augmentations¶

Waveform augmentations for tempo estimation training.

All augmentations operate on (1, T) float32 tensors. The composed TempoAugmenter is the public entry point; use build_augmenter() to construct one from a Hydra config section.

Functions

`build_augmenter`(cfg)	Build a `TempoAugmenter` from a Hydra augmentations config section.
`build_beat_phase_augmenter`(cfg)	Build a `BeatPhaseAugmenter` from a Hydra augmentations config section.

Classes

`AddNoise`([std])	Add white Gaussian noise at a fixed standard deviation.
`AugmentedBeatDataset`(subset, augmenter, ...)	Wraps a `BeatDataset` and applies `BeatPhaseAugmenter` at item-access time.
`AugmentedDataset`(subset, augmenter, ...)	Wraps a dataset and applies `TempoAugmenter` at item-access time.
`BeatPhaseAugmenter`([time_stretch, gain, noise])	Composes waveform+target augmentations for beat-phase training.
`FrameTimeStretch`([min_rate, max_rate])	Randomly stretch or compress audio and resample the frame-level target to match.
`RandomGain`([min_db, max_db])	Scale amplitude by a random gain drawn uniformly from a dB range.
`TempoAugmenter`([time_stretch, gain, noise])	Composes waveform augmentations and returns an updated (wav, tempo) pair.
`TimeStretch`([min_rate, max_rate])	Randomly stretch or compress audio and scale the tempo label accordingly.

class AddNoise(std=0.005)[source]¶

Bases: object

Add white Gaussian noise at a fixed standard deviation.

Parameters:: std (float) – Noise standard deviation (relative to full-scale ±1 audio).

class AugmentedBeatDataset(subset, augmenter, sample_rate, n_samples, n_frames)[source]¶

Bases: Dataset

Wraps a BeatDataset and applies BeatPhaseAugmenter at item-access time.

Use this to augment only the training split while leaving the validation split unchanged — the two are independent dataset instances (see build_beat_dataloaders()), so wrapping one has no effect on the other.

Parameters:

subset – A BeatDataset (or torch.utils.data.Subset of one).
augmenter (BeatPhaseAugmenter) – The augmenter to apply.
sample_rate (int) – Sample rate of the audio (passed to the augmenter).
n_samples (int) – Fixed clip length in samples (passed to the augmenter).
n_frames (int) – Fixed target length in frames (passed to the augmenter).

class AugmentedDataset(subset, augmenter, sample_rate, n_samples)[source]¶

Bases: Dataset

Wraps a dataset and applies TempoAugmenter at item-access time.

Use this to augment only the training split while leaving the validation split unchanged — the two are independent dataset instances (see build_dataloaders()), so wrapping one has no effect on the other.

Parameters:

subset – A TempoDataset (or torch.utils.data.Subset of one).
augmenter (TempoAugmenter) – The augmenter to apply.
sample_rate (int) – Sample rate of the audio (passed to the augmenter).
n_samples (int) – Fixed clip length in samples (passed to the augmenter).

class BeatPhaseAugmenter(time_stretch=None, gain=None, noise=None)[source]¶

Bases: object

Composes waveform+target augmentations for beat-phase training.

Applied in order:

FrameTimeStretch — changes length; re-crops/pads both the waveform and the target back to their expected fixed lengths.
RandomGain
AddNoise

Parameters:

time_stretch (FrameTimeStretch | None)
gain (RandomGain | None)
noise (AddNoise | None)

class FrameTimeStretch(min_rate=0.85, max_rate=1.15)[source]¶

Bases: object

Randomly stretch or compress audio and resample the frame-level target to match.

Same resampling mechanism as TimeStretch (treating the waveform as recorded at sr * rate and resampling back to sr), but instead of rescaling a scalar label, resamples the target’s time axis by the same rate so beat/one/last/mask events stay aligned with the stretched audio’s frames. Frame count and sample count scale identically under a constant hop length, so using one shared rate for both is what keeps them in sync.

Parameters:

min_rate (float) – Minimum speed multiplier (< 1 = slower, lower tempo).
max_rate (float) – Maximum speed multiplier (> 1 = faster, higher tempo).

class RandomGain(min_db=-6.0, max_db=6.0)[source]¶

Bases: object

Scale amplitude by a random gain drawn uniformly from a dB range.

Parameters:

min_db (float) – Lower bound of the gain range in dB (negative = quieter).
max_db (float) – Upper bound of the gain range in dB (positive = louder).

class TempoAugmenter(time_stretch=None, gain=None, noise=None)[source]¶

Bases: object

Composes waveform augmentations and returns an updated (wav, tempo) pair.

Applied in order: 1. TimeStretch — changes length; re-crops/pads back to n_samples. 2. RandomGain 3. AddNoise

Parameters:

time_stretch (TimeStretch | None)
gain (RandomGain | None)
noise (AddNoise | None)

class TimeStretch(min_rate=0.85, max_rate=1.15)[source]¶

Bases: object

Randomly stretch or compress audio and scale the tempo label accordingly.

Implemented via resampling: treating the waveform as if it were recorded at sr * rate Hz and resampling back to sr changes its duration by 1 / rate, which is equivalent to speeding up (rate > 1) or slowing down (rate < 1). Pitch also shifts as a side-effect, which is acceptable for tempo estimation.

Parameters:

min_rate (float) – Minimum speed multiplier (< 1 = slower, lower tempo).
max_rate (float) – Maximum speed multiplier (> 1 = faster, higher tempo).

build_augmenter(cfg)[source]¶

Build a TempoAugmenter from a Hydra augmentations config section.

Returns None if cfg.enabled is false or no individual augmentation is enabled, so callers can skip wrapping altogether.

Parameters:: cfg (DictConfig)
Return type:: TempoAugmenter | None

build_beat_phase_augmenter(cfg)[source]¶

Build a BeatPhaseAugmenter from a Hydra augmentations config section.

Mirrors build_augmenter(), using the frame-target-aware FrameTimeStretch in place of TimeStretch.

Returns None if cfg.enabled is false or no individual augmentation is enabled, so callers can skip wrapping altogether.

Parameters:: cfg (DictConfig)
Return type:: BeatPhaseAugmenter | None