{
  "id": 133685,
  "title": "Inference time",
  "url": "/competitions/bengaliai-cv19/discussion/133685",
  "author_name": "",
  "post_date": "2020-03-03T20:56:01.659634300Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I use train data to estimate inference time. Blending 2 se_resnext101_32x4d models on train data with image size 128x128 without tta takes me 970s with Kaggle kernels with gpu.  I wonder if I'm missing something that could speed this up.</p>\n\n<p>OK, speeding up things may not be where I should focus, getting better models is where I should focus.</p>",
  "messages": [
    {
      "id": "762832",
      "postDate": "03/03/2020 20:56:01",
      "content": "<p>I use train data to estimate inference time. Blending 2 se_resnext101_32x4d models on train data with image size 128x128 without tta takes me 970s with Kaggle kernels with gpu.  I wonder if I'm missing something that could speed this up.</p>\n\n<p>OK, speeding up things may not be where I should focus, getting better models is where I should focus.</p>",
      "rawMarkdown": "I use train data to estimate inference time. Blending 2 se_resnext101_32x4d models on train data with image size 128x128 without tta takes me 970s with Kaggle kernels with gpu.  I wonder if I'm missing something that could speed this up.\n\nOK, speeding up things may not be where I should focus, getting better models is where I should focus.",
      "votes": null
    },
    {
      "id": "762839",
      "postDate": "03/03/2020 21:06:58",
      "content": "<p>Inference is quite fast, what is slow is loading the parquet files, but you only do this once. Your 8 minutes per model seem quite fast, you can blend a lot if you want.</p>",
      "rawMarkdown": "Inference is quite fast, what is slow is loading the parquet files, but you only do this once. Your 8 minutes per model seem quite fast, you can blend a lot if you want.",
      "votes": null
    },
    {
      "id": "762859",
      "postDate": "03/03/2020 21:56:06",
      "content": "<p>Indeed, if I have something good to blend...</p>",
      "rawMarkdown": "Indeed, if I have something good to blend...",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 762839,
      "author_name": "philippsinger",
      "author_url": "",
      "post_date": "03/03/2020 21:06:58",
      "content": "<p>Inference is quite fast, what is slow is loading the parquet files, but you only do this once. Your 8 minutes per model seem quite fast, you can blend a lot if you want.</p>",
      "votes": null,
      "replies": [
        {
          "id": 762859,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "03/03/2020 21:56:06",
          "content": "<p>Indeed, if I have something good to blend...</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "762832": "I use train data to estimate inference time. Blending 2 se_resnext101_32x4d models on train data with image size 128x128 without tta takes me 970s with Kaggle kernels with gpu.  I wonder if I'm missing something that could speed this up.\n\nOK, speeding up things may not be where I should focus, getting better models is where I should focus.",
    "762839": "Inference is quite fast, what is slow is loading the parquet files, but you only do this once. Your 8 minutes per model seem quite fast, you can blend a lot if you want.",
    "762859": "Indeed, if I have something good to blend..."
  },
  "source": "meta"
}