{
  "id": 227701,
  "title": "Number of Train images",
  "url": "/competitions/hpa-single-cell-image-classification/discussion/227701",
  "author_name": "GitMach",
  "post_date": "2021-03-21T20:36:26.473000",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi, </p>\n<p>Simple question, simple answer PLS.<br>\nHow many images have you trained ?<br>\nThank you for your feedback</p>",
  "messages": [
    {
      "id": 1247633,
      "postDate": "2021-03-21T22:30:17.360Z",
      "content": "<p>All training has been on local machines.  I have 64GB in CPU RAM and additional 200GB in swap file so I can load up memory with no worries - batch sizes on GPU is memory constraint for me - I have dual GPU and used mixed precision and most models are trained with batch size of 64.</p>\n<p>I created single cell RGB images  - the number of these is the only clean reference that I have.  My single cell images are 512x512.</p>\n<p>When training I started with only those  images that contained one label.  I drop all single cells that are on the border of the original large image.  This leaves me with 160, 580 single cell images to train.  I get 205,858 single cell images if I selected the images that have one or two labels (still dropping those on the original image edge).    I get 308,746 single cell images if I select all the images regardless of number of labels (yep - still dropping the original edge cells).</p>\n<p>If I include all the images and the single cells that are on the border this gives me 491,092 single cell images to train.  I have not yet trained any models with this total.</p>",
      "rawMarkdown": "All training has been on local machines.  I have 64GB in CPU RAM and additional 200GB in swap file so I can load up memory with no worries - batch sizes on GPU is memory constraint for me - I have dual GPU and used mixed precision and most models are trained with batch size of 64.\n\nI created single cell RGB images  - the number of these is the only clean reference that I have.  My single cell images are 512x512.\n\nWhen training I started with only those  images that contained one label.  I drop all single cells that are on the border of the original large image.  This leaves me with 160, 580 single cell images to train.  I get 205,858 single cell images if I selected the images that have one or two labels (still dropping those on the original image edge).    I get 308,746 single cell images if I select all the images regardless of number of labels (yep - still dropping the original edge cells).\n\nIf I include all the images and the single cells that are on the border this gives me 491,092 single cell images to train.  I have not yet trained any models with this total.",
      "votes": 1,
      "replies": [
        {
          "id": 1247999,
          "postDate": "2021-03-22T08:53:02.300Z",
          "content": "<p>Good approach, starting by selecting only images with one Cell. <br>\nThank you for sharing.</p>",
          "rawMarkdown": "Good approach, starting by selecting only images with one Cell. \nThank you for sharing."
        }
      ]
    },
    {
      "id": 1247567,
      "postDate": "2021-03-21T20:36:26.473Z",
      "content": "<p>Hi, </p>\n<p>Simple question, simple answer PLS.<br>\nHow many images have you trained ?<br>\nThank you for your feedback</p>",
      "rawMarkdown": "Hi, \n\nSimple question, simple answer PLS.\nHow many images have you trained ?\nThank you for your feedback",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 1247633,
      "author_name": "PC Jimmmy",
      "author_url": "",
      "post_date": "2021-03-21T22:30:17.360000",
      "content": "<p>All training has been on local machines.  I have 64GB in CPU RAM and additional 200GB in swap file so I can load up memory with no worries - batch sizes on GPU is memory constraint for me - I have dual GPU and used mixed precision and most models are trained with batch size of 64.</p>\n<p>I created single cell RGB images  - the number of these is the only clean reference that I have.  My single cell images are 512x512.</p>\n<p>When training I started with only those  images that contained one label.  I drop all single cells that are on the border of the original large image.  This leaves me with 160, 580 single cell images to train.  I get 205,858 single cell images if I selected the images that have one or two labels (still dropping those on the original image edge).    I get 308,746 single cell images if I select all the images regardless of number of labels (yep - still dropping the original edge cells).</p>\n<p>If I include all the images and the single cells that are on the border this gives me 491,092 single cell images to train.  I have not yet trained any models with this total.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1247999,
          "author_name": "GitMach",
          "author_url": "",
          "post_date": "2021-03-22T08:53:02.300000",
          "content": "<p>Good approach, starting by selecting only images with one Cell. <br>\nThank you for sharing.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1247633": "All training has been on local machines.  I have 64GB in CPU RAM and additional 200GB in swap file so I can load up memory with no worries - batch sizes on GPU is memory constraint for me - I have dual GPU and used mixed precision and most models are trained with batch size of 64.\n\nI created single cell RGB images  - the number of these is the only clean reference that I have.  My single cell images are 512x512.\n\nWhen training I started with only those  images that contained one label.  I drop all single cells that are on the border of the original large image.  This leaves me with 160, 580 single cell images to train.  I get 205,858 single cell images if I selected the images that have one or two labels (still dropping those on the original image edge).    I get 308,746 single cell images if I select all the images regardless of number of labels (yep - still dropping the original edge cells).\n\nIf I include all the images and the single cells that are on the border this gives me 491,092 single cell images to train.  I have not yet trained any models with this total.",
    "1247567": "Hi, \n\nSimple question, simple answer PLS.\nHow many images have you trained ?\nThank you for your feedback"
  }
}