- read
train_metadata.csv - select only labels with good rating
- slice clips into fragments with
strideandlengthor cut based on detection model - randomly drop samples classes with too much data
- "upsample" classes with too little data
- coarsen coordinates
- add date coarsened up to season (month?, 1/8 of y?)
- read
train_soundscape_labels.csv - add date coarsened date
- add coarsened coordinates
- assume rating is '5'
- assume all labels are primary (?)
- join datasets (short+long+external?)
- treat secondary labels with lower value than 1 (?)
- add folds
- use rating as sample weight (?)
- use secondary labels with label value < 1 and linear activation (?)
- do something about class imbalance
- do augmentation with mixing of random fragments, (predict all labels mixed up)
- try "Cosine Annealing Scheduler with warmup"
- use sounscapes data for fine-tuning
- Baseline: ensemble of 2d and 1d convnets
- Perceiver (?)
- Predictions close by in time of the same bird should be treaded with more confidence
- Try LSTM that takes predicted stream and trains on soundscapes to correct predictions
- Try different ensembling methods: averaging/sqrt(sum(squares))/voting
- Collect more data, particularly on rare classes
- Use only soundscapes for validation to be closer to test dataset
- Assign default rating value more optimally (is there correllation between length/other stuff and rating?)
- Implement SWA https://arxiv.org/pdf/1803.05407.pdf, https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/
- Create synthetic data with GANs
- "Perceiver: General Perception with Iterative Attention" https://arxiv.org/pdf/2103.03206.pdf
- "Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains" https://arxiv.org/pdf/2006.10739.pdf
- https://www.kaggle.com/stefankahl/birdclef2021-exploring-the-data
- http://dcase.community/challenge2018/task-bird-audio-detection-results
- http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Lasseck_76.pdf
- SWA https://arxiv.org/pdf/1803.05407.pdf, https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/
- https://github.com/iver56/audiomentations
- https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html, https://github.com/DemisEom/SpecAugment
- https://enzokro.dev/spectrogram_normalizations/2020/09/10/Normalizing-spectrograms-for-deep-learning.html