Skip to content

Latest commit

 

History

History
28 lines (16 loc) · 1.6 KB

File metadata and controls

28 lines (16 loc) · 1.6 KB

Rainforest Connection Species Audio Detection

Given animal audio clip, correctly classify them

A post-competition organization of the code used in the competition. Because i find the resources provided by Kaggle (30h free TPU) to be sufficient, most of the code are written in the kernel notebook in kaggle platform (although i am not really a fan of using it). I started using the baseline code written by yosshi999 and made changes along the way, implementing my own model architectures.

What's new?

This solution places at top 9%. However, I did manage to have a single model that reaches 0.891. More diversification might result in a better score. My implementations includes some of the following:

  1. random resized crop as augmentation
  2. custom channel attention layer + spatial attention layer + CNN blocks inspired by a winner from a similar competition. The model design can be found in rainforest_audio/model.py

Run

Run all 7 densenet notebook and generate the data. Place them in their respective data folder. Also, get one of a public notebook. All the csv are uploaded in the /data/ folder.

To generate the submission, run:

$ make run_submission

Post-competition

Mask-loss did not worked for me. Have yet to try re-training all my models with Lsoft objective function. Thing that did not work are placed in archive folder within the notebook.

I have also added a summary of learning points and experience in a pdf document based on an internal sharing. It is found in Rainforest Audio Detection.pdf