{
  "id": 350293,
  "title": "Help with starting out",
  "url": "/competitions/hubmap-organ-segmentation/discussion/350293",
  "author_name": "",
  "post_date": "2022-09-05T04:01:47.445295100Z",
  "votes": 11,
  "comment_count": 6,
  "views": 0,
  "content": "<p>It's been 2-3 weeks since we started working on this competition and so far we haven't had any success.<br>\nThis is not a post to lament but ask for suggestions and get some highly needed help.<br>\nWe've tried patches vs resizing approach. Trained UNet from scratch, pre-trained encoders… but we aren't getting anything at all. Best CV we've had is ~0.15, which isn't even worth submitting. The images are either blank or just learn to do edge-detection on the input.<br>\nAny guide to what we might be doing wrong?</p>",
  "messages": [
    {
      "id": "1926712",
      "postDate": "09/05/2022 04:01:47",
      "content": "<p>It's been 2-3 weeks since we started working on this competition and so far we haven't had any success.<br>\nThis is not a post to lament but ask for suggestions and get some highly needed help.<br>\nWe've tried patches vs resizing approach. Trained UNet from scratch, pre-trained encoders… but we aren't getting anything at all. Best CV we've had is ~0.15, which isn't even worth submitting. The images are either blank or just learn to do edge-detection on the input.<br>\nAny guide to what we might be doing wrong?</p>",
      "rawMarkdown": "It's been 2-3 weeks since we started working on this competition and so far we haven't had any success.\nThis is not a post to lament but ask for suggestions and get some highly needed help.\nWe've tried patches vs resizing approach. Trained UNet from scratch, pre-trained encoders... but we aren't getting anything at all. Best CV we've had is ~0.15, which isn't even worth submitting. The images are either blank or just learn to do edge-detection on the input.\nAny guide to what we might be doing wrong?",
      "votes": null
    },
    {
      "id": "1926931",
      "postDate": "09/05/2022 08:54:19",
      "content": "<p>Are you doing augmentations on the dataset?</p>",
      "rawMarkdown": "Are you doing augmentations on the dataset?",
      "votes": null
    },
    {
      "id": "1927216",
      "postDate": "09/05/2022 13:07:25",
      "content": "<p>No<br>\nI thought that first i should try to overfit model on the existing data then i will move on to augmentations.<br>\nBecause i suspect that the model isn't learning anything </p>",
      "rawMarkdown": "No\nI thought that first i should try to overfit model on the existing data then i will move on to augmentations.\nBecause i suspect that the model isn't learning anything",
      "votes": null
    },
    {
      "id": "1928921",
      "postDate": "09/06/2022 17:38:58",
      "content": "<p>Loss function bce or iou ?  Normalisation?</p>",
      "rawMarkdown": "Loss function bce or iou ?  Normalisation?",
      "votes": null
    },
    {
      "id": "1929247",
      "postDate": "09/06/2022 23:28:08",
      "content": "<p>I used bce<br>\nAnd as for normalisation i just divided the image by 255, to bring the pixel values between 0 and 1</p>",
      "rawMarkdown": "I used bce\nAnd as for normalisation i just divided the image by 255, to bring the pixel values between 0 and 1",
      "votes": null
    },
    {
      "id": "1929662",
      "postDate": "09/07/2022 09:02:36",
      "content": "<p>Check if the model  param.requires_grad is True in your models (at least in your decoder) Another possibility is with threshold, but i guess you might have already tried to plot the probabilities from your model output. A simple unet (resnet head) with  12-16 epochs and a batch size of 16 should give you over 0.5 local and around 0.4-0.5 in public LB </p>",
      "rawMarkdown": "Check if the model  param.requires_grad is True in your models (at least in your decoder) Another possibility is with threshold, but i guess you might have already tried to plot the probabilities from your model output. A simple unet (resnet head) with  12-16 epochs and a batch size of 16 should give you over 0.5 local and around 0.4-0.5 in public LB",
      "votes": null
    },
    {
      "id": "1929879",
      "postDate": "09/07/2022 12:12:41",
      "content": "<p>Is it possible that everyone is using something other than Kaggle notebooks because i don't think i am.able to go past 4 batch size.<br>\nAnd with threshold you mean the value above and below which i prepare the mask for?</p>",
      "rawMarkdown": "Is it possible that everyone is using something other than Kaggle notebooks because i don't think i am.able to go past 4 batch size.\nAnd with threshold you mean the value above and below which i prepare the mask for?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1926931,
      "author_name": "julianmukaj",
      "author_url": "",
      "post_date": "09/05/2022 08:54:19",
      "content": "<p>Are you doing augmentations on the dataset?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1927216,
          "author_name": "bhavesjain",
          "author_url": "",
          "post_date": "09/05/2022 13:07:25",
          "content": "<p>No<br>\nI thought that first i should try to overfit model on the existing data then i will move on to augmentations.<br>\nBecause i suspect that the model isn't learning anything </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1928921,
      "author_name": "riyasvk",
      "author_url": "",
      "post_date": "09/06/2022 17:38:58",
      "content": "<p>Loss function bce or iou ?  Normalisation?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1929247,
          "author_name": "bhavesjain",
          "author_url": "",
          "post_date": "09/06/2022 23:28:08",
          "content": "<p>I used bce<br>\nAnd as for normalisation i just divided the image by 255, to bring the pixel values between 0 and 1</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1929662,
          "author_name": "riyasvk",
          "author_url": "",
          "post_date": "09/07/2022 09:02:36",
          "content": "<p>Check if the model  param.requires_grad is True in your models (at least in your decoder) Another possibility is with threshold, but i guess you might have already tried to plot the probabilities from your model output. A simple unet (resnet head) with  12-16 epochs and a batch size of 16 should give you over 0.5 local and around 0.4-0.5 in public LB </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1929879,
          "author_name": "bhavesjain",
          "author_url": "",
          "post_date": "09/07/2022 12:12:41",
          "content": "<p>Is it possible that everyone is using something other than Kaggle notebooks because i don't think i am.able to go past 4 batch size.<br>\nAnd with threshold you mean the value above and below which i prepare the mask for?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1926712": "It's been 2-3 weeks since we started working on this competition and so far we haven't had any success.\nThis is not a post to lament but ask for suggestions and get some highly needed help.\nWe've tried patches vs resizing approach. Trained UNet from scratch, pre-trained encoders... but we aren't getting anything at all. Best CV we've had is ~0.15, which isn't even worth submitting. The images are either blank or just learn to do edge-detection on the input.\nAny guide to what we might be doing wrong?",
    "1926931": "Are you doing augmentations on the dataset?",
    "1927216": "No\nI thought that first i should try to overfit model on the existing data then i will move on to augmentations.\nBecause i suspect that the model isn't learning anything",
    "1928921": "Loss function bce or iou ?  Normalisation?",
    "1929247": "I used bce\nAnd as for normalisation i just divided the image by 255, to bring the pixel values between 0 and 1",
    "1929662": "Check if the model  param.requires_grad is True in your models (at least in your decoder) Another possibility is with threshold, but i guess you might have already tried to plot the probabilities from your model output. A simple unet (resnet head) with  12-16 epochs and a batch size of 16 should give you over 0.5 local and around 0.4-0.5 in public LB",
    "1929879": "Is it possible that everyone is using something other than Kaggle notebooks because i don't think i am.able to go past 4 batch size.\nAnd with threshold you mean the value above and below which i prepare the mask for?"
  },
  "source": "meta"
}