{
  "id": 163697,
  "title": "Accuracy changes on moving from cpu to gpu",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/163697",
  "author_name": "",
  "post_date": "2020-07-03T05:06:10.106737200Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>So I came across a very weird error. I am training the last block Resnext50 for this competition. I trained it on a CPU for 2 epochs, ended up with a training accuracy of 0.81 and a validation accuracy of 0.75 after which the notebook crashed. When I did the same thing on a GPU, the accuracy kept oscillating around 0.5. Does anyone have an idea about how to solve this?</p>",
  "messages": [
    {
      "id": "913268",
      "postDate": "07/03/2020 05:06:10",
      "content": "<p>So I came across a very weird error. I am training the last block Resnext50 for this competition. I trained it on a CPU for 2 epochs, ended up with a training accuracy of 0.81 and a validation accuracy of 0.75 after which the notebook crashed. When I did the same thing on a GPU, the accuracy kept oscillating around 0.5. Does anyone have an idea about how to solve this?</p>",
      "rawMarkdown": "So I came across a very weird error. I am training the last block Resnext50 for this competition. I trained it on a CPU for 2 epochs, ended up with a training accuracy of 0.81 and a validation accuracy of 0.75 after which the notebook crashed. When I did the same thing on a GPU, the accuracy kept oscillating around 0.5. Does anyone have an idea about how to solve this?",
      "votes": null
    },
    {
      "id": "913360",
      "postDate": "07/03/2020 07:09:03",
      "content": "<p>The performance of number of nodes on each layer is different on CPU then GPU. Its basically due to there architectures. So , while training it on GPU , you need to keep this in your mind. Along with , you also need to work upon adding some other layers to for proper preprocessing of inforamtion you gain from ResNext50. I hope this will help you in improving performance.</p>",
      "rawMarkdown": "The performance of number of nodes on each layer is different on CPU then GPU. Its basically due to there architectures. So , while training it on GPU , you need to keep this in your mind. Along with , you also need to work upon adding some other layers to for proper preprocessing of inforamtion you gain from ResNext50. I hope this will help you in improving performance.",
      "votes": null
    },
    {
      "id": "913368",
      "postDate": "07/03/2020 07:17:52",
      "content": "<p><a href=\"/prashantarorat\">@prashantarorat</a>  Thanks for the reply. Could you elaborate as to how this error can be solved in terms of pytorch code?</p>",
      "rawMarkdown": "prashantarorat  Thanks for the reply. Could you elaborate as to how this error can be solved in terms of pytorch code?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 913360,
      "author_name": "prashantarorat",
      "author_url": "",
      "post_date": "07/03/2020 07:09:03",
      "content": "<p>The performance of number of nodes on each layer is different on CPU then GPU. Its basically due to there architectures. So , while training it on GPU , you need to keep this in your mind. Along with , you also need to work upon adding some other layers to for proper preprocessing of inforamtion you gain from ResNext50. I hope this will help you in improving performance.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 913368,
      "author_name": "aryaman1999",
      "author_url": "",
      "post_date": "07/03/2020 07:17:52",
      "content": "<p><a href=\"/prashantarorat\">@prashantarorat</a>  Thanks for the reply. Could you elaborate as to how this error can be solved in terms of pytorch code?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "913268": "So I came across a very weird error. I am training the last block Resnext50 for this competition. I trained it on a CPU for 2 epochs, ended up with a training accuracy of 0.81 and a validation accuracy of 0.75 after which the notebook crashed. When I did the same thing on a GPU, the accuracy kept oscillating around 0.5. Does anyone have an idea about how to solve this?",
    "913360": "The performance of number of nodes on each layer is different on CPU then GPU. Its basically due to there architectures. So , while training it on GPU , you need to keep this in your mind. Along with , you also need to work upon adding some other layers to for proper preprocessing of inforamtion you gain from ResNext50. I hope this will help you in improving performance.",
    "913368": "prashantarorat  Thanks for the reply. Could you elaborate as to how this error can be solved in terms of pytorch code?"
  },
  "source": "meta"
}