absolution-omega / app /mini_model.py
2Btainment's picture
v0.6: Arrangement Studio, Demucs htdemucs (CPU), bucket persistence
607c4f0 verified
Raw History Blame Contribute Delete
13.3 kB
"""OMEGA-TAG Mini v0.5: tiny, OFFLINE multilabel spectral-event model.
Model shipped with project uses *synthetic* two-stage training: isolated/pair
pretraining, then overlapping-mixture fine-tuning. It has NEVER been trained
on real labelled music, so it must not claim reliable instrument identities.
All scores are uncalibrated classifier activations, not probabilities.
No remote downloads or unsafe pickle deserialization.
"""
from __future__ import annotations
from pathlib import Path
import hashlib
import json
import math
import numpy as np
import soundfile as sf
from scipy.signal import resample_poly
LABELS = ("bass_sustain", "kick_like", "snare_like", "hi_hat_like",
"harmonic_mid", "noise_texture")
FRIENDLY = {
"bass_sustain": "Tieffrequente, anhaltende Töne / Bass-Verdacht",
"kick_like": "Tieffrequente Impulse / Kick-Verdacht",
"snare_like": "Breitbandige Mitten-Impulse / Snare-Verdacht",
"hi_hat_like": "Helle, kurze Impulse / Hi-Hat-Verdacht",
"harmonic_mid": "Harmonische Mitten / Melodie-Verdacht",
"noise_texture": "Rauschähnliche Textur (kann musikalisch sein)",
}
SR = 8000
N = 4096
MODEL_FILE = Path(__file__).resolve().parent / "models" / "omega_tag_mini_v1.npz"
BANDS = ((25,80),(80,180),(180,400),(400,900),(900,1900),(1900,3500),(3500,3950))
def features(x: np.ndarray) -> np.ndarray:
"""32 stable feature components from 0.512s mono resampled audio.
Features are descriptive and deliberately do not identify instruments
through semantic embedding / pretrained audio foundation features.
"""
x = np.asarray(x, dtype=np.float64).reshape(-1)
if len(x) != N:
raise ValueError(f"Expected {N} samples, got {len(x)}")
if not np.isfinite(x).all():
raise ValueError("Non-finite features input")
# Fixed frames. Use per-frame relative energies for spectral activity.
windows = np.stack([x[i:i+1024] for i in range(0, N-1024+1, 384)], axis=0)
fft = np.abs(np.fft.rfft(windows*np.hanning(1024), axis=-1))**2+1e-15
hz = np.fft.rfftfreq(1024, 1./SR)
bands=np.stack([fft[:,(hz>=lo)&(hz<hi)].sum(axis=1) for lo,hi in BANDS],axis=1)
frame_energy = fft.sum(axis=1,keepdims=True)
proportion=bands/frame_energy
avg=proportion.mean(axis=0)
variability=proportion.std(axis=0)
logenergy=np.log10(np.maximum(bands.mean(axis=0),1e-10))
logtotal=np.log10(max(float(np.mean(x*x)),1e-10))
flatness=(np.exp(np.mean(np.log(fft),axis=1))/np.mean(fft,axis=1)).mean()
zcr=np.mean(np.signbit(x[1:])!=np.signbit(x[:-1]))
power=np.mean(windows*windows,axis=1)
onset=np.maximum(np.diff(power),0)
novelty=float(np.max(onset)/(np.mean(power)+1e-12))
fluctuation=float(np.std(power)/(np.mean(power)+1e-12))
crest=float(np.max(np.abs(x))/(math.sqrt(float(np.mean(x*x)))+1e-10))
spectral_flux=np.maximum(np.diff(np.sqrt(fft),axis=0),0).sum()/np.sqrt(fft).sum()
# 7+7+7+7 = 28, then 7 metadata => 35
result=np.r_[avg,variability,np.clip(logenergy,-10,7),
np.clip(logenergy-logenergy.mean(),-12,12),
logtotal, flatness,zcr,np.log1p(novelty),
np.log1p(fluctuation), np.log1p(crest),spectral_flux]
return np.asarray(result,dtype=np.float32)
FEATURE_COUNT = len(features(np.zeros(N,dtype=np.float32)))
def synthetic_signal(label: str, rng: np.random.Generator) -> np.ndarray:
"""Generated test/training signals, NOT copyrighted music nor real stems."""
t=np.arange(N)/SR
x=np.zeros(N,dtype=np.float64)
if label=="bass_sustain":
freq=rng.uniform(46,160)
for harmonic in (1,2,3):
x+=np.sin(2*np.pi*(harmonic*freq*t+rng.uniform(0,1)))*(.6/harmonic**1.3)
x*=.4+.6*np.sin(2*np.pi*rng.uniform(.3,2)*t+rng.uniform(0,6))**2
elif label=="kick_like":
for loc in rng.choice(np.arange(0,N-850,300),size=int(rng.integers(1,4)),replace=False):
k=np.arange(min(1500,N-loc)); time=k/SR
phase=2*np.pi*(50*time+90*(1-np.exp(-time*32))/32)
env=np.exp(-time*rng.uniform(17,35))
x[loc:loc+len(k)] += np.sin(phase)*env
elif label=="snare_like":
for loc in rng.choice(np.arange(0,N-450,280),size=int(rng.integers(1,4)),replace=False):
k=np.arange(min(850,N-loc)); time=k/SR
pulse=rng.standard_normal(len(k))
pulse=np.r_[0,np.diff(pulse)]
x[loc:loc+len(k)] += pulse*np.exp(-time*27)*.3
elif label=="hi_hat_like":
for loc in rng.choice(np.arange(0,N-150,170),size=int(rng.integers(2,7)),replace=False):
k=np.arange(min(380,N-loc));time=k/SR
noise=rng.standard_normal(len(k))
noise=np.r_[0,np.diff(noise)]
x[loc:loc+len(k)]+=noise*np.exp(-time*85)*.35
elif label=="harmonic_mid":
for freq in rng.uniform(200,900,size=int(rng.integers(1,4))):
x+=np.sin(2*np.pi*(freq*t+rng.uniform(0,1)))*.4
x*=.3+.7*(np.sin(np.pi*np.clip(t/rng.uniform(.25,.85),0,1))**2)
elif label=="noise_texture":
noise=rng.standard_normal(N)
if rng.random()<.6:
noise=np.convolve(noise,np.ones(37)/37,mode='same')
x += noise*.3
return x.astype(np.float32)
def synthetic_case(rng:np.random.Generator, phase: str) -> tuple[np.ndarray,np.ndarray]:
max_sources=2 if phase=="pretrain" else 5
count=int(rng.integers(1,max_sources+1))
indices=rng.choice(len(LABELS),size=count,replace=False)
y=np.zeros(len(LABELS),dtype=np.float32)
x=np.zeros(N,dtype=np.float32)
for idx in indices:
y[idx]=1.
x+=synthetic_signal(LABELS[idx],rng)*rng.uniform(.2,1.0)
# Apply nonstationary gain, external hum/noise and polarity changes to test
# robustness; all remain synthetic.
if rng.random()<.4:x*=np.linspace(rng.uniform(.35,1),rng.uniform(.35,1),N)
if rng.random()<.5:x+=rng.standard_normal(N).astype(np.float32)*rng.uniform(0,.016)
if rng.random()<.5:x*=-1
peak=max(float(np.max(np.abs(x))),1e-8)
x*=float(rng.uniform(.15,.8))/peak
return x,y
def generate_training(n:int,seed:int,phase:str):
rng=np.random.default_rng(seed)
X=np.empty((n,FEATURE_COUNT),np.float32)
Y=np.empty((n,len(LABELS)),np.float32)
for i in range(n):
x,y=synthetic_case(rng,phase)
X[i]=features(x);Y[i]=y
return X,Y
def _sigmoid(z): return 1/(1+np.exp(-np.clip(z,-35,35)))
def _net(X,W1,b1,W2,b2):
h=np.tanh(X@W1+b1)
return h,_sigmoid(h@W2+b2)
def train_model(destination:Path=MODEL_FILE,seed:int=25,epochs1:int=90,epochs2:int=105,
pretrain_count:int=650,finetune_count:int=1100) -> dict:
"""Offline reproducible 2-stage neural training, using only synthetic audio.
Stage 2 is real parameter fine-tuning of stage 1 weights but NOT transfer
learning from a commercial foundation model or real instrument recordings.
"""
rng=np.random.default_rng(seed)
xp,yp=generate_training(pretrain_count,seed,"pretrain")
xf,yf=generate_training(finetune_count,seed+1,"finetune")
xt,yt=generate_training(300,seed+2,"finetune")
mu=xp.mean(axis=0)
sigma=np.maximum(xp.std(axis=0),.08)
xp=np.clip((xp-mu)/sigma,-8,8)
xf=np.clip((xf-mu)/sigma,-8,8)
xt=np.clip((xt-mu)/sigma,-8,8)
hidden=32
W1=(rng.standard_normal((FEATURE_COUNT,hidden))*.15).astype(np.float64)
b1=np.zeros(hidden)
W2=(rng.standard_normal((hidden,len(LABELS)))*.15).astype(np.float64)
b2=np.zeros(len(LABELS))
params=[W1,b1,W2,b2]
mt=[np.zeros_like(p) for p in params]
vt=[np.zeros_like(p) for p in params]
step=0
def loop(X,Y,epochs,lr):
nonlocal step
for epoch in range(epochs):
for batchidx in np.array_split(rng.permutation(len(X)),max(1,len(X)//96)):
data,labels=X[batchidx],Y[batchidx]
h,prob=_net(data,*params)
grad=(prob-labels)/len(batchidx)
g2=h.T@grad
gb2=grad.sum(axis=0)
gh=(grad@W2.T)*(1-h*h)
g1=data.T@gh
gb1=gh.sum(axis=0)
for idx, g in enumerate((g1,gb1,g2,gb2)):
step+=1 if idx==0 else 0
mt[idx][:]=.9*mt[idx]+.1*g
vt[idx][:]=.999*vt[idx]+.001*g*g
mhat=mt[idx]/(1-.9**step)
vhat=vt[idx]/(1-.999**step)
params[idx]-=lr*mhat/(np.sqrt(vhat)+1e-8)
loop(xp,yp,epochs1,.012)
loop(xf,yf,epochs2,.006)
h,pred=_net(xt,*params)
estimates=pred>=.5
precision=float(np.sum((estimates==1)&(yt==1))/max(1,np.sum(estimates==1)))
recall=float(np.sum((estimates==1)&(yt==1))/max(1,np.sum(yt==1)))
acc=float(np.mean(estimates==yt))
destination.parent.mkdir(parents=True,exist_ok=True)
np.savez_compressed(destination,
W1=W1.astype(np.float32),b1=b1.astype(np.float32),
W2=W2.astype(np.float32),b2=b2.astype(np.float32),
mu=mu.astype(np.float32),sigma=sigma.astype(np.float32),
model_schema=np.array("omega_mini_synthetic_v1"),
labels=np.asarray(LABELS))
return {"stage1":"synthetic isolated/pair pretraining",
"stage2":"fine-tuning on synthetic overlapping mixtures",
"heldout_synthetic_precision":round(precision,3),
"heldout_synthetic_recall":round(recall,3),
"heldout_synthetic_label_accuracy":round(acc,3),
"validation_dataset":"300 held-out synthetic mixtures, NOT real-world benchmark",
"checkpoint":str(destination),
"checkpoint_sha256":hashlib.sha256(destination.read_bytes()).hexdigest(),
"architecture":f"feature({FEATURE_COUNT}) -> tanh(32) -> sigmoid({len(LABELS)})",
"train_examples":pretrain_count+finetune_count}
def load_model(path:Path=MODEL_FILE)->dict:
if not path.is_file():
raise FileNotFoundError("Mini model missing. Run python -m app.mini_model --train")
with np.load(path,allow_pickle=False) as f:
params={k:f[k].copy() for k in ("W1","b1","W2","b2","mu","sigma")}
schema=str(f["model_schema"].item())
labels=tuple(str(x) for x in f["labels"])
if schema!="omega_mini_synthetic_v1" or labels!=LABELS:
raise ValueError("Unsupported/incompatible mini model weights")
if params["W1"].shape!=(FEATURE_COUNT,32) or params["W2"].shape!=(32,len(LABELS)):
raise ValueError("Corrupt mini model dimensions")
if not all(np.isfinite(v).all() for v in params.values()):
raise ValueError("Non-finite model weights")
return params
def predict_features(x:np.ndarray,model:dict)->np.ndarray:
scaled=np.clip((x-model["mu"])/model["sigma"],-8,8)
return _net(scaled,model["W1"],model["b1"],model["W2"],model["b2"])[1]
def inspect_file(path:Path,max_windows:int=12, model_path:Path=MODEL_FILE)->dict:
"""Inspect strategically sampled 0.512s segments; bounded RAM, any duration."""
model=load_model(model_path)
scores=[]
with sf.SoundFile(str(path)) as source:
if source.frames<=0:
raise ValueError("Empty audio")
count=min(max_windows,max(1,math.ceil(source.frames/source.samplerate/.65)))
positions=np.linspace(0,max(0,source.frames-int(.55*source.samplerate)),count,dtype=int)
for pos in positions:
source.seek(int(pos))
block=source.read(min(source.frames-int(pos),int(.65*source.samplerate)),
dtype="float32",always_2d=True)
if not np.isfinite(block).all():raise ValueError("Invalid audio samples")
mono=block.mean(axis=1)
if len(mono)>0:
g=math.gcd(source.samplerate,SR)
mono=resample_poly(mono,SR//g,source.samplerate//g)
if len(mono)<N:mono=np.pad(mono,(0,N-len(mono)))
mono=mono[:N]
scores.append(predict_features(features(mono),model))
pred=np.stack(scores)
score=np.mean(pred,axis=0)*.6+np.max(pred,axis=0)*.4
return {
"model":"OMEGA-TAG Mini v1",
"model_kind":"tiny_neural_network__synthetic_pretrained_and_finetuned",
"checkpoint_sha256":hashlib.sha256(Path(model_path).read_bytes()).hexdigest(),
"input_sha256":hashlib.sha256(Path(path).read_bytes()).hexdigest(),
"sampled_windows":len(scores),
"scope":"SYNTHETIC SIGNAL TEXTURE TAGGING; no verified instrument identification",
"score_definition":"uncalibrated neural activation; NOT real-world probability",
"detections":[{"id":label,"label":FRIENDLY[label],"score":round(float(prob),4),
"suggested":bool(prob>=.55),
"note":"Hypothesis only. May be wrong for real recordings."}
for label,prob in zip(LABELS,score)],
"limitations":["No pretrained music foundation model",
"No isolated real music-stem training data",
"Scores are not calibrated for real recordings",
"Short sampled windows may miss rare or late events"],
}
if __name__=="__main__":
import sys
if sys.argv[1:]==["--train"]:
details=train_model()
print(json.dumps(details,indent=2))
else:
print("Use: python -m app.mini_model --train")