Add model card: provenance, files, precision guidance
Browse files
README.md
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: RSNuts/Universal_Source_Separation
|
| 4 |
+
tags:
|
| 5 |
+
- mlx
|
| 6 |
+
- audio
|
| 7 |
+
- audio-to-audio
|
| 8 |
+
- source-separation
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# USS ResUNet30 (MLX)
|
| 12 |
+
|
| 13 |
+
Universal source separation — extracts an arbitrary sound class from a mixture.
|
| 14 |
+
Converted to MLX from the [ByteDance USS](https://github.com/bytedance/uss) code
|
| 15 |
+
and the published
|
| 16 |
+
[RSNuts/Universal_Source_Separation](https://huggingface.co/RSNuts/Universal_Source_Separation)
|
| 17 |
+
checkpoint.
|
| 18 |
+
|
| 19 |
+
## Files
|
| 20 |
+
|
| 21 |
+
| file | size | notes |
|
| 22 |
+
| --- | ---: | --- |
|
| 23 |
+
| `resunet30_fp32.safetensors` | 102 MiB | recommended |
|
| 24 |
+
| `resunet30_fp16.safetensors` | 51 MiB | half the download |
|
| 25 |
+
|
| 26 |
+
## Which precision
|
| 27 |
+
|
| 28 |
+
Measured on an Apple M1 Pro (16 GB), 30 s of audio, three timed runs. Quality is
|
| 29 |
+
fp16 against this model's own fp32 output — the cost of the precision change, not
|
| 30 |
+
a score for the model.
|
| 31 |
+
|
| 32 |
+
| precision | speed | peak memory | quality vs fp32 |
|
| 33 |
+
| --- | ---: | ---: | ---: |
|
| 34 |
+
| fp32 | 27.2x | 1203 MiB | reference |
|
| 35 |
+
| fp16 | 25.8x | 1259 MiB | 57.3 dB |
|
| 36 |
+
|
| 37 |
+
**fp32 is the one to use.** This is the unusual case where the smaller checkpoint
|
| 38 |
+
is both slower and heavier at run time — what dominates the footprint is the
|
| 39 |
+
forward pass, not the weights. Take fp16 only if halving the download matters more
|
| 40 |
+
than that, and note that 57.3 dB is audible on close listening.
|
| 41 |
+
|
| 42 |
+
## Conditioning
|
| 43 |
+
|
| 44 |
+
USS is conditioned on a 527-dimensional vector over the
|
| 45 |
+
[AudioSet](https://research.google.com/audioset/) classes, which selects what the
|
| 46 |
+
separator extracts. It is an input rather than a model weight, and it is not
|
| 47 |
+
shipped here: each vector is a normalised multi-hot indicator over a class group —
|
| 48 |
+
a zero vector with `1/n` at each of the `n` chosen indices — so a class list fully
|
| 49 |
+
determines it.
|
| 50 |
+
|
| 51 |
+
The AudioSet ontology is CC BY-SA 4.0, which is a different licence from this
|
| 52 |
+
model's.
|
| 53 |
+
|
| 54 |
+
## Use
|
| 55 |
+
|
| 56 |
+
Consumed by [AudioToolSwift](https://github.com/starkdmi) as
|
| 57 |
+
`ModelRepository.uss`, which pins this repository to an exact revision and
|
| 58 |
+
verifies the SHA-256 of every file it downloads.
|