Here another wave competition after :
\nUsually, moving from wave data to images would lead to better performance. One could be tempted to go the other way: feeding the time series to a recurrent neural net or any time aware neural model (Wavenet, Recurrent CNN, Temporal CNN …) but the fact is that, advances in deep image processing are such that transforming the waves into images would \"always\" give better results. Raw wave-based models would often suffer from deadly overfitting.
\nThere are many choices going from time series data (waves, sounds, …) to images:
\nAfter computing these images, you can feed them to any image model, usually one of:
\nInteresting alternative and new models could also perform very well : ViT, DeiT, CaiT, NFNet.
\nBut, be aware : spectograms are NOT like other images ! Essentially because they include some time dimension. To help models dealing with this specificity, one can use a SED model or a well designed pooling strategy.
\nThat's all for now. And you, how are you dealing with this competition ?
", "messages": [ { "id": "1374056", "postDate": "07/03/2021 01:34:24", "content": "Here another wave competition after :
\nUsually, moving from wave data to images would lead to better performance. One could be tempted to go the other way: feeding the time series to a recurrent neural net or any time aware neural model (Wavenet, Recurrent CNN, Temporal CNN …) but the fact is that, advances in deep image processing are such that transforming the waves into images would \"always\" give better results. Raw wave-based models would often suffer from deadly overfitting.
\nThere are many choices going from time series data (waves, sounds, …) to images:
\nAfter computing these images, you can feed them to any image model, usually one of:
\nInteresting alternative and new models could also perform very well : ViT, DeiT, CaiT, NFNet.
\nBut, be aware : spectograms are NOT like other images ! Essentially because they include some time dimension. To help models dealing with this specificity, one can use a SED model or a well designed pooling strategy.
\nThat's all for now. And you, how are you dealing with this competition ?
", "rawMarkdown": "Here another wave competition after :\n* [Cornell Birdcall Identification](https://www.kaggle.com/c/birdsong-recognition)\n* [Rainforest Connection Species Audio Detection](https://www.kaggle.com/c/rfcx-species-audio-detection)\n* [BirdCLEF 2021 Birdcall Identification]( https://www.kaggle.com/c/birdclef-2021)\n\nUsually, moving from wave data to images would lead to better performance. One could be tempted to go the other way: feeding the time series to a recurrent neural net or any time aware neural model (Wavenet, Recurrent CNN, Temporal CNN ...) but the fact is that, advances in deep image processing are such that transforming the waves into images would \"always\" give better results. Raw wave-based models would often suffer from deadly overfitting.\n\nThere are many choices going from time series data (waves, sounds, ...) to images:\n* [Fourier transform](https://en.wikipedia.org/wiki/Fourier_transform)\n* [Mel Spectogram](https://fairyonice.github.io/implement-the-spectrogram-from-scratch-in-python.html)\n* [MFCC: Mel-Frequency Cepstral Coefficients](https://en.wikipedia.org/wiki/Mel-frequency_cepstrum)\n* [PCEN: Per-Channel Energy Normalization](https://arxiv.org/pdf/1607.05666.pdf)\n\nAfter computing these images, you can feed them to any image model, usually one of:\n* The **EfficientNet** family : **ns**, **ap**, **em**, **v2** ...\n* The **ResNet** family : **ResNet**, **ResNest**, **ResNext** ...\n\nInteresting alternative and new models could also perform very well : ViT, DeiT, CaiT, NFNet.\n\nBut, be aware : spectograms are **NOT** like other images ! Essentially because they include some time dimension. To help models dealing with this specificity, one can use a **SED** model or a well designed pooling strategy.\n\n\nThat's all for now. And you, how are you dealing with this competition ?", "votes": null }, { "id": "1375744", "postDate": "07/04/2021 12:53:21", "content": "one can use a SED model what is this SED model you are referring to?
Spectral energy distribution (SED) model is being referred here, below link is helpful for more details
\nhttp://sed.sao.ru/sed_E_models.html
\n@kneroma : Kindly correct me if wrong here.
", "rawMarkdown": "Spectral energy distribution (SED) model is being referred here, below link is helpful for more details\n\nhttp://sed.sao.ru/sed_E_models.html\n\n@kneroma : Kindly correct me if wrong here.", "votes": null }, { "id": "1376209", "postDate": "07/05/2021 00:13:07", "content": "@saurabhbagchi, it may also mean sound detection model from previous birdcall competitions. I could be wrong though.
", "rawMarkdown": "saurabhbagchi, it may also mean sound detection model from previous birdcall competitions. I could be wrong though.", "votes": null }, { "id": "1376333", "postDate": "07/05/2021 04:06:34", "content": "That appears to be more correct @pukkinming ! Thanks for pointing out!
", "rawMarkdown": "That appears to be more correct @pukkinming ! Thanks for pointing out!", "votes": null }, { "id": "1376947", "postDate": "07/05/2021 13:07:55", "content": "@saurabhbagchi thanks for trying to reply, but I was talking about Sound Event Detection (SED) models. These models are able to make a good use of the time axis.
", "rawMarkdown": "saurabhbagchi thanks for trying to reply, but I was talking about Sound Event Detection (**SED**) models. These models are able to make a good use of the time axis.", "votes": null }, { "id": "1379477", "postDate": "07/07/2021 12:04:39", "content": "Was just reading some winning solution of a past 2 years ago signal processing competition , people were using Lightgbm and features related to Wave , it seems now Cnn is a go to tool for Wave Data
", "rawMarkdown": "Was just reading some winning solution of a past 2 years ago signal processing competition , people were using Lightgbm and features related to Wave , it seems now Cnn is a go to tool for Wave Data", "votes": null }, { "id": "1380064", "postDate": "07/07/2021 19:03:14", "content": "Looks this competition is not same as Audio competition as you mentioned @kneroma , the given GW signal is not like audio waveform and the sample rate is very low as 2048Hz, do you think MelSpectrogram is still the best presentation for this competition, or CQT is a better option?
", "rawMarkdown": "Looks this competition is not same as Audio competition as you mentioned @kneroma , the given GW signal is not like audio waveform and the sample rate is very low as 2048Hz, do you think MelSpectrogram is still the best presentation for this competition, or CQT is a better option?", "votes": null }, { "id": "1563263", "postDate": "10/28/2021 06:52:18", "content": "Hey,
\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.
\nThank you all,
\nRegards,
\nChris
one can use a SED model what is this SED model you are referring to?
Spectral energy distribution (SED) model is being referred here, below link is helpful for more details
\nhttp://sed.sao.ru/sed_E_models.html
\n@kneroma : Kindly correct me if wrong here.
", "votes": null, "replies": [] }, { "id": 1376209, "author_name": "pukkinming", "author_url": "", "post_date": "07/05/2021 00:13:07", "content": "@saurabhbagchi, it may also mean sound detection model from previous birdcall competitions. I could be wrong though.
", "votes": null, "replies": [] }, { "id": 1376333, "author_name": "saurabhbagchi", "author_url": "", "post_date": "07/05/2021 04:06:34", "content": "That appears to be more correct @pukkinming ! Thanks for pointing out!
", "votes": null, "replies": [] }, { "id": 1376947, "author_name": "kneroma", "author_url": "", "post_date": "07/05/2021 13:07:55", "content": "@saurabhbagchi thanks for trying to reply, but I was talking about Sound Event Detection (SED) models. These models are able to make a good use of the time axis.
", "votes": null, "replies": [] } ] }, { "id": 1379477, "author_name": "sayedathar11", "author_url": "", "post_date": "07/07/2021 12:04:39", "content": "Was just reading some winning solution of a past 2 years ago signal processing competition , people were using Lightgbm and features related to Wave , it seems now Cnn is a go to tool for Wave Data
", "votes": null, "replies": [] }, { "id": 1380064, "author_name": "superchenhao", "author_url": "", "post_date": "07/07/2021 19:03:14", "content": "Looks this competition is not same as Audio competition as you mentioned @kneroma , the given GW signal is not like audio waveform and the sample rate is very low as 2048Hz, do you think MelSpectrogram is still the best presentation for this competition, or CQT is a better option?
", "votes": null, "replies": [] }, { "id": 1563263, "author_name": "zerafachris", "author_url": "", "post_date": "10/28/2021 06:52:18", "content": "Hey,
\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.
\nThank you all,
\nRegards,
\nChris