By Tobias Toft Christensen, Mikkel Heber Hahn Petersen, and Anders Hansen Warming.
This repository is related to a paper presenting an acoustic scene classification method, which uses transfer learning on a VGGish pre-trained model. Transfer learning is a method where knowledge from solving one problem is gained and stored, and can subsequently be used and applied to a related problem. The performance of this method is evaluated on the TUT Acoustic Scenes 2017 data set. A data set collected in Finland by Tampere University of technology. The data collection has received funding from the European Research Council and is part of a DCASE \textit{(Detection and Classification of Acoustic Scenes and Events)} 2017 challenge. The project is related to the DTU course 02456 Deep Learning.
The project are written in Python programming language and some of the scripts is formatted into Jupyter Notebooks.
This repository contains a folder "tfRecordsReal" with the tfRecords for the training data, the validation data and the test data. Further more the resulting model is compiled and saved in the folder 3ClassModel and 15ClassModel, respectively for the 3 class and the 15 class classification problem. This repository does NOT contain the pre-trained VGGish model (ref: https://github.com/tensorflow/models/tree/master/research/audioset), but can be obtained from the link. The pre-trained VGGish model is due to the tfRecords, not needed to reproduce the project results.
Finally this repository contain a Jupyter Notebook file (AcousticSceneClassifier.ipynb) with our model.
To reproduce the results illustrated in the Article the two trained models are saved, and can be restored. The two saved models are respectively the 3 class problem and the 15 class problem. Only test classification accuracy for the 3 class problem is illustrated in the Jupyter Notebook. The test for the 15 class problem is preformed the same way, just tuned to 15 classes. The code can be found in the repository under "restoreTest15Classes.py".