.</p>",
      "rawMarkdown": "I would say that TensorFlow is better to fit big models on TPU and use bigger images. However, I believe that with PyTorch you are much more in control of what you're doing.",
      "votes": null
    },
    {
      "id": "1706418",
      "postDate": "02/27/2022 13:35:14",
      "content": "<p>I believe I replicated the best TF TPU kernel in Pytorch. I even added more tricks but I can't beat its score on LB.</p>",
      "rawMarkdown": "I believe I replicated the best TF TPU kernel in Pytorch. I even added more tricks but I can't beat its score on LB.",
      "votes": null
    },
    {
      "id": "1707615",
      "postDate": "02/28/2022 16:03:25",
      "content": "<p>Thanks for your reply. </p>",
      "rawMarkdown": "Thanks for your reply.",
      "votes": null
    },
    {
      "id": "1707623",
      "postDate": "02/28/2022 16:08:33",
      "content": "<p>Thanks for your reply. Maybe I should learn tf again. 😂</p>",
      "rawMarkdown": "Thanks for your reply. Maybe I should learn tf again. 😂",
      "votes": null
    },
    {
      "id": "1707624",
      "postDate": "02/28/2022 16:10:32",
      "content": "<p>Thanks for your reply. I wish you a good score on LB. 😊</p>",
      "rawMarkdown": "Thanks for your reply. I wish you a good score on LB. 😊",
      "votes": null
    },
    {
      "id": "1713306",
      "postDate": "03/05/2022 21:25:22",
      "content": "<p>It looks like TPU usage is the main reason indeed. It seems that some are reporting slight difference between TF/Keras with TPU vs similar code using PyTorch and <a href=\"https://github.com/pytorch/xla\" target=\"_blank\">XLA</a>. If anyone knows some of the reasons, it would be great. 👌</p>",
      "rawMarkdown": "It looks like TPU usage is the main reason indeed. It seems that some are reporting slight difference between TF/Keras with TPU vs similar code using PyTorch and [XLA](https://github.com/pytorch/xla). If anyone knows some of the reasons, it would be great. 👌",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1706358,
      "author_name": "dragonzhang",
      "author_url": "",
      "post_date": "02/27/2022 12:26:41",
      "content": "<p>I think that It depends on hardware.  </p>\n<p>In general,  TF/TPU seems much faster than  pytorch.</p>\n<p>However Pytorch users do leading top on LB.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1707615,
          "author_name": "yangranran",
          "author_url": "",
          "post_date": "02/28/2022 16:03:25",
          "content": "<p>Thanks for your reply. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1706392,
      "author_name": "wolfy73",
      "author_url": "",
      "post_date": "02/27/2022 13:14:23",
      "content": "<p>I would say that TensorFlow is better to fit big models on TPU and use bigger images. However, I believe that with PyTorch you are much more in control of what you're doing.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1707623,
          "author_name": "yangranran",
          "author_url": "",
          "post_date": "02/28/2022 16:08:33",
          "content": "<p>Thanks for your reply. Maybe I should learn tf again. 😂</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1706418,
      "author_name": "aerdem4",
      "author_url": "",
      "post_date": "02/27/2022 13:35:14",
      "content": "<p>I believe I replicated the best TF TPU kernel in Pytorch. I even added more tricks but I can't beat its score on LB.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1707624,
          "author_name": "yangranran",
          "author_url": "",
          "post_date": "02/28/2022 16:10:32",
          "content": "<p>Thanks for your reply. I wish you a good score on LB. 😊</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1713306,
      "author_name": "yassinealouini",
      "author_url": "",
      "post_date": "03/05/2022 21:25:22",
      "content": "<p>It looks like TPU usage is the main reason indeed. It seems that some are reporting slight difference between TF/Keras with TPU vs similar code using PyTorch and <a href=\"https://github.com/pytorch/xla\" target=\"_blank\">XLA</a>. If anyone knows some of the reasons, it would be great. 👌</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1706289": "To use tpu?",
    "1706358": "I think that It depends on hardware.  \n\nIn general,  TF/TPU seems much faster than  pytorch.\n\nHowever Pytorch users do leading top on LB.",
    "1706392": "I would say that TensorFlow is better to fit big models on TPU and use bigger images. However, I believe that with PyTorch you are much more in control of what you're doing.",
    "1706418": "I believe I replicated the best TF TPU kernel in Pytorch. I even added more tricks but I can't beat its score on LB.",
    "1707615": "Thanks for your reply.",
    "1707623": "Thanks for your reply. Maybe I should learn tf again. 😂",
    "1707624": "Thanks for your reply. I wish you a good score on LB. 😊",
    "1713306": "It looks like TPU usage is the main reason indeed. It seems that some are reporting slight difference between TF/Keras with TPU vs similar code using PyTorch and [XLA](https://github.com/pytorch/xla). If anyone knows some of the reasons, it would be great. 👌"
  },
  "source": "meta"
}