PyUp Safety actively tracks 282,845 Python packages for vulnerabilities and notifies you when to upgrade.
Major (breaking) change * New base classes * `ProcessorGroup` and `LossGroup` inherit from `DAGLayer` (dags.py) * All ddsp.training layers [decoders, encoders, preprocessors] inherit from `DictLayer` (nn.py) * Renamed classes to more precise terms * `Additive` -> `Harmonic` * `DefaultPreprocessor` -> `F0LoudnessPreprocessor` * `TranscribingAutoencoder` -> `Inverse Synthesis` * New experimental `MidiAutoencoder` model (WIP) * `Evaluator` classes in eval_util (now configurable from gin instead of a big long series of if statements) * Minor bug fixes
* Cloud training scripts * Model API refactor (no more `model.get_controls()`, `model()` now returns a dictionary of output tensors instead of audio. Audio can be retrieved with `model.get_audio_from_outputs(outputs)` * Separate files for each model * Minor bug fixes
Release for reproducing the results from the 2020 ICML SAS workshop paper (https://openreview.net/forum?id=RlVTYWhsky7). WIP code from the paper added with EXPERIMENTAL disclaimers. Gin configs and details provided in `ddsp/training/gin/papers/icml2020` v.0.10.0 * Custom cumsum operation to avoid phase accumulation errors for generating long sequences. * Script to automatically update old gin configs.
Add custom cumsum function that doesn't accumulate phase errors like tf.cumsum.
* Updated pitch detection metrics (RPA, RCA) * Sinusoidal Synthesizer * Warm starting models (model_dir -> save_dir, restore_dir)
Small fixes to bugs introduced by refactor :).
Some bug fixes and a refactor of train_util and eval_util.
* New data normalization in the demo colab notebooks. * Tiny model config. * Most (but not all) of the variable sample rate PRs. * Tests and bug fixes.
Simplify and refactor RnnFcDecoder. * Requires old models to add a single line to their operative gin configs, or --gin_param, `RnnFcDecoder.input_keys = ('f0_scaled', 'ld_scaled')`
* Models now use self._loss_dict to keep track of losses, and not the built-in keras self.losses (so that we can keep track of each loss name without needing a synced parallel list).
* Allow memory growth flag for GPUs with less memory. * Use latest CREPE * Remove custom TPU cumsum function * Bug fixes to colab * Compare f0 predictions with f0 ground truth * Creating datasets with different sample rates
Update code to use tensorflow 2 and python 3.
Code used in the initial ICLR 2020 paper (https://openreview.net/forum?id=B1x1ma4tDr). `ddsp/` works for tf1 and tf2, while `ddsp/training/` is written with the tf1 Estimator API.