{
  "id": 192409,
  "title": "My model is making the same prediction for all inputs.",
  "url": "/competitions/rsna-str-pulmonary-embolism-detection/discussion/192409",
  "author_name": "DarkCube",
  "post_date": "2020-10-21T11:38:23.530000",
  "votes": 0,
  "comment_count": 5,
  "views": 0,
  "content": "<p>can someone please help me on this?<br>\nMy model is giving the same output no matter what the input is. I think its some sort of vanishing gradient problem.</p>\n<p>I'm using EfficientNetB0 followed by a single sigmoid neuron to predict image-level labels. I'm using Ian's jpeg dataset. The data is normalized to the [0, 1] range. I tried a 2 softmax neuron and the problem persisted. I tried imagenet's weights and I tried no weights. I tried implementing binary cross-entropy from scratch. I tried decreasing the learning rate. Nothing worked. I'm about to try a cats-and-dogs dataset and train the model on it and see if the problem is in the data.</p>\n<h3>Additional info:</h3>\n<ol>\n<li>I'm splitting data by patient</li>\n<li>I'm training my model only on patients that have a positive exam for pe</li>\n<li>I'm training by patients; splitting the CT scans into batches of 20 and then doing a training step.</li>\n</ol>\n<p>what might be the problem? is there anything I haven't tried yet?</p>\n<p>Edit: I tried a cat and dog dataset and the problem persists. So it can't be that.<br>\nEdit: it's okay. I solved it. Thank you for trying to help me.</p>",
  "messages": [
    {
      "id": 1056042,
      "postDate": "2020-10-21T11:39:42.777Z",
      "content": "<p><a href=\"https://www.kaggle.com/darkcube/fork-of-rsna-pe-competition-stage-1-training\" target=\"_blank\">here is my notebook</a></p>",
      "rawMarkdown": "[here is my notebook](https://www.kaggle.com/darkcube/fork-of-rsna-pe-competition-stage-1-training)",
      "replies": [
        {
          "id": 1056150,
          "postDate": "2020-10-21T13:09:27.750Z",
          "content": "<p>Did you attach the correct notebook? I don't see your codes.</p>",
          "rawMarkdown": "Did you attach the correct notebook? I don't see your codes.",
          "votes": 1
        },
        {
          "id": 1056158,
          "postDate": "2020-10-21T13:16:56.250Z",
          "content": "<p>Oh, I didn't save the notebook.</p>\n<p>You should now be able to see it.</p>",
          "rawMarkdown": "Oh, I didn't save the notebook.\n\nYou should now be able to see it."
        }
      ]
    },
    {
      "id": 1056041,
      "postDate": "2020-10-21T11:38:23.530Z",
      "content": "<p>can someone please help me on this?<br>\nMy model is giving the same output no matter what the input is. I think its some sort of vanishing gradient problem.</p>\n<p>I'm using EfficientNetB0 followed by a single sigmoid neuron to predict image-level labels. I'm using Ian's jpeg dataset. The data is normalized to the [0, 1] range. I tried a 2 softmax neuron and the problem persisted. I tried imagenet's weights and I tried no weights. I tried implementing binary cross-entropy from scratch. I tried decreasing the learning rate. Nothing worked. I'm about to try a cats-and-dogs dataset and train the model on it and see if the problem is in the data.</p>\n<h3>Additional info:</h3>\n<ol>\n<li>I'm splitting data by patient</li>\n<li>I'm training my model only on patients that have a positive exam for pe</li>\n<li>I'm training by patients; splitting the CT scans into batches of 20 and then doing a training step.</li>\n</ol>\n<p>what might be the problem? is there anything I haven't tried yet?</p>\n<p>Edit: I tried a cat and dog dataset and the problem persists. So it can't be that.<br>\nEdit: it's okay. I solved it. Thank you for trying to help me.</p>",
      "rawMarkdown": "can someone please help me on this?\nMy model is giving the same output no matter what the input is. I think its some sort of vanishing gradient problem.\n\nI'm using EfficientNetB0 followed by a single sigmoid neuron to predict image-level labels. I'm using Ian's jpeg dataset. The data is normalized to the [0, 1] range. I tried a 2 softmax neuron and the problem persisted. I tried imagenet's weights and I tried no weights. I tried implementing binary cross-entropy from scratch. I tried decreasing the learning rate. Nothing worked. I'm about to try a cats-and-dogs dataset and train the model on it and see if the problem is in the data.\n\n### Additional info:\n1. I'm splitting data by patient\n2. I'm training my model only on patients that have a positive exam for pe\n3. I'm training by patients; splitting the CT scans into batches of 20 and then doing a training step.\n\nwhat might be the problem? is there anything I haven't tried yet?\n\nEdit: I tried a cat and dog dataset and the problem persists. So it can't be that.\nEdit: it's okay. I solved it. Thank you for trying to help me."
    },
    {
      "id": 1056346,
      "postDate": "2020-10-21T16:01:25.763Z",
      "rawMarkdown": "",
      "isDeleted": true,
      "replies": [
        {
          "id": 1056349,
          "postDate": "2020-10-21T16:06:24.563Z",
          "content": "<p>I think so yeah. I modified the train.csv so it can better suit what I need but I think I got the labels correctly. I'll make sure to check though</p>",
          "rawMarkdown": "I think so yeah. I modified the train.csv so it can better suit what I need but I think I got the labels correctly. I'll make sure to check though"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1056042,
      "author_name": "DarkCube",
      "author_url": "",
      "post_date": "2020-10-21T11:39:42.777000",
      "content": "<p><a href=\"https://www.kaggle.com/darkcube/fork-of-rsna-pe-competition-stage-1-training\" target=\"_blank\">here is my notebook</a></p>",
      "votes": 0,
      "replies": [
        {
          "id": 1056150,
          "author_name": "Yee Ng",
          "author_url": "",
          "post_date": "2020-10-21T13:09:27.750000",
          "content": "<p>Did you attach the correct notebook? I don't see your codes.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1056158,
          "author_name": "DarkCube",
          "author_url": "",
          "post_date": "2020-10-21T13:16:56.250000",
          "content": "<p>Oh, I didn't save the notebook.</p>\n<p>You should now be able to see it.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1056346,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-10-21T16:01:25.763000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 1056349,
          "author_name": "DarkCube",
          "author_url": "",
          "post_date": "2020-10-21T16:06:24.563000",
          "content": "<p>I think so yeah. I modified the train.csv so it can better suit what I need but I think I got the labels correctly. I'll make sure to check though</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1056042": "[here is my notebook](https://www.kaggle.com/darkcube/fork-of-rsna-pe-competition-stage-1-training)",
    "1056041": "can someone please help me on this?\nMy model is giving the same output no matter what the input is. I think its some sort of vanishing gradient problem.\n\nI'm using EfficientNetB0 followed by a single sigmoid neuron to predict image-level labels. I'm using Ian's jpeg dataset. The data is normalized to the [0, 1] range. I tried a 2 softmax neuron and the problem persisted. I tried imagenet's weights and I tried no weights. I tried implementing binary cross-entropy from scratch. I tried decreasing the learning rate. Nothing worked. I'm about to try a cats-and-dogs dataset and train the model on it and see if the problem is in the data.\n\n### Additional info:\n1. I'm splitting data by patient\n2. I'm training my model only on patients that have a positive exam for pe\n3. I'm training by patients; splitting the CT scans into batches of 20 and then doing a training step.\n\nwhat might be the problem? is there anything I haven't tried yet?\n\nEdit: I tried a cat and dog dataset and the problem persists. So it can't be that.\nEdit: it's okay. I solved it. Thank you for trying to help me.",
    "1056346": ""
  }
}