A flexible frequency-band splitter for music source separation, organized around a single separator family that can express BS-style, mel-style, and custom layouts.
Instead of treating BS-style and mel-band layouts as separate Python entry points, this package treats them as different band-layout configurations of one core design centered on BandSplitRotator.
The codebase is implemented in PyTorch, fully typed (py.typed), and designed for modular reuse so options such as PoPE, custom filter banks, and optional SageAttention acceleration can live on one aligned constructor surface.
Quick fix: size mismatch when loading a checkpoint
If loading an older BS-style checkpoint or configuration raises a size-mismatch error, check mask_estimator_depth in the configuration.
Some earlier configuration files effectively used mask_estimator_depth=1 even when stored as 2 because a later subtraction was applied. This package removes that subtraction, so the direct equivalent is:
- set
mask_estimator_depth=1
Updating that value resolves the most common mismatch quickly.
Why this architecture helps in practice
- Forward-looking architecture: A single model family makes it easier to adopt new ideas, such as PoPE, HyperACE, or custom band-split definitions, while keeping interfaces aligned with established ecosystems.
- Universal configuration surface:
BandSplitRotator exposes downstream attention, transformer, mask-estimator, optional HyperACE segmentation-branch, STFT, and loss settings on one constructor.
- Rich tooling and ecosystem: The package provides strong typing (
py.typed), modular APIs, and rich docstrings focusing on usage, literature citations, and migration paths.
Easy to migrate
Transitioning from older codebases is straightforward because most downstream identifiers stay the same and the data flow is highly similar.
If you’re changing from an existing codebase, replace the older import path with BandSplitRotator, then choose the desired band layout with freqs_per_bands, sample_rate together with num_bands, or mask_filter_bank.
What is unified here
The key design idea is that the difference between the BS-style front end and the mel-band front end
is treated as a band-layout problem, not as a reason to maintain two unrelated model families.
hunterFormsBS.bandSplitRotator.BandSplitRotator is the primary unified entry point and the only public separator class exported by this package.hunterFormsBS.bands.BandSplit, hunterFormsBS.mask.MaskEstimator, hunterFormsBS.mask.MLP, hunterFormsBS.loss.lossComputation, hunterFormsBS.attend.Attention, and hunterFormsBS.transform.Transformer hold the reusable typed building blocks shared across all band-layout modes.hunterFormsBS.hyperACE.SegmModel supplies the optional segmentation-style mask-estimation branch used when use_hyperACE=True.- Attention and transformer options such as
attn_dropout, ff_dropout, flash_attn, sage_attention, and scale keep the same identifiers as they move from model constructors into downstream blocks.
- All user-configurable separator settings live on
BandSplitRotator, including optional HyperACE branch settings under the segm_* prefix.
At the band level, the model only needs a band-membership map, called mask_filter_bank in the
codebase. You can think of that map as a Boolean matrix
$$
F \in {0, 1}^{B \times N_f}
$$
where $B$ is the number of bands and $N_f$ is the number of STFT frequency bins.
- In a non-overlapping BS-style layout, each frequency bin belongs to exactly one band, so
$$
\forall f,; \sum_b F_{b,f} = 1.
$$
- In an overlapping mel-style layout, some frequency bins belong to more than one band, so
$$
\exists f \text{ such that } \sum_b F_{b,f} > 1.
$$
When bands overlap, the reconstructed mask for a frequency bin is averaged across the contributing
bands:
$$
\hat{M}{f,t} = \frac{1}{S_f} \sum{b : F_{b,f} = 1} \hat{M}^{(b)}{f,t},
\qquad
S_f = \sum_b F{b,f}.
$$
That is why this package makes it easy to move between overlapping and non-overlapping bands, and to
change how bands are distributed across the frequency axis. The architectural difference lives in the
filter bank, not in two separate theories of the model.
Which band-layout mode you should use
| Use this configuration |
When |
Key parameters |
BandSplitRotator with the default non-overlapping layout |
You want a BS-style frequency partition or a close comparison with non-overlapping upstream layouts. |
freqs_per_bands, optional num_bands |
BandSplitRotator with automatic mel-band construction |
You want a mel-band layout computed at runtime. |
sample_rate, num_bands, melscale_fbanks_mel_scale, melscale_fbanks_norm |
BandSplitRotator with explicit custom bands |
You want a checkpoint-specific or research-specific layout. |
mask_filter_bank |
Attention and optional acceleration
flash_attn=True requests PyTorch scaled-dot-product-attention backends when the active device
supports that path. sage_attention=True asks downstream Attend blocks to call
sageattention.sageattn; install SageAttention manually
before enabling it because hunterFormsBS does not install that package.
BandSplitRotator can choose RoPE or PoPE with use_pope, so the same model family can switch
positional encoders without changing entry points.
Automatic and custom mask_filter_bank construction
Most users never need this section. When mask_filter_bank is None and both sample_rate and
num_bands are provided, BandSplitRotator builds mel-band layouts at runtime with
torchaudio.functional.melscale_fbanks. The constructor parameters
melscale_fbanks_mel_scale and melscale_fbanks_norm are forwarded to torchaudio, so you can
choose the mel-scale formula and normalization rule used for automatic mel-band construction.
For the package default mel-band layout, keep sample_rate=44100, stft_n_fft=2048,
num_bands=60, melscale_fbanks_mel_scale='slaney', and melscale_fbanks_norm='slaney'.
When sample_rate and num_bands are not both provided, BandSplitRotator derives the
non-overlapping BS-style Boolean band-membership map from freqs_per_bands.
If a checkpoint uses a different band layout, pass mask_filter_bank explicitly as a Boolean
torch.Tensor with shape (band, freq).
All optional segmentation-branch settings also live on BandSplitRotator under the segm_*
prefix.
Package map
hunterFormsBS.__init__
- Direct export:
BandSplitRotator
- Purpose: small top-level namespace for the primary separator model.
hunterFormsBS.bandSplitRotator
- Main symbols:
BandSplitRotator
- Purpose: unified separator that can build BS-style, mel-style, or custom band layouts from one
model family, with downstream attention, transformer, STFT, mask-estimator, optional
segmentation-branch, and loss options on the constructor.
hunterFormsBS.bands
- Main symbols:
BandSplit, DEFAULT_FREQS_PER_BANDS
- Purpose: band projection layer and the BS-style default frequency-bin partition.
hunterFormsBS.loss
- Main symbols:
lossComputation
- Purpose: multi-resolution STFT training-loss helper.
hunterFormsBS.mask
- Main symbols:
MaskEstimator, MLP
- Purpose: mask-estimation heads and band-local affine blocks.
hunterFormsBS.attend
- Main symbols:
Attend, Attention
- Purpose: shared attention core with RoPE / PoPE, PyTorch SDPA, and optional SageAttention support.
hunterFormsBS.transform
- Main symbols:
FeedForward, Transformer
- Purpose: position-wise feedforward block and stacked attention-and-feedforward transformer.
hunterFormsBS.hyperACE
- Main symbols:
SegmModel, HyperACE, Backbone, Decoder, ProgressiveUpsampleHead
- Purpose: optional segmentation-style mask-estimation branch and its reusable building blocks.
hunterFormsBS.theTypes
- Main symbols:
ParametersComputeLoss, FlashAttentionConfig, ParametersAttention, ParametersMaskEstimator, ParametersSTFT, ParametersTransformer
- Purpose: typed configuration records used across the package.
Architecture in one sentence
The stable separator path is
raw audio → STFT → band gathering → BandSplit → hierarchical attention → MaskEstimator → mask
followed by overlap-aware mask averaging when needed, complex masking in the STFT domain, and inverse
STFT reconstruction back to waveform audio.
When use_hyperACE=True, each MaskEstimator head also adds the optional segmentation-style branch
before the final mask accumulation step.
Top-level exports
The top-level package namespace currently re-exports the primary model that new users most often
need:
Supporting modules stay under explicit submodule imports so the main namespace remains small and
comparisons with upstream repos stay easy to read.
Reference materials
Gaussian Error Linear Units (GELUs)
Language Modeling with Gated Convolutional Networks
Attention Is All You Need
Root Mean Square Layer Normalization
FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration
SageAttention2: Efficient Attention with Thorough Outlier Smoothing and Per-thread INT4 Quantization
SageAttention2++: A More Efficient Implementation of SageAttention2
Einops: Clear and Reliable Tensor Manipulations with Einstein-like Notation
Sound Demixing Challenge 2023 Music Demixing Track Technical Report: TFC-TDF-UNet v3
Music Source Separation Based on a Lightweight Deep Learning Framework (DTTNET: DUAL-PATH TFC-TDF UNET)
YOLOv13: Real-Time Object Detection with Hypergraph-Enhanced Adaptive Visual Perception
Decoupling the “What” and “Where” With Polar Coordinate Positional Embeddings
Packages and documentation
My recovery
By law, authors retain their copyrights unless they grant exceptions. The Attribution-NonCommercial 4.0 International Deed outlines how you may use this work.
The restriction on commercial use (without a license) limits advantage-taking, unjust enrichment, and exploitation. You would be morally wrong if you were to enrich yourself by using this work while I am denied healthcare, housing, security of property, and the safety of my person.
To obtain and retain the right to use this work, you incur the cost of attributing the work to me. Your nominal cost might accrue as a benefit that mitigates my suffering or saves my life.
![dependabot[bot]](https://avatars.githubusercontent.com/in/29110?v=4)
评论