{
  "id": 362305,
  "title": "Was anybody able to speed up training in keras?",
  "url": "/competitions/tabular-playground-series-oct-2022/discussion/362305",
  "author_name": "",
  "post_date": "2022-10-26T15:40:22.216809500Z",
  "votes": 6,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi Folks!</p>\n<p>I spent a significant amount of time translating the code from<a href=\"https://www.kaggle.com/code/paddykb/tps-2022-10-fastai\" target=\"_blank\">TPS-2022-10 Fastai</a> into a Keras/Tensorflow dataset, and at the best performance it takes around 2 minutes per epoch.</p>\n<p>I've been wondering if there is an additional abstraction on these frameworks that results in the performance drop.</p>",
  "messages": [
    {
      "id": "2004922",
      "postDate": "10/26/2022 15:40:22",
      "content": "<p>Hi Folks!</p>\n<p>I spent a significant amount of time translating the code from<a href=\"https://www.kaggle.com/code/paddykb/tps-2022-10-fastai\" target=\"_blank\">TPS-2022-10 Fastai</a> into a Keras/Tensorflow dataset, and at the best performance it takes around 2 minutes per epoch.</p>\n<p>I've been wondering if there is an additional abstraction on these frameworks that results in the performance drop.</p>",
      "rawMarkdown": "Hi Folks!\n\nI spent a significant amount of time translating the code from[TPS-2022-10 Fastai](https://www.kaggle.com/code/paddykb/tps-2022-10-fastai) into a Keras/Tensorflow dataset, and at the best performance it takes around 2 minutes per epoch.\n\nI've been wondering if there is an additional abstraction on these frameworks that results in the performance drop.",
      "votes": null
    },
    {
      "id": "2005853",
      "postDate": "10/27/2022 09:21:40",
      "content": "<p>I noticed this too. I assumed the difference is due to using NumPy vs torch for augmentation. Most of the time is wasted moving data around.</p>\n<p>I did consider using cython to speed this up a bit - but for trying out features, and architectures, I only used 10% of the data - which is quick enough for experimenting.</p>",
      "rawMarkdown": "I noticed this too. I assumed the difference is due to using NumPy vs torch for augmentation. Most of the time is wasted moving data around.\n\nI did consider using cython to speed this up a bit - but for trying out features, and architectures, I only used 10% of the data - which is quick enough for experimenting.",
      "votes": null
    },
    {
      "id": "2006391",
      "postDate": "10/27/2022 15:06:11",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/paddykb\" target=\"_blank\">@paddykb</a> you made this competition much easier. So probably this is the right opportunity to learn about new frameworks and their advantages.</p>",
      "rawMarkdown": "Thanks @paddykb you made this competition much easier. So probably this is the right opportunity to learn about new frameworks and their advantages.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2005853,
      "author_name": "paddykb",
      "author_url": "",
      "post_date": "10/27/2022 09:21:40",
      "content": "<p>I noticed this too. I assumed the difference is due to using NumPy vs torch for augmentation. Most of the time is wasted moving data around.</p>\n<p>I did consider using cython to speed this up a bit - but for trying out features, and architectures, I only used 10% of the data - which is quick enough for experimenting.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2006391,
          "author_name": "jcaliz",
          "author_url": "",
          "post_date": "10/27/2022 15:06:11",
          "content": "<p>Thanks <a href=\"https://www.kaggle.com/paddykb\" target=\"_blank\">@paddykb</a> you made this competition much easier. So probably this is the right opportunity to learn about new frameworks and their advantages.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2004922": "Hi Folks!\n\nI spent a significant amount of time translating the code from[TPS-2022-10 Fastai](https://www.kaggle.com/code/paddykb/tps-2022-10-fastai) into a Keras/Tensorflow dataset, and at the best performance it takes around 2 minutes per epoch.\n\nI've been wondering if there is an additional abstraction on these frameworks that results in the performance drop.",
    "2005853": "I noticed this too. I assumed the difference is due to using NumPy vs torch for augmentation. Most of the time is wasted moving data around.\n\nI did consider using cython to speed this up a bit - but for trying out features, and architectures, I only used 10% of the data - which is quick enough for experimenting.",
    "2006391": "Thanks @paddykb you made this competition much easier. So probably this is the right opportunity to learn about new frameworks and their advantages."
  },
  "source": "meta"
}