Releases: TUT-ARG/DCASE2017-baseline-system
Releases · TUT-ARG/DCASE2017-baseline-system
Introducing more Keras related functionality
- Introduced recognizer classes: SceneRecognizer, EventRecognizer
- Introduced Keras callbacks: ProgressLoggerCallback, ProgressPlotterCallback, StopperCallback, StasherCallback
- Added new learner classes: SceneClassifierKerasSequential, EventDetectorKerasSequential
- Full support for Keras sequential API
- Epoch-by-epoch processing, with external metric evaluation (using sed_eval)
- DataProcessor class to embed feature processing and data processing chains
- Feature generators to allow loading feature data from disk during training procedure
- Added possibility to use constant and simple math equations in the keras based learner parameters
Minor updates
- feature_stacker parameter hash added to the data paths. This breaks backward compatibility, previously stored models and results has to be re-calculated or moved to correct paths.
Reproducibility update
- Added Docker container to reproduce baseline results
- Fixed training data ordering (might have been different on different computer setups)
- Added Windows compatibility
First release
- DCASE2017 Challenge baseline system for Tasks 1, 2, and 3
- Development dataset downloading for DCASE2017 Challenge Tasks 1, 2, and 3