{
  "id": 170943,
  "title": "Did Anyone Have Success Using Accumulated Gradients?",
  "url": "/competitions/alaska2-image-steganalysis/discussion/170943",
  "author_name": "",
  "post_date": "2020-07-29T15:52:43.139553200Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hey all,</p>\n\n<p>I am fairly new when it comes to computer vision competitions. One of the things I tried to use was accumulated gradients because I only have a 2080ti with 12GB of memory. </p>\n\n<p>I tried to implement some of the winning solutions - mainly experimenting with different efficient net architectures B0 - B6 and changing the first stride but my results are still only in the 0.91X range after 40 epochs (maybe I still need to train longer but this took 7 days!)? </p>\n\n<p>I am wondering if the community has any feedback or general knowledge to share.</p>",
  "messages": [
    {
      "id": "950778",
      "postDate": "07/29/2020 15:52:43",
      "content": "<p>Hey all,</p>\n\n<p>I am fairly new when it comes to computer vision competitions. One of the things I tried to use was accumulated gradients because I only have a 2080ti with 12GB of memory. </p>\n\n<p>I tried to implement some of the winning solutions - mainly experimenting with different efficient net architectures B0 - B6 and changing the first stride but my results are still only in the 0.91X range after 40 epochs (maybe I still need to train longer but this took 7 days!)? </p>\n\n<p>I am wondering if the community has any feedback or general knowledge to share.</p>",
      "rawMarkdown": "Hey all,\n\nI am fairly new when it comes to computer vision competitions. One of the things I tried to use was accumulated gradients because I only have a 2080ti with 12GB of memory. \n\nI tried to implement some of the winning solutions - mainly experimenting with different efficient net architectures B0 - B6 and changing the first stride but my results are still only in the 0.91X range after 40 epochs (maybe I still need to train longer but this took 7 days!)? \n\nI am wondering if the community has any feedback or general knowledge to share.",
      "votes": null
    },
    {
      "id": "951221",
      "postDate": "07/30/2020 02:06:55",
      "content": "<p>You may try to use TPU. It can train one epoch in 15~60 mins. For me each model take over 100 epochs. </p>",
      "rawMarkdown": "You may try to use TPU. It can train one epoch in 15~60 mins. For me each model take over 100 epochs.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 951221,
      "author_name": "wuliaokaola",
      "author_url": "",
      "post_date": "07/30/2020 02:06:55",
      "content": "<p>You may try to use TPU. It can train one epoch in 15~60 mins. For me each model take over 100 epochs. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "950778": "Hey all,\n\nI am fairly new when it comes to computer vision competitions. One of the things I tried to use was accumulated gradients because I only have a 2080ti with 12GB of memory. \n\nI tried to implement some of the winning solutions - mainly experimenting with different efficient net architectures B0 - B6 and changing the first stride but my results are still only in the 0.91X range after 40 epochs (maybe I still need to train longer but this took 7 days!)? \n\nI am wondering if the community has any feedback or general knowledge to share.",
    "951221": "You may try to use TPU. It can train one epoch in 15~60 mins. For me each model take over 100 epochs."
  },
  "source": "meta"
}