General organization

[YannisBilly]

• [x] Read and read more
• [x] Download the data
• [x] Download annotations
• [x] Make a parser with slicing in time domain
• [x] Make a Data Loader
• [x] Search for SOTA, r2g models and download them
• [x] Fine-Tune on one song to test if this can be a think
• [ ] Weighted cross entropy must be used as the output is sparse (several thousand no beat but 70-80 beats per image.
• [ ] Import focal_tversky loss for multi-label classification from https://github.com/nabsabraham/focal-tversky-unet/blob/master/losses.py

[YannisBilly]

I will implement Transfer Learning from ImageNet based on the works of: Palanisamy, K., Singhania, D., & Yao, A. (2020). Rethinking CNN Models for Audio Classification (arXiv:2007.11154). arXiv. http://arxiv.org/abs/2007.11154

I will try to make the case that the probem can be seen as a single classificaiton problem with Transfer Learning from the Image Domain.

I will test if quantization like an image can play a role.

[YannisBilly]

If we see Beat Tracking as a binary classification problem it can be seen also as a Outlier Estimation (as a beat is an outlier with respect to the content of a song).