{
  "id": 189043,
  "title": "Disappearing Inference Time Limitations",
  "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/189043",
  "author_name": "",
  "post_date": "2020-10-06T12:58:54.824588800Z",
  "votes": 5,
  "comment_count": 2,
  "views": 0,
  "content": "<p>At first it looks like as a code competition, there is even a code requirements page with run time limits (e.g. GPU Notebook &lt;= 9 hours run-time). Even the Lyft webinar mentions limited inference time.<br>\nOn the other hand since the test set is public, we are allowed to just submit the csv. It was also confirmed by <a href=\"https://www.kaggle.com/iglovikov\" target=\"_blank\">@iglovikov</a> <a href=\"https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/177414#1034502\" target=\"_blank\">https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/177414#1034502</a><br>\nWhat is the reason of the discrepancy?<br>\np.s. I tried to catch up with the forum yesterday sorry if it got already answered.</p>",
  "messages": [
    {
      "id": "1039244",
      "postDate": "10/06/2020 12:58:54",
      "content": "<p>At first it looks like as a code competition, there is even a code requirements page with run time limits (e.g. GPU Notebook &lt;= 9 hours run-time). Even the Lyft webinar mentions limited inference time.<br>\nOn the other hand since the test set is public, we are allowed to just submit the csv. It was also confirmed by <a href=\"https://www.kaggle.com/iglovikov\" target=\"_blank\">@iglovikov</a> <a href=\"https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/177414#1034502\" target=\"_blank\">https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/177414#1034502</a><br>\nWhat is the reason of the discrepancy?<br>\np.s. I tried to catch up with the forum yesterday sorry if it got already answered.</p>",
      "rawMarkdown": "At first it looks like as a code competition, there is even a code requirements page with run time limits (e.g. GPU Notebook <= 9 hours run-time). Even the Lyft webinar mentions limited inference time.\n\nOn the other hand since the test set is public, we are allowed to just submit the csv. It was also confirmed by @iglovikov https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/177414#1034502\n\n\nWhat is the reason of the discrepancy?\n\n\n\np.s. I tried to catch up with the forum yesterday sorry if it got already answered.",
      "votes": null
    },
    {
      "id": "1039850",
      "postDate": "10/06/2020 20:38:22",
      "content": "<p>The plan was to have limited inference time. I liked this experience at Deepfake.</p>\n<p>But we messed up with some technical details and it was a risk not to launch the completion in time =&gt; we got back to the standard format: train offline, predict offline.</p>\n<p>I hope, next time, everything will work as planned.</p>",
      "rawMarkdown": "The plan was to have limited inference time. I liked this experience at Deepfake.\n\nBut we messed up with some technical details and it was a risk not to launch the completion in time => we got back to the standard format: train offline, predict offline.\n\nI hope, next time, everything will work as planned.",
      "votes": null
    },
    {
      "id": "1039874",
      "postDate": "10/06/2020 21:07:42",
      "content": "<p>Thanks, that was my guess. I understand that limited inference would be more practical. <br>\nEven the test set could be hidden as in real word you would not know in advance all the objects you would like to predict.</p>\n<p>To be honest from a competitor point of view it is way more convenient to see the test set and be able to upload csv files. This competition already has some extra <br>\ncomplexity I would rather not struggle with the kernel environment too :)</p>",
      "rawMarkdown": "Thanks, that was my guess. I understand that limited inference would be more practical. \nEven the test set could be hidden as in real word you would not know in advance all the objects you would like to predict.\n\nTo be honest from a competitor point of view it is way more convenient to see the test set and be able to upload csv files. This competition already has some extra \ncomplexity I would rather not struggle with the kernel environment too :)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1039850,
      "author_name": "iglovikov",
      "author_url": "",
      "post_date": "10/06/2020 20:38:22",
      "content": "<p>The plan was to have limited inference time. I liked this experience at Deepfake.</p>\n<p>But we messed up with some technical details and it was a risk not to launch the completion in time =&gt; we got back to the standard format: train offline, predict offline.</p>\n<p>I hope, next time, everything will work as planned.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1039874,
          "author_name": "gaborfodor",
          "author_url": "",
          "post_date": "10/06/2020 21:07:42",
          "content": "<p>Thanks, that was my guess. I understand that limited inference would be more practical. <br>\nEven the test set could be hidden as in real word you would not know in advance all the objects you would like to predict.</p>\n<p>To be honest from a competitor point of view it is way more convenient to see the test set and be able to upload csv files. This competition already has some extra <br>\ncomplexity I would rather not struggle with the kernel environment too :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1039244": "At first it looks like as a code competition, there is even a code requirements page with run time limits (e.g. GPU Notebook <= 9 hours run-time). Even the Lyft webinar mentions limited inference time.\n\nOn the other hand since the test set is public, we are allowed to just submit the csv. It was also confirmed by @iglovikov https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/177414#1034502\n\n\nWhat is the reason of the discrepancy?\n\n\n\np.s. I tried to catch up with the forum yesterday sorry if it got already answered.",
    "1039850": "The plan was to have limited inference time. I liked this experience at Deepfake.\n\nBut we messed up with some technical details and it was a risk not to launch the completion in time => we got back to the standard format: train offline, predict offline.\n\nI hope, next time, everything will work as planned.",
    "1039874": "Thanks, that was my guess. I understand that limited inference would be more practical. \nEven the test set could be hidden as in real word you would not know in advance all the objects you would like to predict.\n\nTo be honest from a competitor point of view it is way more convenient to see the test set and be able to upload csv files. This competition already has some extra \ncomplexity I would rather not struggle with the kernel environment too :)"
  },
  "source": "meta"
}