KO
|
EN
gitlite — search
Search
#typescript
#ai-agents
#deepseek-harness
#dsh-plugin
#open-source
#ai
#cli
#dsh
#claude-code
#codex
#developer-tools
#react
PolyDDSP
★ 22
Open GitHub ↗
Polyphonic generalisation of DDSP
Download README (.md)
Explore Similar Repositories
continual-learning-ogdplus
:
Code for "Generalisation Guarantees for Continual Learning with Orthogonal Gradient Descent" (ICML 2020 - Lifelong Learning Workshop)
stitching-is-combinatorial-generalisation
:
[ICLR 2024] Closing the Gap between TD Learning and Supervised Learning - A Generalisation Point of View.
disent-and-gen
:
Code from the article: "The Role of Disentanglement in Generalisation" (ICLR, 2021).
TED
:
Official repository for "Temporal Disentanglement of Representations for Improved Generalisation in Reinforcement Learning".
rl-iter
:
Repository for Iterated Relearning: The Impact of Non-stationarity on Generalisation in Deep Reinforcement Learning
// repository documentation
Was this content helpful?
★ 0
(0 ratings)
Select Rating:
★
★
★
★
★
Submit Feedback
Recent Feedback
×
Download README
Do you want to download the
README.md
file for
PolyDDSP
?
Download (.md)
# PolyDDSP A PyTorch reimplementation of PolyDDSP — a polyphonic extension of [DDSP](https://arxiv.org/abs/2001.04643) that resynthesises polyphonic audio through a bank of differentiable harmonic + filtered-noise voices. Paper: Baker, T., Climent, R., & Chen, K. *PolyDDSP: A Lightweight and Polyphonic Differentiable Digital Signal Processing Library.* Music in the AI Era (CMMR 2023), Springer LNCS 13770 — https://zenodo.org/records/10113134 ## Architecture - Pitch extraction is a frozen Basic Pitch, using the full pre-mean per-frame note amplitudes and pitch bend values, with loudest-voice allocation. - All voices share a single set of decoder weights and are batch processed, so voice count is a runtime choice. - Each voice takes its loudness from one extracted loudness envelope plus its own framed velocity from Basic Pitch. Timbre encoding is global. - Audio can be generated at any sample rate by changing the synthesis resolution. The experiments below ran at 16 kHz to match the original DDSP. Key modules: `polyddsp/model/polyddsp.py` (top level), `decoder.py`, `additive.py`, `noise.py`, `reverb.py`, `z.py`, `loudness.py`, `note_extraction.py`, `voice_allocation.py`. ## Install ```bash uv sync export POLYDDSP_DATA_DIR=/path/to/data # datasets, default /data/polyddsp export POLYDDSP_OUT_DIR=/path/to/runs # run outputs, default outputs/ ``` ## Data | Dataset | Download | Path under `$POLYDDSP_DATA_DIR` | |---|---|---| | GuitarSet | [zenodo 3371780](https://zenodo.org/records/3371780) (`audio_mono-pickup_mix.zip`) | `guitarset/**/*_mix.wav` | | MAESTRO | [magenta.tensorflow.org](https://magenta.tensorflow.org/datasets/maestro) | `maestro/bach_2018/**/*.wav` | MAESTRO is restricted to `bach_2018` to keep the piano and recording environment consistent. Files are split 80/20 train/val from `run.seed`. ## Usage ```bash # 1. Transcribe the dataset to a pitch cache. Required before training. python -m polyddsp.preprocess --n-voices 6 \ --glob '**/*_mix.wav' --root $POLYDDSP_DATA_DIR/guitarset # 2. Train python -m polyddsp.train experiment=guitarset # 3. Evaluate a checkpoint python -m polyddsp.eval experiment=guitarset ckpt=outputs/<run>/best.pt # 4. Render audio from a checkpoint python -m polyddsp.infer ckpt=outputs/<run>/best.pt input=clip.wav out=resynth.wav ``` - `--n-voices` must match the config's `model.n_voices`. `--glob` defaults to `**/*.mp3`, so pass it explicitly; for MAESTRO use `--glob '**/*.wav' --root $POLYDDSP_DATA_DIR/maestro/bach_2018`. - The cache is a per-file `.bp_v{V}_sr16000_hop64.f0.pt` sidecar, and is skipped if newer than its audio. - `infer` transcribes in-process, so it needs no cache and takes any audio file. - Config is Hydra (`configs/config.yaml` + `configs/experiment/*.yaml`); any key can be overridden on the CLI, e.g. `train.batch_size=8`. Checkpoints, the resolved config, `log.jsonl` and metric summaries go to `$POLYDDSP_OUT_DIR/<run.name>/`. Pass `wandb.mode=disabled` to run offline. ## Results Held-out val split, 100k steps at batch size 12 on 4-second crops. All metrics lower-is-better; machine-readable copy in `results/final_metrics.json`. | | GuitarSet | MAESTRO | |---|---|---| | `n_voices` | 6 | 6 | | `use_z` | true | false | | `use_reverb` | false | true | | `use_noise` | true | false | | MSS | 0.642 | 0.423 | | CLAP-FD | 0.0495 | 0.0523 | | FAD | 0.438 | 1.235 | | multipitch MSE | 0.000956 | 0.00117 | | MFCC L1 | 1.607 | 1.675 | | loudness L1 | 1.780 | 2.408 | ```bash python -m polyddsp.train experiment=guitarset python -m polyddsp.train experiment=maestro experiment.model.n_voices=6 ``` MAESTRO ships `n_voices: 10`; the run above used 6, which fits in 24 GB and was never exceeded by simultaneous notes in this subset. ## Performance Faster than realtime on CPU, scaling with voice count: roughly 15x at V=1, 3.4x at V=6 and 2x at V=10. ## Known limitations - Basic Pitch note timing drifts on long files, by about 0.3 s per minute of audio. See `tests/test_bp_timing_drift.py`. - The model is not streamable, purely because there is no streamable polyphonic pitch detector to put in front of it. Everything downstream generalises to streaming. ## Tests ```bash uv run pytest ``` The CLAP and FAD metric tests download model checkpoints on first run. ## License Code is MIT (see `LICENSE`). The paper is CC BY 4.0.
