{
  "id": 233539,
  "title": "Preprocessing data & models",
  "url": "/competitions/birdclef-2021/discussion/233539",
  "author_name": "",
  "post_date": "2021-04-19T19:41:29.226909500Z",
  "votes": 6,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hello everyone! I had a quick question about the competition submission rules! This is my first code competition on Kaggle so maybe the answer is simple.</p>\n<p>I see that external data is allowed if it is public and accessible. This means we can use pre-processed spectrograms or pre-trained models when submitting.</p>\n<p>However, are we also allowed to use saved processed versions of these datasets and models? If I generate a spectrogram dataset but leave it private can I still use these images in my final submission if the original dataset is public?</p>\n<p>And is this also true for model weights? (Can I train my model on my computer and then save the weights to a private dataset for the final submission? What if I did this but also tuned the weights during the final submission?)</p>\n<p>Thank you!</p>",
  "messages": [
    {
      "id": "1278358",
      "postDate": "04/19/2021 19:41:29",
      "content": "<p>Hello everyone! I had a quick question about the competition submission rules! This is my first code competition on Kaggle so maybe the answer is simple.</p>\n<p>I see that external data is allowed if it is public and accessible. This means we can use pre-processed spectrograms or pre-trained models when submitting.</p>\n<p>However, are we also allowed to use saved processed versions of these datasets and models? If I generate a spectrogram dataset but leave it private can I still use these images in my final submission if the original dataset is public?</p>\n<p>And is this also true for model weights? (Can I train my model on my computer and then save the weights to a private dataset for the final submission? What if I did this but also tuned the weights during the final submission?)</p>\n<p>Thank you!</p>",
      "rawMarkdown": "Hello everyone! I had a quick question about the competition submission rules! This is my first code competition on Kaggle so maybe the answer is simple.\n\nI see that external data is allowed if it is public and accessible. This means we can use pre-processed spectrograms or pre-trained models when submitting.\n\nHowever, are we also allowed to use saved processed versions of these datasets and models? If I generate a spectrogram dataset but leave it private can I still use these images in my final submission if the original dataset is public?\n\nAnd is this also true for model weights? (Can I train my model on my computer and then save the weights to a private dataset for the final submission? What if I did this but also tuned the weights during the final submission?)\n\nThank you!",
      "votes": null
    },
    {
      "id": "1279914",
      "postDate": "04/21/2021 11:15:15",
      "content": "<p>Yes, you can of course train your own local model and then use the weights for your final submission. It's different if it's someone else's model (e.g., Mobile Net pre-trained on ImageNet), these weights need to be publicly accessible. Same goes for datasets (e.g., Google AudioSet). If it's publicly available, you can use it (but you'll have to disclose it!).</p>",
      "rawMarkdown": "Yes, you can of course train your own local model and then use the weights for your final submission. It's different if it's someone else's model (e.g., Mobile Net pre-trained on ImageNet), these weights need to be publicly accessible. Same goes for datasets (e.g., Google AudioSet). If it's publicly available, you can use it (but you'll have to disclose it!).",
      "votes": null
    },
    {
      "id": "1280100",
      "postDate": "04/21/2021 15:03:44",
      "content": "<p>Thank you so much for clarification! This is extremely helpful since I was previously worried about the time limits for my pre-processing!</p>\n<p>Thank you!</p>",
      "rawMarkdown": "Thank you so much for clarification! This is extremely helpful since I was previously worried about the time limits for my pre-processing!\n\nThank you!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1279914,
      "author_name": "stefankahl",
      "author_url": "",
      "post_date": "04/21/2021 11:15:15",
      "content": "<p>Yes, you can of course train your own local model and then use the weights for your final submission. It's different if it's someone else's model (e.g., Mobile Net pre-trained on ImageNet), these weights need to be publicly accessible. Same goes for datasets (e.g., Google AudioSet). If it's publicly available, you can use it (but you'll have to disclose it!).</p>",
      "votes": null,
      "replies": [
        {
          "id": 1280100,
          "author_name": "aarons0",
          "author_url": "",
          "post_date": "04/21/2021 15:03:44",
          "content": "<p>Thank you so much for clarification! This is extremely helpful since I was previously worried about the time limits for my pre-processing!</p>\n<p>Thank you!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1278358": "Hello everyone! I had a quick question about the competition submission rules! This is my first code competition on Kaggle so maybe the answer is simple.\n\nI see that external data is allowed if it is public and accessible. This means we can use pre-processed spectrograms or pre-trained models when submitting.\n\nHowever, are we also allowed to use saved processed versions of these datasets and models? If I generate a spectrogram dataset but leave it private can I still use these images in my final submission if the original dataset is public?\n\nAnd is this also true for model weights? (Can I train my model on my computer and then save the weights to a private dataset for the final submission? What if I did this but also tuned the weights during the final submission?)\n\nThank you!",
    "1279914": "Yes, you can of course train your own local model and then use the weights for your final submission. It's different if it's someone else's model (e.g., Mobile Net pre-trained on ImageNet), these weights need to be publicly accessible. Same goes for datasets (e.g., Google AudioSet). If it's publicly available, you can use it (but you'll have to disclose it!).",
    "1280100": "Thank you so much for clarification! This is extremely helpful since I was previously worried about the time limits for my pre-processing!\n\nThank you!"
  },
  "source": "meta"
}