Wired SCNet into global power settings.

This commit is contained in:
seclusionenthusiast committed 2026-03-03 22:40:28 +09:00
1 parent a877061aab
commit c5f8c62a90
2 files changed
+34 -10

No files matched your search

+16 -3
View File
@@ -259,10 +259,23 @@ class SCNetSeparator:
# Public API
# ------------------------------------------------------------------
def separate_into_parts(self, input_path: str, output_dir: str) -> Dict[str, str]:
def separate_into_parts(
self,
input_path: str,
output_dir: str,
chunk_size_override: Optional[int] = None,
num_overlap_override: Optional[int] = None,
) -> Dict[str, str]:
"""
Separate audio into stems.
Args:
chunk_size_override: Override the model-config chunk size (samples).
Smaller values reduce VRAM; larger improve context.
num_overlap_override: Override the model-config num_overlap.
Higher values improve boundary quality at the cost
of more computation.
Returns:
Dict mapping stem name → output file path.
"""
@@ -271,8 +284,8 @@ class SCNetSeparator:
sample_rate: int = cfg.get("audio", {}).get("sample_rate", 44100)
inf_cfg: dict = cfg.get("inference", {})
chunk_size: int = inf_cfg.get("chunk_size", 485100)
num_overlap: int = inf_cfg.get("num_overlap", 2)
chunk_size: int = chunk_size_override if chunk_size_override is not None else inf_cfg.get("chunk_size", 485100)
num_overlap: int = num_overlap_override if num_overlap_override is not None else inf_cfg.get("num_overlap", 2)
# Stem order from training config; fall back to model attribute
instruments = cfg.get("training", {}).get("instruments") or list(model.sources)
+18 -7
View File
@@ -115,18 +115,21 @@ _SEP_POWER_PRESETS: Dict[str, Dict] = {
'demucs': {'segment': 5, 'overlap': 0.1},
'sepacap': {'chunk_samples': 24_000}, # 1 s @ 24 kHz
'sepreformer': {'chunk_samples': 24_000}, # 3 s @ 8 kHz
'scnet': {'chunk_size': 132_300, 'num_overlap': 2}, # ~3 s @ 44.1 kHz
},
'standard': {
'roformer': {'batch_size': None, 'num_overlap': None},
'demucs': {'segment': None, 'overlap': None},
'sepacap': {'chunk_samples': 48_000}, # 2 s @ 24 kHz
'sepreformer': {'chunk_samples': 48_000}, # 6 s @ 8 kHz
'scnet': {'chunk_size': None, 'num_overlap': None}, # model config defaults
},
'performance': {
'roformer': {'batch_size': 8, 'num_overlap': 4},
'demucs': {'segment': 20, 'overlap': 0.5},
'sepacap': {'chunk_samples': 96_000}, # 4 s @ 24 kHz
'sepreformer': {'chunk_samples': 96_000}, # 12 s @ 8 kHz
'scnet': {'chunk_size': None, 'num_overlap': 4}, # larger overlap for quality
},
}
@@ -396,6 +399,8 @@ def run_separation(
job_id, input_path, model, normalised_format, stems_dir,
mp3_bitrate=mp3_bitrate, metadata=metadata,
stem_prefix=proj_name,
chunk_size_override=_power['scnet'].get('chunk_size'),
num_overlap_override=_power['scnet'].get('num_overlap'),
)
elif model in BSR_SW_MODEL_IDS:
stem_files = _run_bsroformer_sw(
@@ -1381,32 +1386,38 @@ def _run_scnet(
mp3_bitrate: str = '256k',
metadata: Optional[Dict] = None,
stem_prefix: Optional[str] = None,
chunk_size_override: Optional[int] = None,
num_overlap_override: Optional[int] = None,
) -> Dict[str, str]:
"""Run SCNet XL separation (state-of-the-art 2024-2025)."""
from .scnet_separator import SCNetSeparator
jm.update_job(
job_id, status=jm.JobStatus.RUNNING,
progress=f"Loading SCNet model ({model})...",
)
if jm.should_cancel(job_id):
raise RuntimeError("Cancelled by user")
sep = SCNetSeparator(
model_id=model,
cpu_separation=CPU_SEPARATION,
output_format=output_format,
mp3_bitrate=mp3_bitrate,
)
jm.update_job(job_id, progress="Separating with SCNet... (this may take a while)")
temp_out = JOBS_DIR / job_id / "scnet_out"
temp_out.mkdir(parents=True, exist_ok=True)
try:
stem_paths = sep.separate_into_parts(input_path, str(temp_out))
stem_paths = sep.separate_into_parts(
input_path, str(temp_out),
chunk_size_override=chunk_size_override,
num_overlap_override=num_overlap_override,
)
if jm.should_cancel(job_id):
raise RuntimeError("Cancelled by user")