{
  "id": 46770,
  "title": "Model evaluation on Raspberry Pi3",
  "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/46770",
  "author_name": "",
  "post_date": "2018-01-03T02:23:34.594316600Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<ul>\n<li>The target hardware is Raspberry Pi3 model B with Raspbian</li>\n<li>The evaluation is performed with the Google benchmark script for RPi3 on the pre-trained model in the tutorial. 100 runs is performed for each graph, with 100 warmup runs and 0 delays between each run and benchmarks</li>\n<li><p>The following data gives the final timing results and memory consumption of the three model in tutorial</p>\n\n<ul><li>Model &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Avg Timing(ms)&nbsp;&nbsp;&nbsp;&nbsp;Memory(Bytes)</li>\n<li>conv&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;87.966&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;1621752</li>\n<li>low-latency conv&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;17.169 &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;142048</li>\n<li>low-latency svdf&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;27.361&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;3144480  </li></ul></li>\n</ul>",
  "messages": [
    {
      "id": "264421",
      "postDate": "01/03/2018 02:23:34",
      "content": "<ul>\n<li>The target hardware is Raspberry Pi3 model B with Raspbian</li>\n<li>The evaluation is performed with the Google benchmark script for RPi3 on the pre-trained model in the tutorial. 100 runs is performed for each graph, with 100 warmup runs and 0 delays between each run and benchmarks</li>\n<li><p>The following data gives the final timing results and memory consumption of the three model in tutorial</p>\n\n<ul><li>Model &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Avg Timing(ms)&nbsp;&nbsp;&nbsp;&nbsp;Memory(Bytes)</li>\n<li>conv&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;87.966&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;1621752</li>\n<li>low-latency conv&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;17.169 &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;142048</li>\n<li>low-latency svdf&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;27.361&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;3144480  </li></ul></li>\n</ul>",
      "rawMarkdown": "The target hardware is Raspberry Pi3 model B with Raspbian\n  - The evaluation is performed with the Google benchmark script for RPi3 on the pre-trained model in the tutorial. 100 runs is performed for each graph, with 100 warmup runs and 0 delays between each run and benchmarks\n  - The following data gives the final timing results and memory consumption of the three model in tutorial\n\n- Model &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Avg Timing(ms)&nbsp;&nbsp;&nbsp;&nbsp;Memory(Bytes)\n- conv&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;87.966&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;1621752\n- low-latency conv&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;17.169 &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;142048\n- low-latency svdf&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;27.361&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;3144480",
      "votes": null
    },
    {
      "id": "265363",
      "postDate": "01/05/2018 10:37:43",
      "content": "<p>Thanks for publishing. I got the following summaries (Python 3.4.2, Raspbian GNU/Linux 8.0 (jessie)) :</p>\n\n<ul>\n<li>Model  | Avg Timing(ms) | Memory(Bytes)  | FLOPs | size(Bytes) |</li>\n<li>conv    |                  74.132 | 1.621.752 | 402.91M | 3.710.976 |</li>\n<li>low_latency_conv   |   17.656      |   142.048 | 11.23M | 3.805.184 |</li>\n<li>single_fc |  7.976     | 116.472 | 94.08K |  192.512 |</li>\n</ul>\n\n<p>I didn't expect that the conv model would be suitable for the special price but the Pi3 is quite fast. In fact you could run the model twice per sample and still be within the 175ms limit.</p>",
      "rawMarkdown": "Thanks for publishing. I got the following summaries (Python 3.4.2, Raspbian GNU/Linux 8.0 (jessie)) :\n\n  - Model  | Avg Timing(ms) | Memory(Bytes)  | FLOPs | size(Bytes) |\n  - conv    |                  74.132 | 1.621.752 | 402.91M | 3.710.976 |\n  - low_latency_conv   |   17.656      |   142.048 | 11.23M | 3.805.184 |\n  - single_fc |  7.976     | 116.472 | 94.08K |  192.512 |\n\nI didn't expect that the conv model would be suitable for the special price but the Pi3 is quite fast. In fact you could run the model twice per sample and still be within the 175ms limit.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 265363,
      "author_name": "seesee",
      "author_url": "",
      "post_date": "01/05/2018 10:37:43",
      "content": "<p>Thanks for publishing. I got the following summaries (Python 3.4.2, Raspbian GNU/Linux 8.0 (jessie)) :</p>\n\n<ul>\n<li>Model  | Avg Timing(ms) | Memory(Bytes)  | FLOPs | size(Bytes) |</li>\n<li>conv    |                  74.132 | 1.621.752 | 402.91M | 3.710.976 |</li>\n<li>low_latency_conv   |   17.656      |   142.048 | 11.23M | 3.805.184 |</li>\n<li>single_fc |  7.976     | 116.472 | 94.08K |  192.512 |</li>\n</ul>\n\n<p>I didn't expect that the conv model would be suitable for the special price but the Pi3 is quite fast. In fact you could run the model twice per sample and still be within the 175ms limit.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "264421": "The target hardware is Raspberry Pi3 model B with Raspbian\n  - The evaluation is performed with the Google benchmark script for RPi3 on the pre-trained model in the tutorial. 100 runs is performed for each graph, with 100 warmup runs and 0 delays between each run and benchmarks\n  - The following data gives the final timing results and memory consumption of the three model in tutorial\n\n- Model &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Avg Timing(ms)&nbsp;&nbsp;&nbsp;&nbsp;Memory(Bytes)\n- conv&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;87.966&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;1621752\n- low-latency conv&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;17.169 &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;142048\n- low-latency svdf&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;27.361&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;3144480",
    "265363": "Thanks for publishing. I got the following summaries (Python 3.4.2, Raspbian GNU/Linux 8.0 (jessie)) :\n\n  - Model  | Avg Timing(ms) | Memory(Bytes)  | FLOPs | size(Bytes) |\n  - conv    |                  74.132 | 1.621.752 | 402.91M | 3.710.976 |\n  - low_latency_conv   |   17.656      |   142.048 | 11.23M | 3.805.184 |\n  - single_fc |  7.976     | 116.472 | 94.08K |  192.512 |\n\nI didn't expect that the conv model would be suitable for the special price but the Pi3 is quite fast. In fact you could run the model twice per sample and still be within the 175ms limit."
  },
  "source": "meta"
}