{
  "id": 43871,
  "title": "For the Raspberry Pi track, what does it mean that the only dependency must be Tensorflow?",
  "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/43871",
  "author_name": "",
  "post_date": "2017-11-20T21:08:03.349472700Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>For the Raspberry Pi track, what does it mean that the only dependency must be Tensorflow? I think that this statement needs some clarification. Does that mean that I can only import python standard library and Tensorflow, but not numpy and other libraries? Or does it mean something else?</p>",
  "messages": [
    {
      "id": "246275",
      "postDate": "11/20/2017 21:08:03",
      "content": "<p>For the Raspberry Pi track, what does it mean that the only dependency must be Tensorflow? I think that this statement needs some clarification. Does that mean that I can only import python standard library and Tensorflow, but not numpy and other libraries? Or does it mean something else?</p>",
      "rawMarkdown": "For the Raspberry Pi track, what does it mean that the only dependency must be Tensorflow? I think that this statement needs some clarification. Does that mean that I can only import python standard library and Tensorflow, but not numpy and other libraries? Or does it mean something else?",
      "votes": null
    },
    {
      "id": "246637",
      "postDate": "11/21/2017 13:45:19",
      "content": "<p>Good point! There are two aspects to this. For the latency evaluation, the trained model must be delivered as a TensorFlow frozen GraphDef file, which will mean it can't rely on any non-TF code, since the freezing process won't work if it does.</p>\n\n<p>For training, the script shouldn't rely on any Python dependencies that are not already required by TensorFlow's pip package. Unfortunately I don't have a complete list here (and I'm away from my computer over Thanksgiving) but numpy is required by TF, so that's safe to use.</p>\n\n<p>I set up these requirements in the hope that the winning model can be turned into an official TensorFlow example. Hopefully that makes sense?</p>",
      "rawMarkdown": "Good point! There are two aspects to this. For the latency evaluation, the trained model must be delivered as a TensorFlow frozen GraphDef file, which will mean it can't rely on any non-TF code, since the freezing process won't work if it does.\n\nFor training, the script shouldn't rely on any Python dependencies that are not already required by TensorFlow's pip package. Unfortunately I don't have a complete list here (and I'm away from my computer over Thanksgiving) but numpy is required by TF, so that's safe to use.\n\nI set up these requirements in the hope that the winning model can be turned into an official TensorFlow example. Hopefully that makes sense?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 246637,
      "author_name": "petewarden",
      "author_url": "",
      "post_date": "11/21/2017 13:45:19",
      "content": "<p>Good point! There are two aspects to this. For the latency evaluation, the trained model must be delivered as a TensorFlow frozen GraphDef file, which will mean it can't rely on any non-TF code, since the freezing process won't work if it does.</p>\n\n<p>For training, the script shouldn't rely on any Python dependencies that are not already required by TensorFlow's pip package. Unfortunately I don't have a complete list here (and I'm away from my computer over Thanksgiving) but numpy is required by TF, so that's safe to use.</p>\n\n<p>I set up these requirements in the hope that the winning model can be turned into an official TensorFlow example. Hopefully that makes sense?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "246275": "For the Raspberry Pi track, what does it mean that the only dependency must be Tensorflow? I think that this statement needs some clarification. Does that mean that I can only import python standard library and Tensorflow, but not numpy and other libraries? Or does it mean something else?",
    "246637": "Good point! There are two aspects to this. For the latency evaluation, the trained model must be delivered as a TensorFlow frozen GraphDef file, which will mean it can't rely on any non-TF code, since the freezing process won't work if it does.\n\nFor training, the script shouldn't rely on any Python dependencies that are not already required by TensorFlow's pip package. Unfortunately I don't have a complete list here (and I'm away from my computer over Thanksgiving) but numpy is required by TF, so that's safe to use.\n\nI set up these requirements in the hope that the winning model can be turned into an official TensorFlow example. Hopefully that makes sense?"
  },
  "source": "meta"
}