{
  "id": 222611,
  "title": "Running out  of GPU memory",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/222611",
  "author_name": "",
  "post_date": "2021-02-28T07:50:29.603661500Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Earlier I was getting a T4 GPU on which I was able to run an EfficientUNet B5 with a batch size of 2 and image size 512x512.<br>\nNow, I am getting a P100 GPU (which has the same 16GB VRAM). But I am running out of memory even with an EfficientUNet B3 and image size 320x320 with the same batch size.</p>\n<p>As much as I know, I haven't changed anything else. I checked that the GPU I am getting has 0mb of memory used in nvidia-smi before running anything.</p>",
  "messages": [
    {
      "id": "1220648",
      "postDate": "02/28/2021 07:50:29",
      "content": "<p>Earlier I was getting a T4 GPU on which I was able to run an EfficientUNet B5 with a batch size of 2 and image size 512x512.<br>\nNow, I am getting a P100 GPU (which has the same 16GB VRAM). But I am running out of memory even with an EfficientUNet B3 and image size 320x320 with the same batch size.</p>\n<p>As much as I know, I haven't changed anything else. I checked that the GPU I am getting has 0mb of memory used in nvidia-smi before running anything.</p>",
      "rawMarkdown": "Earlier I was getting a T4 GPU on which I was able to run an EfficientUNet B5 with a batch size of 2 and image size 512x512.\nNow, I am getting a P100 GPU (which has the same 16GB VRAM). But I am running out of memory even with an EfficientUNet B3 and image size 320x320 with the same batch size.\n\nAs much as I know, I haven't changed anything else. I checked that the GPU I am getting has 0mb of memory used in nvidia-smi before running anything.",
      "votes": null
    },
    {
      "id": "1221469",
      "postDate": "03/01/2021 02:44:14",
      "content": "<p>you may want to check dtype. np.float32 will take more bits than np.uint8.</p>",
      "rawMarkdown": "you may want to check dtype. np.float32 will take more bits than np.uint8.",
      "votes": null
    },
    {
      "id": "1221760",
      "postDate": "03/01/2021 08:47:39",
      "content": "<p>Just realized my random crop transform was commented, I was using full-sized images. -_-</p>",
      "rawMarkdown": "Just realized my random crop transform was commented, I was using full-sized images. -_-",
      "votes": null
    },
    {
      "id": "1241187",
      "postDate": "03/17/2021 00:01:56",
      "content": "<p>Yes that could kill it. Did you adjust down? Also, maybe don't adjust at the crop, adjust well ahead of that with the images sized down.</p>",
      "rawMarkdown": "Yes that could kill it. Did you adjust down? Also, maybe don't adjust at the crop, adjust well ahead of that with the images sized down.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1221469,
      "author_name": "igor14497",
      "author_url": "",
      "post_date": "03/01/2021 02:44:14",
      "content": "<p>you may want to check dtype. np.float32 will take more bits than np.uint8.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1221760,
          "author_name": "pranshu15",
          "author_url": "",
          "post_date": "03/01/2021 08:47:39",
          "content": "<p>Just realized my random crop transform was commented, I was using full-sized images. -_-</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1241187,
          "author_name": "crained",
          "author_url": "",
          "post_date": "03/17/2021 00:01:56",
          "content": "<p>Yes that could kill it. Did you adjust down? Also, maybe don't adjust at the crop, adjust well ahead of that with the images sized down.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1220648": "Earlier I was getting a T4 GPU on which I was able to run an EfficientUNet B5 with a batch size of 2 and image size 512x512.\nNow, I am getting a P100 GPU (which has the same 16GB VRAM). But I am running out of memory even with an EfficientUNet B3 and image size 320x320 with the same batch size.\n\nAs much as I know, I haven't changed anything else. I checked that the GPU I am getting has 0mb of memory used in nvidia-smi before running anything.",
    "1221469": "you may want to check dtype. np.float32 will take more bits than np.uint8.",
    "1221760": "Just realized my random crop transform was commented, I was using full-sized images. -_-",
    "1241187": "Yes that could kill it. Did you adjust down? Also, maybe don't adjust at the crop, adjust well ahead of that with the images sized down."
  },
  "source": "meta"
}