{
  "id": 135471,
  "title": "Is there any suitable network architecture for TPU?",
  "url": "/competitions/flower-classification-with-tpus/discussion/135471",
  "author_name": "",
  "post_date": "2020-03-14T02:14:19.832752900Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Both accuracy-wise and training/inference speed-wise.\nI'm not familiar with TPU, but I've noticed that people tend to use EfficientNet for TPU.\nOf course it seems natural choice because EfficientNet is created by google, but is there any theoretical or practical reason to choose EfficientNet for TPU?\nDoes it perform relatively better than SeResNeXt on TPU??</p>",
  "messages": [
    {
      "id": "771312",
      "postDate": "03/14/2020 02:14:19",
      "content": "<p>Both accuracy-wise and training/inference speed-wise.\nI'm not familiar with TPU, but I've noticed that people tend to use EfficientNet for TPU.\nOf course it seems natural choice because EfficientNet is created by google, but is there any theoretical or practical reason to choose EfficientNet for TPU?\nDoes it perform relatively better than SeResNeXt on TPU??</p>",
      "rawMarkdown": "Both accuracy-wise and training/inference speed-wise.\nI'm not familiar with TPU, but I've noticed that people tend to use EfficientNet for TPU.\nOf course it seems natural choice because EfficientNet is created by google, but is there any theoretical or practical reason to choose EfficientNet for TPU?\nDoes it perform relatively better than SeResNeXt on TPU??",
      "votes": null
    },
    {
      "id": "775674",
      "postDate": "03/16/2020 23:52:30",
      "content": "<p>I think people in this competition were shooting for best classification performance. The TPU is just an accelerator. It makes training go faster. Models should be portable between TPU and GPU and I am not aware of any model that would be designed for TPUs specifically.</p>",
      "rawMarkdown": "I think people in this competition were shooting for best classification performance. The TPU is just an accelerator. It makes training go faster. Models should be portable between TPU and GPU and I am not aware of any model that would be designed for TPUs specifically.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 775674,
      "author_name": "mgorner",
      "author_url": "",
      "post_date": "03/16/2020 23:52:30",
      "content": "<p>I think people in this competition were shooting for best classification performance. The TPU is just an accelerator. It makes training go faster. Models should be portable between TPU and GPU and I am not aware of any model that would be designed for TPUs specifically.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "771312": "Both accuracy-wise and training/inference speed-wise.\nI'm not familiar with TPU, but I've noticed that people tend to use EfficientNet for TPU.\nOf course it seems natural choice because EfficientNet is created by google, but is there any theoretical or practical reason to choose EfficientNet for TPU?\nDoes it perform relatively better than SeResNeXt on TPU??",
    "775674": "I think people in this competition were shooting for best classification performance. The TPU is just an accelerator. It makes training go faster. Models should be portable between TPU and GPU and I am not aware of any model that would be designed for TPUs specifically."
  },
  "source": "meta"
}