Code for transfer learning and fine-tuning with YAMNet

This repo requires the code and weights file from the official yamnet directory to be downloaded: https://github.com/tensorflow/models/tree/master/research/audioset/yamnet.

• generate YAMNet input.ipynb pre-computes spectrograms from audio files and stores them as .npy these are then loaded during training using a custom Tensorflow data generator defined in datagen_yamnet.py. Pre-computing the spectrograms greatly increases the speed of training / experimentation.

• fine-tune YAMNet.ipynb sets up spectrogram data generators, loads AudioSet weights into YAMNet and adds a custom network top, and runs transfer learning and fine-tuning.

References

• https://github.com/tensorflow/models/tree/master/research/audioset/yamnet

• Gemmeke, J. et. al., AudioSet: An ontology and human-labelled dataset for audio events, ICASSP 2017

• Hershey, S. et. al., CNN Architectures for Large-Scale Audio Classification, ICASSP 2017