{
  "id": 183456,
  "title": "TPUs Never Die 😈",
  "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/183456",
  "author_name": "Sarthak khandelwal",
  "post_date": "2020-09-16T17:49:22.007000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Its really beneficial when we have resources that we can train our model with. Anyone in the entire machine learning stack would feel a need to have a good amount of resources to work with. And here on kaggle, we are provided with <strong>TPUs</strong> which helps making computations faster and memory efficient so that we can train larger model architectures for comparatively large <strong>batch sizes</strong>.  <strong>TPUs</strong> also allow us to attach and detach data to it very efficiently.</p>\n<p>In this scenario too we feel a need to work with <strong>TPU</strong> to train our model on the <strong>Lyft Dataset</strong>. For such a big dataset, it would be very beneficial if we're working with <strong>TPU</strong> to prepare and train our model.</p>\n<p>Hence I decided to prepare a <a href=\"https://www.kaggle.com/forwet/tpus-never-die\" target=\"_blank\">notebook</a> for setting up the <strong>TPU</strong> with <strong>Pytorch</strong> and working on training of a model with it. I tried my best in describing the things that I knew about <strong>TPU</strong> and <strong>Pytorch</strong>.</p>\n<p>I recently have started working with <strong>Pytorch</strong> specifically for this competition and hence there could be a possibility of loopholes or unstructured code in my notebook and hence please feel free to correct me out there. Thanks!</p>",
  "messages": [
    {
      "id": 1013464,
      "postDate": "2020-09-16T17:49:22.007Z",
      "content": "<p>Its really beneficial when we have resources that we can train our model with. Anyone in the entire machine learning stack would feel a need to have a good amount of resources to work with. And here on kaggle, we are provided with <strong>TPUs</strong> which helps making computations faster and memory efficient so that we can train larger model architectures for comparatively large <strong>batch sizes</strong>.  <strong>TPUs</strong> also allow us to attach and detach data to it very efficiently.</p>\n<p>In this scenario too we feel a need to work with <strong>TPU</strong> to train our model on the <strong>Lyft Dataset</strong>. For such a big dataset, it would be very beneficial if we're working with <strong>TPU</strong> to prepare and train our model.</p>\n<p>Hence I decided to prepare a <a href=\"https://www.kaggle.com/forwet/tpus-never-die\" target=\"_blank\">notebook</a> for setting up the <strong>TPU</strong> with <strong>Pytorch</strong> and working on training of a model with it. I tried my best in describing the things that I knew about <strong>TPU</strong> and <strong>Pytorch</strong>.</p>\n<p>I recently have started working with <strong>Pytorch</strong> specifically for this competition and hence there could be a possibility of loopholes or unstructured code in my notebook and hence please feel free to correct me out there. Thanks!</p>",
      "rawMarkdown": "Its really beneficial when we have resources that we can train our model with. Anyone in the entire machine learning stack would feel a need to have a good amount of resources to work with. And here on kaggle, we are provided with **TPUs** which helps making computations faster and memory efficient so that we can train larger model architectures for comparatively large **batch sizes**.  **TPUs** also allow us to attach and detach data to it very efficiently.\n\nIn this scenario too we feel a need to work with **TPU** to train our model on the **Lyft Dataset**. For such a big dataset, it would be very beneficial if we're working with **TPU** to prepare and train our model.\n\nHence I decided to prepare a [notebook](https://www.kaggle.com/forwet/tpus-never-die) for setting up the **TPU** with **Pytorch** and working on training of a model with it. I tried my best in describing the things that I knew about **TPU** and **Pytorch**.\n\nI recently have started working with **Pytorch** specifically for this competition and hence there could be a possibility of loopholes or unstructured code in my notebook and hence please feel free to correct me out there. Thanks!",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1013464": "Its really beneficial when we have resources that we can train our model with. Anyone in the entire machine learning stack would feel a need to have a good amount of resources to work with. And here on kaggle, we are provided with **TPUs** which helps making computations faster and memory efficient so that we can train larger model architectures for comparatively large **batch sizes**.  **TPUs** also allow us to attach and detach data to it very efficiently.\n\nIn this scenario too we feel a need to work with **TPU** to train our model on the **Lyft Dataset**. For such a big dataset, it would be very beneficial if we're working with **TPU** to prepare and train our model.\n\nHence I decided to prepare a [notebook](https://www.kaggle.com/forwet/tpus-never-die) for setting up the **TPU** with **Pytorch** and working on training of a model with it. I tried my best in describing the things that I knew about **TPU** and **Pytorch**.\n\nI recently have started working with **Pytorch** specifically for this competition and hence there could be a possibility of loopholes or unstructured code in my notebook and hence please feel free to correct me out there. Thanks!"
  }
}