I released the denoise spectrograms dataset.
\nhttps://www.kaggle.com/takamichitoda/rfcx-denoise-melspec
This dataset is calculated from this notebook.
\nhttps://www.kaggle.com/takamichitoda/spectrogram-generation-with-denoise
\n(This notebook made from Theo Viel's notebook. Thanks.)
This competition train dataset has a lot of noise but the test set looks clean.
\nSo, I think that if I use the cleaned train set, the LB is improved.
I'm just starting to investigate, but I'll share my knowledge in this thread.
\nUPDATE: 2021/01/26
\nI finished training but the result is not good.
\n| \n | CV | \nLB | \n
|---|---|---|
| My Best Model | \n0.9260 | \n0.903 | \n
| My Best Model with denoise | \n0.8954 | \n0.858 | \n
I think tolerating noise is more important than denoise.
", "messages": [ { "id": "1168530", "postDate": "01/25/2021 03:09:18", "content": "I released the denoise spectrograms dataset.
\nhttps://www.kaggle.com/takamichitoda/rfcx-denoise-melspec
This dataset is calculated from this notebook.
\nhttps://www.kaggle.com/takamichitoda/spectrogram-generation-with-denoise
\n(This notebook made from Theo Viel's notebook. Thanks.)
This competition train dataset has a lot of noise but the test set looks clean.
\nSo, I think that if I use the cleaned train set, the LB is improved.
I'm just starting to investigate, but I'll share my knowledge in this thread.
\nUPDATE: 2021/01/26
\nI finished training but the result is not good.
\n| \n | CV | \nLB | \n
|---|---|---|
| My Best Model | \n0.9260 | \n0.903 | \n
| My Best Model with denoise | \n0.8954 | \n0.858 | \n
I think tolerating noise is more important than denoise.
", "rawMarkdown": "I released the denoise spectrograms dataset.\nhttps://www.kaggle.com/takamichitoda/rfcx-denoise-melspec\n\nThis dataset is calculated from this notebook.\nhttps://www.kaggle.com/takamichitoda/spectrogram-generation-with-denoise\n(This notebook made from [Theo Viel's notebook](https://www.kaggle.com/theoviel/spectrogram-generation). Thanks.)\n\nThis competition train dataset has a lot of noise but the test set looks clean. \nSo, I think that if I use the cleaned train set, the LB is improved.\n\nI'm just starting to investigate, but I'll share my knowledge in this thread.\n\n---\nUPDATE: 2021/01/26\n\nI finished training but the result is not good.\n\n||CV|LB|\n|--|--|--|\n|My Best Model|0.9260|0.903|\n|My Best Model with denoise|0.8954|0.858|\n\nI think tolerating noise is more important than denoise.", "votes": null }, { "id": "1170347", "postDate": "01/26/2021 06:53:45", "content": "But a good attempt all the same! Wonder why this happens…Maybe we are removing some original data as well when we remove noise or maybe the background noise plays a role in determining the bird call..certain birds found in certain habitat (background noise of various insects and even birds could vary based on target in question)..After all birds of same feather flock together :)
", "rawMarkdown": "But a good attempt all the same! Wonder why this happens...Maybe we are removing some original data as well when we remove noise or maybe the background noise plays a role in determining the bird call..certain birds found in certain habitat (background noise of various insects and even birds could vary based on target in question)..After all birds of same feather flock together :)", "votes": null }, { "id": "1171194", "postDate": "01/26/2021 17:15:48", "content": "Interesting work. I think as @allohvk mentioned maybe some unseen signal is being destroyed. I would also expect such a simple denoising method would likely be able to be simulated by the CNN's anyway so maybe some of that is implicit in the model already.
", "rawMarkdown": "Interesting work. I think as @allohvk mentioned maybe some unseen signal is being destroyed. I would also expect such a simple denoising method would likely be able to be simulated by the CNN's anyway so maybe some of that is implicit in the model already.", "votes": null }, { "id": "1171619", "postDate": "01/27/2021 01:50:38", "content": "the denoise is energy based, but sometimes the bird we want is not the most loud in the train sample (other species are louder). I think that might give us problem
", "rawMarkdown": "the denoise is energy based, but sometimes the bird we want is not the most loud in the train sample (other species are louder). I think that might give us problem", "votes": null }, { "id": "1172130", "postDate": "01/27/2021 09:17:15", "content": "Thank you comment!
\n\n\nmaybe the background noise plays a role in determining the bird call
\n
I didn't think of it, but it may be true.
\nThank you for sharing your idea.
\nThe difficulty is that we can't detect audio that we shouldn't use.
I think adding noise to test data may work better than removing noise from train data. So, I'm trying that now.
", "rawMarkdown": "Thank you comment!\n\n> maybe the background noise plays a role in determining the bird call\n\nI didn't think of it, but it may be true.\nThank you for sharing your idea.\nThe difficulty is that we can't detect audio that we shouldn't use.\n\nI think adding noise to test data may work better than removing noise from train data. So, I'm trying that now.", "votes": null }, { "id": "1172206", "postDate": "01/27/2021 09:54:27", "content": "Thank you comment.
\nI think so too, so I'm training tolerating noise model.
Yes, it bothers me with processes other than denoising too…
", "rawMarkdown": "Yes, it bothers me with processes other than denoising too...", "votes": null }, { "id": "1172235", "postDate": "01/27/2021 10:07:23", "content": "Thanks for trying and sharing the (negative) results. You are saving time for others, which is unvaluable.
\nI had a hunch that you would remove some signal along denoising, but seeing it is very helpful.
", "rawMarkdown": "Thanks for trying and sharing the (negative) results. You are saving time for others, which is unvaluable.\n\n I had a hunch that you would remove some signal along denoising, but seeing it is very helpful.", "votes": null }, { "id": "1172510", "postDate": "01/27/2021 12:06:12", "content": "Thank you.
\nThis did not work, but there are many other processes which should be tried in this competition.
\nI continue to improve my model.
Thank you for sharing your insights!
\nHow about running a noise removal script on the test data?
\nEven if the test data is somewhat clean, I felt it would be appropriate to run the same process on the test data as on the training data.
Thank you comment and mention.
\n\n\nHow about running a noise removal script on the test data?
\n
I tried it. The result is LB=0.857.
\nIt has downed a little.
But a good attempt all the same! Wonder why this happens…Maybe we are removing some original data as well when we remove noise or maybe the background noise plays a role in determining the bird call..certain birds found in certain habitat (background noise of various insects and even birds could vary based on target in question)..After all birds of same feather flock together :)
", "votes": null, "replies": [ { "id": 1172130, "author_name": "takamichitoda", "author_url": "", "post_date": "01/27/2021 09:17:15", "content": "Thank you comment!
\n\n\nmaybe the background noise plays a role in determining the bird call
\n
I didn't think of it, but it may be true.
\nThank you for sharing your idea.
\nThe difficulty is that we can't detect audio that we shouldn't use.
I think adding noise to test data may work better than removing noise from train data. So, I'm trying that now.
", "votes": null, "replies": [] } ] }, { "id": 1171194, "author_name": "ryches", "author_url": "", "post_date": "01/26/2021 17:15:48", "content": "Interesting work. I think as @allohvk mentioned maybe some unseen signal is being destroyed. I would also expect such a simple denoising method would likely be able to be simulated by the CNN's anyway so maybe some of that is implicit in the model already.
", "votes": null, "replies": [ { "id": 1172206, "author_name": "takamichitoda", "author_url": "", "post_date": "01/27/2021 09:54:27", "content": "Thank you comment.
\nI think so too, so I'm training tolerating noise model.
the denoise is energy based, but sometimes the bird we want is not the most loud in the train sample (other species are louder). I think that might give us problem
", "votes": null, "replies": [ { "id": 1172213, "author_name": "takamichitoda", "author_url": "", "post_date": "01/27/2021 09:58:39", "content": "Yes, it bothers me with processes other than denoising too…
", "votes": null, "replies": [] } ] }, { "id": 1172235, "author_name": "cpmpml", "author_url": "", "post_date": "01/27/2021 10:07:23", "content": "Thanks for trying and sharing the (negative) results. You are saving time for others, which is unvaluable.
\nI had a hunch that you would remove some signal along denoising, but seeing it is very helpful.
", "votes": null, "replies": [ { "id": 1172510, "author_name": "takamichitoda", "author_url": "", "post_date": "01/27/2021 12:06:12", "content": "Thank you.
\nThis did not work, but there are many other processes which should be tried in this competition.
\nI continue to improve my model.
Thank you for sharing your insights!
\nHow about running a noise removal script on the test data?
\nEven if the test data is somewhat clean, I felt it would be appropriate to run the same process on the test data as on the training data.
Thank you comment and mention.
\n\n\nHow about running a noise removal script on the test data?
\n
I tried it. The result is LB=0.857.
\nIt has downed a little.