Spectogram classification

Description

This project includes all the code used during a private kaggle competition. This solution got 7th place on both public and private leaderboard.

Requirements

• python 3.6
• pytorch 1.0
• tensorboardX

Best solution description

1. Spectograms are split to 32 x 32 parts and the task is to classify which spectogram the part belongs to (there 3 parts).
2. Data is normalised to range [0, 1]
3. Achitecture consists of convolutional feature extractor and capsulenet head optimised according to this article.

Additional info

For additional information / used data / trained models - send me a private message or an e-mail at kacp.kania@gmail.com.