Skip to content

Inference

inference

High level orchestrator for model inference

Functions:

Name Description
run_inference_on_file

Runs the full source separation pipeline on a single audio file.
separate

Chunk, predict and stitch.

run_inference_on_file 

run_inference_on_file(
    mixture: Audio[RawAudioTensor],
    config: Config,
    model: Module,
) -> dict[StemName, RawAudioTensor]

Runs the full source separation pipeline on a single audio file.

Source code in src/splifft/inference.py 
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
def run_inference_on_file(
    mixture: Audio[RawAudioTensor], config: Config, model: nn.Module
) -> dict[StemName, RawAudioTensor]:
    """Runs the full source separation pipeline on a single audio file."""

    mixture_data: RawAudioTensor | NormalizedAudioTensor = mixture.data
    mixture_stats: NormalizationStats | None = None
    if config.inference.normalize_input_audio:
        norm_audio = normalize_audio(mixture)
        mixture_data = norm_audio.audio.data
        mixture_stats = norm_audio.stats

    separated_data = separate(
        mixture_data=mixture_data,
        chunk_cfg=config.chunking,
        model=model,
        batch_size=config.inference.batch_size,
        num_model_stems=len(config.model.output_stem_names),
        chunk_size=config.model.chunk_size,
        use_autocast_dtype=config.inference.use_autocast_dtype,
    )

    denormalized_stems: dict[ModelOutputStemName, RawAudioTensor] = {}
    for i, stem_name in enumerate(config.model.output_stem_names):
        stem_data = separated_data[i, ...]
        if mixture_stats is not None:
            stem_data = denormalize_audio(
                audio_data=NormalizedAudioTensor(stem_data),
                stats=mixture_stats,
            )
            denormalized_stems[stem_name] = stem_data
        else:
            denormalized_stems[stem_name] = RawAudioTensor(stem_data)

    if config.inference.apply_tta:
        raise NotImplementedError

    output_stems = denormalized_stems
    if config.derived_stems:
        output_stems = derive_stems(
            denormalized_stems,
            mixture.data,
            config.derived_stems,
        )

    return output_stems

separate 

separate(
    mixture_data: RawAudioTensor | NormalizedAudioTensor,
    chunk_cfg: ChunkingConfig,
    model: Module,
    batch_size: BatchSize,
    num_model_stems: NumModelStems,
    chunk_size: ChunkSize,
    *,
    use_autocast_dtype: Dtype | None = None,
) -> Tensor

Chunk, predict and stitch.

Source code in src/splifft/inference.py 
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
def separate(
    mixture_data: RawAudioTensor | NormalizedAudioTensor,
    chunk_cfg: ChunkingConfig,
    model: nn.Module,
    batch_size: BatchSize,
    num_model_stems: NumModelStems,
    chunk_size: ChunkSize,
    *,
    use_autocast_dtype: Dtype | None = None,
) -> Tensor:  # FIXME: update type hint.
    """Chunk, predict and stitch."""
    device = mixture_data.device
    original_num_samples = mixture_data.shape[-1]

    hop_size = int(chunk_size * (1 - chunk_cfg.overlap_ratio))

    if chunk_cfg.window_shape == "hann":
        window = torch.hann_window(chunk_size, device=device)
    else:
        raise NotImplementedError(f"{chunk_cfg.window_shape=}")

    chunk_generator = generate_chunks(
        audio_data=mixture_data,
        chunk_size=chunk_size,
        hop_size=hop_size,
        batch_size=batch_size,
        padding_mode=chunk_cfg.padding_mode,
    )
    if tqdm is not None:
        chunk_generator = tqdm(
            chunk_generator,
            desc="Processing chunks",
        )

    processed_chunks = []
    with (
        torch.inference_mode(),
        torch.autocast(
            device_type=device.type,
            enabled=use_autocast_dtype is not None,
            dtype=(
                get_dtype(use_autocast_dtype) if use_autocast_dtype is not None else torch.float32
            ),
        ),
    ):
        for chunk_batch in chunk_generator:
            separated_batch = model(chunk_batch)
            processed_chunks.append(separated_batch)

    return stitch_chunks(
        processed_chunks=processed_chunks,
        num_stems=num_model_stems,
        chunk_size=chunk_size,
        hop_size=hop_size,
        target_num_samples=original_num_samples,
        window=WindowTensor(window),
    )