{
  "id": 166975,
  "title": "Inference is too slow",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/166975",
  "author_name": "",
  "post_date": "2020-07-14T18:02:26.662297600Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I have 4 models with different input size and TTA is implemented in my inference kernel, but the speed of  images preprocessing is too slow, don't know how to speed up the processing. I'm going to try preprocessing all images in advance and store these images in working space, then start predicting. Any good ideas?</p>",
  "messages": [
    {
      "id": "929510",
      "postDate": "07/14/2020 18:02:26",
      "content": "<p>I have 4 models with different input size and TTA is implemented in my inference kernel, but the speed of  images preprocessing is too slow, don't know how to speed up the processing. I'm going to try preprocessing all images in advance and store these images in working space, then start predicting. Any good ideas?</p>",
      "rawMarkdown": "I have 4 models with different input size and TTA is implemented in my inference kernel, but the speed of  images preprocessing is too slow, don't know how to speed up the processing. I'm going to try preprocessing all images in advance and store these images in working space, then start predicting. Any good ideas?",
      "votes": null
    },
    {
      "id": "931019",
      "postDate": "07/15/2020 22:42:31",
      "content": "<p>Are you resizing the images each time during image preprocessing?</p>",
      "rawMarkdown": "Are you resizing the images each time during image preprocessing?",
      "votes": null
    },
    {
      "id": "932187",
      "postDate": "07/16/2020 19:46:24",
      "content": "<p>You can use larger batch sizes during inference compared to training since the GPU doesn't need to store gradients. Keep an eye on how much GPU memory you have free and increase the batch size accordingly</p>",
      "rawMarkdown": "You can use larger batch sizes during inference compared to training since the GPU doesn't need to store gradients. Keep an eye on how much GPU memory you have free and increase the batch size accordingly",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 931019,
      "author_name": "sayakdasgupta",
      "author_url": "",
      "post_date": "07/15/2020 22:42:31",
      "content": "<p>Are you resizing the images each time during image preprocessing?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 932187,
      "author_name": "anjum48",
      "author_url": "",
      "post_date": "07/16/2020 19:46:24",
      "content": "<p>You can use larger batch sizes during inference compared to training since the GPU doesn't need to store gradients. Keep an eye on how much GPU memory you have free and increase the batch size accordingly</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "929510": "I have 4 models with different input size and TTA is implemented in my inference kernel, but the speed of  images preprocessing is too slow, don't know how to speed up the processing. I'm going to try preprocessing all images in advance and store these images in working space, then start predicting. Any good ideas?",
    "931019": "Are you resizing the images each time during image preprocessing?",
    "932187": "You can use larger batch sizes during inference compared to training since the GPU doesn't need to store gradients. Keep an eye on how much GPU memory you have free and increase the batch size accordingly"
  },
  "source": "meta"
}