This library provides "data source" objects that provide training material for my neural network experiments. It also contains functionality to load audio datasets that come in the specific directory structure used at the Department of Computational Perception, and thus by me.
This library is primarily for my personal use, so it is tailored to my development environment. Although I made some effort to make it more general, documentation might be lacking, and there might be assumptions that do not hold in general. I'm sorry.
You might want to look at the following repositories:
- nn: Helper functions for training Lasagne neural networks.
- chordrec: My experimental framework to train neural network-based models for chord recognition. This is the actual reason why the dmgr and nn libraries exist.
Each dataset has the following directory structure:
+-- dataset_name/
+-- audio/
+-- *.flac
+-- annotations/
+-- beats/
+-- *.beats
+-- chords/
+-- *.chords
...
+-- splits/
+-- 8-fold_cv_*.fold
Names of annotation and audio files need to correspond. dmgr does not require
these files to be in separate directories; it will search the whole dataset
directory recursively. However, the crossvalidation fold definitions have
to be in a splits sub-directory. Each .fold file contains a list of file
names (one per line, without file extension) that represent the test files of
this fold. The rest of the files can be used for training.