This week ended another competition by >>Harvard Medical School (HMS) also with time-frequency spectrograms.
\nWhile the HMS competition is substantially different, we could take a lot from the HMS winning solutions to apply in this competition.
This could be a good starting point to see what a winning solution with a similar data type could look like. Of course, the goal in this competition is absolutely different, as well as the data in this competition is comprised of a single audio channel, as opposed to 19 EEG + 1 ECG channels in the HMS.
", "messages": [ { "id": "2748469", "postDate": "04/12/2024 13:15:28", "content": "This week ended another competition by >>Harvard Medical School (HMS) also with time-frequency spectrograms.
\nWhile the HMS competition is substantially different, we could take a lot from the HMS winning solutions to apply in this competition.
This could be a good starting point to see what a winning solution with a similar data type could look like. Of course, the goal in this competition is absolutely different, as well as the data in this competition is comprised of a single audio channel, as opposed to 19 EEG + 1 ECG channels in the HMS.
", "rawMarkdown": "This week ended another competition by >>[Harvard Medical School (HMS)](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/code) also with time-frequency spectrograms. \nWhile the HMS competition is substantially different, we could take a lot from the HMS winning solutions to apply in this competition.\n\n1. [Audio Course](https://huggingface.co/learn/audio-course/en/chapter0/introduction) by HuggingFace - it is a short course and will teach you all the necessary details about digital signals, different types of spectrograms and transformers for audio\n2. Gold Medal solution notebook from HMS Comp. by @cdeotte [link](https://www.kaggle.com/code/cdeotte/single-model-gold-solution-cv-0-24-lb-0-24)\n3. 2-stage learning for the cases if data quality is not uniform in the dataset or if there is a subsample of not-so-well annotated data; by @seanbearden [link](https://www.kaggle.com/code/seanbearden/effnetb0-2-pop-model-train-twice-lb-0-39)\n4. HMS Comp. 1st place solution write-up by Team Sony @sugupoko [link](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/492560)\n5. GPU-accelerated spectrograms with cupy by @sergiosaharovskiy [link](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/487110)\n6. HMS 3rd Place solution write-up by @christofhenkel [link](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/492471)\n6. General introduction to time-frequency analysis methods can be found as links to the YouTube Channel of Mike X Cohen in my earlier discussion in HMS [link](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/466768)\n\n\n\nThis could be a good starting point to see what a winning solution with a similar data type could look like. Of course, the goal in this competition is absolutely different, as well as the data in this competition is comprised of a single audio channel, as opposed to 19 EEG + 1 ECG channels in the HMS.", "votes": null }, { "id": "2750598", "postDate": "04/13/2024 18:48:08", "content": "Thanks for the reference. My published Gold Medal notebook here demonstrates how to create spectrograms inside a PyTorch dataloader. This allows us to apply data augmentation both before spectrogram (to raw waveform) and after spectrogram (to images). This improves CV score and LB score!
\nAnd it allows us to perform quick experiments with different spectrogram hyperparameters.
", "rawMarkdown": "Thanks for the reference. My published Gold Medal notebook [here][1] demonstrates how to create spectrograms inside a PyTorch dataloader. This allows us to apply data augmentation both **before** spectrogram (to raw waveform) and **after** spectrogram (to images). This improves CV score and LB score!\n\nAnd it allows us to perform quick experiments with different spectrogram hyperparameters.\n\n[1]: https://www.kaggle.com/code/cdeotte/single-model-gold-solution-cv-0-24-lb-0-24", "votes": null } ], "comments": [ { "id": 2750598, "author_name": "cdeotte", "author_url": "", "post_date": "04/13/2024 18:48:08", "content": "Thanks for the reference. My published Gold Medal notebook here demonstrates how to create spectrograms inside a PyTorch dataloader. This allows us to apply data augmentation both before spectrogram (to raw waveform) and after spectrogram (to images). This improves CV score and LB score!
\nAnd it allows us to perform quick experiments with different spectrogram hyperparameters.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2748469": "This week ended another competition by >>[Harvard Medical School (HMS)](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/code) also with time-frequency spectrograms. \nWhile the HMS competition is substantially different, we could take a lot from the HMS winning solutions to apply in this competition.\n\n1. [Audio Course](https://huggingface.co/learn/audio-course/en/chapter0/introduction) by HuggingFace - it is a short course and will teach you all the necessary details about digital signals, different types of spectrograms and transformers for audio\n2. Gold Medal solution notebook from HMS Comp. by @cdeotte [link](https://www.kaggle.com/code/cdeotte/single-model-gold-solution-cv-0-24-lb-0-24)\n3. 2-stage learning for the cases if data quality is not uniform in the dataset or if there is a subsample of not-so-well annotated data; by @seanbearden [link](https://www.kaggle.com/code/seanbearden/effnetb0-2-pop-model-train-twice-lb-0-39)\n4. HMS Comp. 1st place solution write-up by Team Sony @sugupoko [link](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/492560)\n5. GPU-accelerated spectrograms with cupy by @sergiosaharovskiy [link](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/487110)\n6. HMS 3rd Place solution write-up by @christofhenkel [link](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/492471)\n6. General introduction to time-frequency analysis methods can be found as links to the YouTube Channel of Mike X Cohen in my earlier discussion in HMS [link](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/466768)\n\n\n\nThis could be a good starting point to see what a winning solution with a similar data type could look like. Of course, the goal in this competition is absolutely different, as well as the data in this competition is comprised of a single audio channel, as opposed to 19 EEG + 1 ECG channels in the HMS.", "2750598": "Thanks for the reference. My published Gold Medal notebook [here][1] demonstrates how to create spectrograms inside a PyTorch dataloader. This allows us to apply data augmentation both **before** spectrogram (to raw waveform) and **after** spectrogram (to images). This improves CV score and LB score!\n\nAnd it allows us to perform quick experiments with different spectrogram hyperparameters.\n\n[1]: https://www.kaggle.com/code/cdeotte/single-model-gold-solution-cv-0-24-lb-0-24" }, "source": "meta" }