{
  "id": 70750,
  "title": "Keras vs Pytorch?",
  "url": "/competitions/airbus-ship-detection/discussion/70750",
  "author_name": "",
  "post_date": "2018-11-07T01:52:32.010702200Z",
  "votes": 5,
  "comment_count": 4,
  "views": 0,
  "content": "<p>For some reason, everyone is using pytorch now...Why?</p>",
  "messages": [
    {
      "id": "416621",
      "postDate": "11/07/2018 01:52:32",
      "content": "<p>For some reason, everyone is using pytorch now...Why?</p>",
      "rawMarkdown": "For some reason, everyone is using pytorch now...Why?",
      "votes": null
    },
    {
      "id": "416869",
      "postDate": "11/07/2018 11:35:48",
      "content": "<p>I can't speak for everyone, but I like PyTorch for the transparency and flexibility. It's like you're working with numpy on steroids that can learn and work on GPU.</p>",
      "rawMarkdown": "I can't speak for everyone, but I like PyTorch for the transparency and flexibility. It's like you're working with numpy on steroids that can learn and work on GPU.",
      "votes": null
    },
    {
      "id": "416916",
      "postDate": "11/07/2018 12:36:34",
      "content": "<p>I have better performance with Pytorch, and it handles multi GPU better than Keras</p>",
      "rawMarkdown": "I have better performance with Pytorch, and it handles multi GPU better than Keras",
      "votes": null
    },
    {
      "id": "416997",
      "postDate": "11/07/2018 15:32:57",
      "content": "<p>I like the flexibility provided by Pytorch and its speed. Dynamic computational graph building is really a great thing, especially for recursive nets, when a tree like structure is build during training or inference depending on the input. The only thing why there can be some drawbacks with Pytorch, and fast.ai in particular, is that those libraries are quite new and the the community of people who are using it is not very big now. So one can have hard time to find an implementation of a particular network and pretrained weights for it. But it gets more popular.</p>",
      "rawMarkdown": "I like the flexibility provided by Pytorch and its speed. Dynamic computational graph building is really a great thing, especially for recursive nets, when a tree like structure is build during training or inference depending on the input. The only thing why there can be some drawbacks with Pytorch, and fast.ai in particular, is that those libraries are quite new and the the community of people who are using it is not very big now. So one can have hard time to find an implementation of a particular network and pretrained weights for it. But it gets more popular.",
      "votes": null
    },
    {
      "id": "727674",
      "postDate": "01/24/2020 00:37:28",
      "content": "<p>I was asking myself the same question. Thanks!</p>",
      "rawMarkdown": "I was asking myself the same question. Thanks!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 416869,
      "author_name": "justuser",
      "author_url": "",
      "post_date": "11/07/2018 11:35:48",
      "content": "<p>I can't speak for everyone, but I like PyTorch for the transparency and flexibility. It's like you're working with numpy on steroids that can learn and work on GPU.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 416916,
      "author_name": "mxdbld",
      "author_url": "",
      "post_date": "11/07/2018 12:36:34",
      "content": "<p>I have better performance with Pytorch, and it handles multi GPU better than Keras</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 416997,
      "author_name": "iafoss",
      "author_url": "",
      "post_date": "11/07/2018 15:32:57",
      "content": "<p>I like the flexibility provided by Pytorch and its speed. Dynamic computational graph building is really a great thing, especially for recursive nets, when a tree like structure is build during training or inference depending on the input. The only thing why there can be some drawbacks with Pytorch, and fast.ai in particular, is that those libraries are quite new and the the community of people who are using it is not very big now. So one can have hard time to find an implementation of a particular network and pretrained weights for it. But it gets more popular.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 727674,
      "author_name": "anacamargos11",
      "author_url": "",
      "post_date": "01/24/2020 00:37:28",
      "content": "<p>I was asking myself the same question. Thanks!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "416621": "For some reason, everyone is using pytorch now...Why?",
    "416869": "I can't speak for everyone, but I like PyTorch for the transparency and flexibility. It's like you're working with numpy on steroids that can learn and work on GPU.",
    "416916": "I have better performance with Pytorch, and it handles multi GPU better than Keras",
    "416997": "I like the flexibility provided by Pytorch and its speed. Dynamic computational graph building is really a great thing, especially for recursive nets, when a tree like structure is build during training or inference depending on the input. The only thing why there can be some drawbacks with Pytorch, and fast.ai in particular, is that those libraries are quite new and the the community of people who are using it is not very big now. So one can have hard time to find an implementation of a particular network and pretrained weights for it. But it gets more popular.",
    "727674": "I was asking myself the same question. Thanks!"
  },
  "source": "meta"
}