Urban Sounds Classification

Final project of Deep Learning Bootcamp that organised by Global AI Hub and Koc Holding.

Preprocessing:

• Images augmented by shifting spectrograms by 50% to reduce class imbalance (applied to small classes)
• Applied min-max normalization (dividing image values by 255)

Hyperparameter search:

• Hyperopt library used for hyperparameter search
• Best parameters:
  • activation : tanh
  • filters: [32,64,64,128]
  • kernel_size: (3, 3)
  • learning_rate: 4.623998518043979e-05
  • optimizer: Adam

Results: Orange is validation result

Test metrics:

Accuracy: 0.8662821054458618

F1 Score: 0.8702771229110035

Precision: 0.8702702403508666

Recall: 0.8710366547578786