{
  "id": 271802,
  "title": "Improving inference speed",
  "url": "/competitions/landmark-retrieval-2021/discussion/271802",
  "author_name": "",
  "post_date": "2021-09-12T17:14:36.383567300Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>In the published notebook, the distance is calculated by merging test data with index one by one, but it is faster to do it all at once.</p>\n<p>In the case of the 1129 public test data, we only need to do knn once, which is 1129 times faster.</p>\n<p>Our team is actually doing that.</p>\n<p>Good luck👍</p>",
  "messages": [
    {
      "id": "1510720",
      "postDate": "09/12/2021 17:14:36",
      "content": "<p>In the published notebook, the distance is calculated by merging test data with index one by one, but it is faster to do it all at once.</p>\n<p>In the case of the 1129 public test data, we only need to do knn once, which is 1129 times faster.</p>\n<p>Our team is actually doing that.</p>\n<p>Good luck👍</p>",
      "rawMarkdown": "In the published notebook, the distance is calculated by merging test data with index one by one, but it is faster to do it all at once.\n\nIn the case of the 1129 public test data, we only need to do knn once, which is 1129 times faster.\n\nOur team is actually doing that.\n\nGood luck👍",
      "votes": null
    },
    {
      "id": "1511029",
      "postDate": "09/13/2021 03:34:25",
      "content": "<p>Besides, faiss is also a good option.<br>\nPlease give it a try.</p>",
      "rawMarkdown": "Besides, faiss is also a good option.\nPlease give it a try.",
      "votes": null
    },
    {
      "id": "1512570",
      "postDate": "09/14/2021 11:18:41",
      "content": "<p>Thank you for your sharing~</p>",
      "rawMarkdown": "Thank you for your sharing~",
      "votes": null
    },
    {
      "id": "1528378",
      "postDate": "09/29/2021 15:40:29",
      "content": "<p>you can use cuda calculation by using pytorch<br>\nsend your data to pytorch dataloader<br>\nafter that, send the dataloader to cuda<br>\nafter that, use cuda calculation like (torch.mm(A,B)) to get a similarity between query and index<br>\nalso faiss API can be a good choice<br>\ngood luck beluga!</p>",
      "rawMarkdown": "you can use cuda calculation by using pytorch\nsend your data to pytorch dataloader\nafter that, send the dataloader to cuda\nafter that, use cuda calculation like (torch.mm(A,B)) to get a similarity between query and index\nalso faiss API can be a good choice\ngood luck beluga!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1511029,
      "author_name": "inoueu1",
      "author_url": "",
      "post_date": "09/13/2021 03:34:25",
      "content": "<p>Besides, faiss is also a good option.<br>\nPlease give it a try.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1512570,
      "author_name": "howiechangchn",
      "author_url": "",
      "post_date": "09/14/2021 11:18:41",
      "content": "<p>Thank you for your sharing~</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1528378,
      "author_name": "deepkim",
      "author_url": "",
      "post_date": "09/29/2021 15:40:29",
      "content": "<p>you can use cuda calculation by using pytorch<br>\nsend your data to pytorch dataloader<br>\nafter that, send the dataloader to cuda<br>\nafter that, use cuda calculation like (torch.mm(A,B)) to get a similarity between query and index<br>\nalso faiss API can be a good choice<br>\ngood luck beluga!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1510720": "In the published notebook, the distance is calculated by merging test data with index one by one, but it is faster to do it all at once.\n\nIn the case of the 1129 public test data, we only need to do knn once, which is 1129 times faster.\n\nOur team is actually doing that.\n\nGood luck👍",
    "1511029": "Besides, faiss is also a good option.\nPlease give it a try.",
    "1512570": "Thank you for your sharing~",
    "1528378": "you can use cuda calculation by using pytorch\nsend your data to pytorch dataloader\nafter that, send the dataloader to cuda\nafter that, use cuda calculation like (torch.mm(A,B)) to get a similarity between query and index\nalso faiss API can be a good choice\ngood luck beluga!"
  },
  "source": "meta"
}