{
  "id": 396363,
  "title": "Load 100+ training files on Kaggle [TPU VM] ",
  "url": "/competitions/icecube-neutrinos-in-deep-ice/discussion/396363",
  "author_name": "",
  "post_date": "2023-03-21T10:01:07.375629100Z",
  "votes": 15,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I was playing around with an excellent <a href=\"https://www.kaggle.com/code/rsmits/tensorflow-lstm-model-training-tpu\" target=\"_blank\">baseline</a> shared by <a href=\"https://www.kaggle.com/rsmits\" target=\"_blank\">@rsmits</a> </p>\n<p>I think the drop in training performance (~1%) on Kaggle TPUs than the model they shared was due to using training files in chunks. So, here is the easiest way to load as many batches as you want at once. </p>\n<p>TPU VM machine provides 330 GB Ram 👀✨</p>\n<p><strong>Step 1</strong>: Switch from TPU v3.8 ---&gt; TPU VM v3.8 instance<br>\n<strong>Step 2</strong>: <code>!pip install /lib/wheels/tensorflow-2.9.1-cp38-cp38-linux_x86_64.whl</code><br>\n<strong>Step 3</strong>: Increase the parameter <code>train_files_delta</code> to a larger number</p>\n<p><strong>I was able to reach CV ~1.01x with that</strong></p>",
  "messages": [
    {
      "id": "2190524",
      "postDate": "03/21/2023 10:01:07",
      "content": "<p>I was playing around with an excellent <a href=\"https://www.kaggle.com/code/rsmits/tensorflow-lstm-model-training-tpu\" target=\"_blank\">baseline</a> shared by <a href=\"https://www.kaggle.com/rsmits\" target=\"_blank\">@rsmits</a> </p>\n<p>I think the drop in training performance (~1%) on Kaggle TPUs than the model they shared was due to using training files in chunks. So, here is the easiest way to load as many batches as you want at once. </p>\n<p>TPU VM machine provides 330 GB Ram 👀✨</p>\n<p><strong>Step 1</strong>: Switch from TPU v3.8 ---&gt; TPU VM v3.8 instance<br>\n<strong>Step 2</strong>: <code>!pip install /lib/wheels/tensorflow-2.9.1-cp38-cp38-linux_x86_64.whl</code><br>\n<strong>Step 3</strong>: Increase the parameter <code>train_files_delta</code> to a larger number</p>\n<p><strong>I was able to reach CV ~1.01x with that</strong></p>",
      "rawMarkdown": "I was playing around with an excellent [baseline](https://www.kaggle.com/code/rsmits/tensorflow-lstm-model-training-tpu) shared by @rsmits \n\nI think the drop in training performance (~1%) on Kaggle TPUs than the model they shared was due to using training files in chunks. So, here is the easiest way to load as many batches as you want at once. \n\nTPU VM machine provides 330 GB Ram 👀✨\n\n**Step 1**: Switch from TPU v3.8 ---> TPU VM v3.8 instance\n**Step 2**: `!pip install /lib/wheels/tensorflow-2.9.1-cp38-cp38-linux_x86_64.whl`\n**Step 3**: Increase the parameter `train_files_delta` to a larger number\n\n**I was able to reach CV ~1.01x with that**",
      "votes": null
    },
    {
      "id": "2201298",
      "postDate": "03/29/2023 07:46:08",
      "content": "<p>thx, I'm learning this👀</p>",
      "rawMarkdown": "thx, I'm learning this👀",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2201298,
      "author_name": "mjx000",
      "author_url": "",
      "post_date": "03/29/2023 07:46:08",
      "content": "<p>thx, I'm learning this👀</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2190524": "I was playing around with an excellent [baseline](https://www.kaggle.com/code/rsmits/tensorflow-lstm-model-training-tpu) shared by @rsmits \n\nI think the drop in training performance (~1%) on Kaggle TPUs than the model they shared was due to using training files in chunks. So, here is the easiest way to load as many batches as you want at once. \n\nTPU VM machine provides 330 GB Ram 👀✨\n\n**Step 1**: Switch from TPU v3.8 ---> TPU VM v3.8 instance\n**Step 2**: `!pip install /lib/wheels/tensorflow-2.9.1-cp38-cp38-linux_x86_64.whl`\n**Step 3**: Increase the parameter `train_files_delta` to a larger number\n\n**I was able to reach CV ~1.01x with that**",
    "2201298": "thx, I'm learning this👀"
  },
  "source": "meta"
}