{ "id": 370423, "title": "Inverse adversarial validation with self-supervised aproach for signal detection", "url": "/competitions/g2net-detecting-continuous-gravitational-waves/discussion/370423", "author_name": "Pizzaboi", "post_date": "2022-12-04T11:35:26.752000", "votes": 9, "comment_count": 2, "views": 0, "content": "

Hi!
\nLooking at that enormous test data I've got an idea:
\nWhat if we run some self-supervised model on the data (BYOL?) and create a denoising encoder for it?
\nBy the reconstruction loss we can:

\n
    \n
  1. Create a well-fitted encoder for test data. As far as we don't use labels there is no leak.
    2. \n
  2. By the loss of the Self-Supervised model validate if the generated data are similar to the test data
    3. \n
  3. Try to find similar pictures as it's done here: https://www.kaggle.com/code/tanreinama/eliminate-noise-using-signal-similarity
    4. \n
\n

The main problem is to find a proper set of augmentations for this self-supervised model. What do you think? Do you have any ideas about valid augmentations?

", "messages": [ { "id": 2054713, "postDate": "2022-12-04T11:35:26.753Z", "content": "

Hi!
\nLooking at that enormous test data I've got an idea:
\nWhat if we run some self-supervised model on the data (BYOL?) and create a denoising encoder for it?
\nBy the reconstruction loss we can:

\n
    \n
  1. Create a well-fitted encoder for test data. As far as we don't use labels there is no leak.
    2. \n
  2. By the loss of the Self-Supervised model validate if the generated data are similar to the test data
    3. \n
  3. Try to find similar pictures as it's done here: https://www.kaggle.com/code/tanreinama/eliminate-noise-using-signal-similarity
    4. \n
\n

The main problem is to find a proper set of augmentations for this self-supervised model. What do you think? Do you have any ideas about valid augmentations?

", "rawMarkdown": "Hi!\nLooking at that enormous test data I've got an idea:\nWhat if we run some self-supervised model on the data (BYOL?) and create a denoising encoder for it? \nBy the reconstruction loss we can:\n1. Create a well-fitted encoder for test data. As far as we don't use labels there is no leak.\n2. By the loss of the Self-Supervised model validate if the generated data are similar to the test data\n3. Try to find similar pictures as it's done here: https://www.kaggle.com/code/tanreinama/eliminate-noise-using-signal-similarity\n\nThe main problem is to find a proper set of augmentations for this self-supervised model. What do you think? Do you have any ideas about valid augmentations?\n", "votes": 9 }, { "id": 2055232, "postDate": "2022-12-04T22:58:16.857Z", "content": "

In fact, I'm doing something similar but not equal with my model several days ago. It's not equal because augmentations aren't needed, at least in my case. I guess people ahead is trying something similar to this too. I said SIC (Spectrogram Image Classifier) wasn't a enough solution before in this forum. It's necessary something else. The approximation idea with the test set is a good try. Good luck! :)

", "rawMarkdown": "In fact, I'm doing something similar but not equal with my model several days ago. It's not equal because augmentations aren't needed, at least in my case. I guess people ahead is trying something similar to this too. I said SIC (Spectrogram Image Classifier) wasn't a enough solution before in this forum. It's necessary something else. The approximation idea with the test set is a good try. Good luck! :)", "votes": 1 }, { "id": 2054730, "postDate": "2022-12-04T12:03:26.143Z", "content": "

I've writen a test-notebook to learn such a thing:
\nhttps://www.kaggle.com/code/asimandia/test-byol-example

", "rawMarkdown": "I've writen a test-notebook to learn such a thing:\nhttps://www.kaggle.com/code/asimandia/test-byol-example", "votes": 2 } ], "comments": [ { "id": 2055232, "author_name": "Zollkron", "author_url": "", "post_date": "2022-12-04T22:58:16.857000", "content": "

In fact, I'm doing something similar but not equal with my model several days ago. It's not equal because augmentations aren't needed, at least in my case. I guess people ahead is trying something similar to this too. I said SIC (Spectrogram Image Classifier) wasn't a enough solution before in this forum. It's necessary something else. The approximation idea with the test set is a good try. Good luck! :)

", "votes": 1, "replies": [] }, { "id": 2054730, "author_name": "Pizzaboi", "author_url": "", "post_date": "2022-12-04T12:03:26.143000", "content": "

I've writen a test-notebook to learn such a thing:
\nhttps://www.kaggle.com/code/asimandia/test-byol-example

", "votes": 2, "replies": [] } ], "raw_markdown_by_id": { "2054713": "Hi!\nLooking at that enormous test data I've got an idea:\nWhat if we run some self-supervised model on the data (BYOL?) and create a denoising encoder for it? \nBy the reconstruction loss we can:\n1. Create a well-fitted encoder for test data. As far as we don't use labels there is no leak.\n2. By the loss of the Self-Supervised model validate if the generated data are similar to the test data\n3. Try to find similar pictures as it's done here: https://www.kaggle.com/code/tanreinama/eliminate-noise-using-signal-similarity\n\nThe main problem is to find a proper set of augmentations for this self-supervised model. What do you think? Do you have any ideas about valid augmentations?\n", "2055232": "In fact, I'm doing something similar but not equal with my model several days ago. It's not equal because augmentations aren't needed, at least in my case. I guess people ahead is trying something similar to this too. I said SIC (Spectrogram Image Classifier) wasn't a enough solution before in this forum. It's necessary something else. The approximation idea with the test set is a good try. Good luck! :)", "2054730": "I've writen a test-notebook to learn such a thing:\nhttps://www.kaggle.com/code/asimandia/test-byol-example" } }