/Diploma-thesis

Title: "Investigation of acoustic scene classification techniques"

Primary LanguageJupyter NotebookMIT LicenseMIT

Diploma-thesis

Title: "Investigation of acoustic scene classification techniques"

Versions used:

  • Keras 2.3.0
  • Tensorflow-gpu 2.1.0
  • Librosa 0.7.1
  • Pysoundfile 0.10.3

Jupyter notebooks

  • Run Training_stage.ipynb to load filenames and labels, pre-process your data (choose either Per-Channel Energy Normalization or log-Mel power spectrograms) and train your model. Multiple CNN architectures are available in the repository (see corresponding Python scripts).

  • Run MLP+ETi.ipynb for feature extraction and Enhanced Temporal Integration. An efficient MLP architecture is provided for training (can be trained on CPU, ~ 1 sec per epoch).

  • Run grad-CAM.ipynb for 'Gradient-weighted Class Activation Mapping' visualization technique. Make sure you use the same pre-processing step throughout this process.

Also, keep an eye on the comments since they provide useful explanations for every step.