v12: signal-proportional phase — TIES SinGLU 3-3

benchmark_v12.py

@@ -0,0 +1,254 @@
+#!/usr/bin/env python3
+"""
+v12: SinGLU + SIGNAL-PROPORTIONAL Phase
+
+v10: φ independent of signal → too hot, destroys structure
+v11: φ tiny fixed scale → too cold, does nothing  
+v12: φ proportional to signal → adapts automatically
+
+    g = Wg·x                           (features)
+    φ = tanh(Wφ·x)                     (modulation signal, bounded [-1,1])
+    core = sin(ω · g · (1 + 0.2·φ))    (signal-proportional phase warping)
+    y = LN( Wo( core ⊙ Wv·x ) )
+
+When g is large: phase effect is large (where the signal is strong)
+When g is small: phase effect is small (where signal is weak)
+On OOD: g·φ tracks signal magnitude, doesn't add uncorrelated noise
+"""
+
+import torch, torch.nn as nn, torch.nn.functional as F
+import numpy as np, math, json
+
+SEEDS = [0, 1, 2]
+def set_seed(s): torch.manual_seed(s); np.random.seed(s)
+
+class VanillaMLP(nn.Module):
+    def __init__(self,di,do,h,n):
+        super().__init__(); L=[]; p=di
+        for _ in range(n): L+=[nn.Linear(p,h),nn.ReLU()]; p=h
+        L.append(nn.Linear(p,do)); self.net=nn.Sequential(*L)
+    def forward(self,x): return self.net(x)
+
+class SinGLULayer(nn.Module):
+    def __init__(self,di,do,mid,w0=30.):
+        super().__init__()
+        self.Wg=nn.Linear(di,mid,bias=False); self.Wv=nn.Linear(di,mid,bias=False)
+        self.Wo=nn.Linear(mid,do,bias=True); self.w0=w0; self.ln=nn.LayerNorm(do)
+        with torch.no_grad():
+            self.Wg.weight.uniform_(-math.sqrt(6/di)/w0,math.sqrt(6/di)/w0)
+            nn.init.xavier_uniform_(self.Wv.weight); nn.init.xavier_uniform_(self.Wo.weight)
+    def forward(self,x): return self.ln(self.Wo(torch.sin(self.w0*self.Wg(x))*self.Wv(x)))
+
+class SinGLUNet(nn.Module):
+    def __init__(self,di,do,h,n,w0=30.):
+        super().__init__(); mid=max(2,int(h*2/3)); L=[]; p=di
+        for _ in range(n): L.append(SinGLULayer(p,h,mid,w0)); p=h
+        L.append(nn.Linear(p,do)); self.layers=nn.ModuleList(L)
+    def forward(self,x):
+        for l in self.layers: x=l(x)
+        return x
+
+class v12Layer(nn.Module):
+    """
+    g = Wg·x
+    φ = tanh(Wφ·x)  
+    core = sin(ω · g · (1 + 0.2·φ))     ← phase proportional to signal
+    y = LN(Wo(core ⊙ Wv·x))
+    """
+    def __init__(self, di, do, mid, w0=30., warp_scale=0.2):
+        super().__init__()
+        self.Wg=nn.Linear(di,mid,bias=False)
+        self.Wv=nn.Linear(di,mid,bias=False)
+        self.Wo=nn.Linear(mid,do,bias=True)
+        self.Wphi=nn.Linear(di,mid,bias=True)
+        self.w0=w0; self.ws=warp_scale; self.ln=nn.LayerNorm(do)
+        with torch.no_grad():
+            self.Wg.weight.uniform_(-math.sqrt(6/di)/w0,math.sqrt(6/di)/w0)
+            nn.init.xavier_uniform_(self.Wv.weight); nn.init.xavier_uniform_(self.Wo.weight)
+            nn.init.zeros_(self.Wphi.weight); nn.init.zeros_(self.Wphi.bias)
+    
+    def forward(self,x):
+        g = self.Wg(x)
+        phi = torch.tanh(self.Wphi(x))
+        warped = g * (1. + self.ws * phi)       # signal-proportional warping
+        core = torch.sin(self.w0 * warped)
+        return self.ln(self.Wo(core * self.Wv(x)))
+    
+    def get_phi(self,x):
+        with torch.no_grad(): return torch.tanh(self.Wphi(x))
+
+class v12Net(nn.Module):
+    def __init__(self,di,do,h,n,w0=30.):
+        super().__init__(); mid=max(2,int(h*2/3)); L=[]; p=di
+        for _ in range(n): L.append(v12Layer(p,h,mid,w0)); p=h
+        L.append(nn.Linear(p,do)); self.layers=nn.ModuleList(L)
+    def forward(self,x):
+        for l in self.layers: x=l(x)
+        return x
+    def get_all_phi(self,x):
+        P=[]; h=x
+        for l in self.layers:
+            if isinstance(l,v12Layer): P.append(l.get_phi(h)); h=l(h)
+            else: h=l(h)
+        return P
+
+# Utils
+def np_(m): return sum(p.numel() for p in m.parameters() if p.requires_grad)
+def fh(di,do,n,t,cls,**kw):
+    lo,hi,b=2,512,2
+    while lo<=hi:
+        mid=(lo+hi)//2; p=np_(cls(di,do,mid,n,**kw))
+        if abs(p-t)<abs(np_(cls(di,do,b,n,**kw))-t): b=mid
+        if p<t: lo=mid+1
+        else: hi=mid-1
+    return b
+
+def tr_r(m,xt,yt,xe,ye,ep,lr,bs=256):
+    o=torch.optim.Adam(m.parameters(),lr=lr); s=torch.optim.lr_scheduler.CosineAnnealingLR(o,T_max=ep)
+    best=float('inf'); n=len(xt)
+    for e in range(ep):
+        m.train(); p=torch.randperm(n)
+        for i in range(0,n,bs):
+            idx=p[i:i+bs]; loss=F.mse_loss(m(xt[idx]),yt[idx])
+            o.zero_grad(); loss.backward(); torch.nn.utils.clip_grad_norm_(m.parameters(),1.); o.step()
+        s.step()
+        if(e+1)%max(1,ep//10)==0:
+            m.eval()
+            with torch.no_grad(): best=min(best,F.mse_loss(m(xe),ye).item())
+    m.eval()
+    with torch.no_grad(): best=min(best,F.mse_loss(m(xe),ye).item())
+    return best
+
+def tr_c(m,xt,yt,xe,ye,ep,lr,bs=256):
+    o=torch.optim.Adam(m.parameters(),lr=lr); s=torch.optim.lr_scheduler.CosineAnnealingLR(o,T_max=ep)
+    best=0; n=len(xt)
+    for e in range(ep):
+        m.train(); p=torch.randperm(n)
+        for i in range(0,n,bs):
+            idx=p[i:i+bs]; loss=F.cross_entropy(m(xt[idx]),yt[idx])
+            o.zero_grad(); loss.backward(); torch.nn.utils.clip_grad_norm_(m.parameters(),1.); o.step()
+        s.step()
+        if(e+1)%max(1,ep//10)==0:
+            m.eval()
+            with torch.no_grad(): best=max(best,(m(xe).argmax(1)==ye).float().mean().item())
+    m.eval()
+    with torch.no_grad(): best=max(best,(m(xe).argmax(1)==ye).float().mean().item())
+    return best
+
+# Data
+def d_cx(n=1000): x=torch.rand(n,4)*2-1; return x,torch.exp(torch.sin(x[:,0]**2+x[:,1]**2)+torch.sin(x[:,2]**2+x[:,3]**2)).unsqueeze(1)
+def d_ne(n=1000): x=torch.rand(n,2)*2-1; return x,(torch.sin(math.pi*(x[:,0]**2+x[:,1]**2))*torch.cos(3*math.pi*x[:,0]*x[:,1])).unsqueeze(1)
+def d_sp(n=1000):
+    t=torch.linspace(0,4*np.pi,n//2); r=torch.linspace(.3,2,n//2)
+    x=torch.cat([torch.stack([r*torch.cos(t),r*torch.sin(t)],1),torch.stack([r*torch.cos(t+np.pi),r*torch.sin(t+np.pi)],1)])+torch.randn(n,2)*.05
+    y=torch.cat([torch.zeros(n//2),torch.ones(n//2)]).long(); p=torch.randperm(n); return x[p],y[p]
+def d_ch(n=1000): x=torch.rand(n,2)*2-1; return x,((torch.sin(3*math.pi*x[:,0])*torch.sin(3*math.pi*x[:,1]))>0).long()
+def d_hf(n=1000): x=torch.linspace(-1,1,n).unsqueeze(1); return x,torch.sin(20*x)+torch.sin(50*x)+.5*torch.sin(100*x)
+def d_mm(n=200): return torch.randn(n,8),torch.randn(n,4)
+def d_ot(n=800): x=torch.rand(n,2)*2-1; return x,(torch.sin(3*math.pi*x[:,0])*torch.cos(3*math.pi*x[:,1])+x[:,0]*x[:,1]).unsqueeze(1)
+def d_oe(n=300): x=torch.rand(n,2)+1; return x,(torch.sin(3*math.pi*x[:,0])*torch.cos(3*math.pi*x[:,1])+x[:,0]*x[:,1]).unsqueeze(1)
+
+def main():
+    print("="*80)
+    print("  v12: SinGLU + SIGNAL-PROPORTIONAL Phase")
+    print("  core = sin(ω · g · (1 + 0.2·tanh(Wφ·x)))  — phase scales with signal")
+    print("="*80)
+    
+    N=3; Ms={'Vanilla':(VanillaMLP,{}),'SinGLU':(SinGLUNet,{'w0':None}),'v12':(v12Net,{'w0':None})}
+    tasks=[
+        ("Complex Fn","r",d_cx,4,1,5000,300,1e-3,30.,750),
+        ("Nested Fn","r",d_ne,2,1,3000,300,1e-3,20.,750),
+        ("Spiral","c",d_sp,2,2,3000,250,1e-3,15.,700),
+        ("Checker","c",d_ch,2,2,3000,250,1e-3,20.,700),
+        ("High-Freq","r",d_hf,1,1,8000,300,1e-3,60.,700),
+        ("Memorize","r",d_mm,8,4,5000,400,1e-3,10.,200),
+    ]
+    R={}; PH={}
+    for tn,tt,df,di,do,bud,ep,lr,w0,sp in tasks:
+        print(f"\n{'━'*80}\n  {tn}\n{'━'*80}")
+        hs={mn:fh(di,do,N,bud,mc,**{k:(w0 if v is None else v) for k,v in mk.items()}) for mn,(mc,mk) in Ms.items()}
+        tr={}
+        for mn,(mc,mk) in Ms.items():
+            kw={k:(w0 if v is None else v) for k,v in mk.items()}; h=hs[mn]; sc=[]
+            for seed in SEEDS:
+                set_seed(seed); x,y=df()
+                if sp>=len(x): xt,yt,xe,ye=x,y,x,y
+                else: xt,yt,xe,ye=x[:sp],y[:sp],x[sp:],y[sp:]
+                set_seed(seed+100); mdl=mc(di,do,h,N,**kw)
+                s=tr_r(mdl,xt,yt,xe,ye,ep,lr) if tt=='r' else tr_c(mdl,xt,yt,xe,ye,ep,lr)
+                sc.append(s)
+                if mn=='v12' and seed==SEEDS[-1]:
+                    mdl.eval()
+                    with torch.no_grad():
+                        pp=mdl.get_all_phi(xe[:100]); ap=torch.cat([p.flatten() for p in pp])
+                        PH[tn]={'m':ap.mean().item(),'s':ap.std().item(),'mn':ap.min().item(),'mx':ap.max().item()}
+            p=np_(mc(di,do,h,N,**kw))
+            tr[mn]={'mean':np.mean(sc),'std':np.std(sc),'scores':sc,'params':p,'hidden':h}
+        ir=tt=='r'
+        best=min(tr,key=lambda k:tr[k]['mean']) if ir else max(tr,key=lambda k:tr[k]['mean'])
+        met="MSE ↓" if ir else "Acc ↑"
+        print(f"\n  {'M':<8} {'H':>3} {'P':>6} {met+' (mean±std)':>26}")
+        print(f"  {'─'*46}")
+        for mn,r in tr.items():
+            m,s=r['mean'],r['std']
+            ms=f"{m:.2e}±{s:.1e}" if(ir and m<.001) else(f"{m:.4f}±{s:.4f}" if ir else f"{m:.1%}±{s:.3f}")
+            print(f"  {mn:<8} {r['hidden']:>3} {r['params']:>6,} {ms:>26}{' ★' if mn==best else ''}")
+        print(f"  → {best}")
+        if tn in PH: d=PH[tn]; print(f"    φ: mean={d['m']:.4f} std={d['s']:.4f} [{d['mn']:.3f},{d['mx']:.3f}]")
+        R[tn]=tr
+    
+    # OOD
+    print(f"\n{'━'*80}\n  OOD: [-1,1] → [1,2]\n{'━'*80}")
+    OD={}
+    for mn,(mc,mk) in Ms.items():
+        kw={k:(20. if v is None else v) for k,v in mk.items()}; h=fh(2,1,N,5000,mc,**kw); ids,ods=[],[]
+        for seed in SEEDS:
+            set_seed(seed); xtr,ytr=d_ot()
+            set_seed(seed+50); xi=torch.rand(200,2)*2-1; yi=(torch.sin(3*math.pi*xi[:,0])*torch.cos(3*math.pi*xi[:,1])+xi[:,0]*xi[:,1]).unsqueeze(1)
+            set_seed(seed+50); xo,yo=d_oe()
+            set_seed(seed+100); mdl=mc(2,1,h,N,**kw)
+            si=tr_r(mdl,xtr,ytr,xi,yi,300,1e-3); mdl.eval()
+            with torch.no_grad(): so=F.mse_loss(mdl(xo),yo).item()
+            ids.append(si); ods.append(so)
+        OD[mn]={'id':np.mean(ids),'ood':np.mean(ods),'deg':np.mean(ods)/max(np.mean(ids),1e-10),'is':np.std(ids),'os':np.std(ods)}
+    bo=min(OD,key=lambda k:OD[k]['ood'])
+    print(f"\n  {'M':<8} {'ID':>12} {'OOD':>12} {'Deg':>7}")
+    print(f"  {'─'*42}")
+    for mn,r in OD.items(): print(f"  {mn:<8} {r['id']:>8.4f}±{r['is']:.3f} {r['ood']:>8.4f}±{r['os']:.3f} {r['deg']:>6.1f}x{' ★' if mn==bo else ''}")
+    print(f"  → {bo}")
+    R['OOD']={mn:{'mean':r['ood'],'std':r['os']} for mn,r in OD.items()}
+    
+    # Summary
+    print(f"\n{'='*80}\n  FINAL SUMMARY\n{'='*80}")
+    wc={k:0 for k in Ms}
+    print(f"\n  {'Task':<14}",end="")
+    for mn in Ms: print(f" {mn:>12}",end="")
+    print(f" {'W':>8}")
+    print(f"  {'─'*50}")
+    for tn,t in R.items():
+        sc={k:v['mean'] for k,v in t.items()}; mx=max(sc.values())
+        ic=mx>.5 and mx<=1 and min(sc.values())>=0
+        if min(sc.values())<.001: ic=False
+        w=min(sc,key=sc.get) if(tn=='OOD' or not ic) else max(sc,key=sc.get)
+        wc[w]+=1
+        row=f"  {tn:<14}"
+        for mn in Ms:
+            s=sc[mn]
+            if ic: row+=f" {s:>11.1%}"
+            elif s<.001: row+=f" {s:>11.2e}"
+            else: row+=f" {s:>11.4f}"
+        row+=f"  ->{w}"; print(row)
+    print(f"\n  {'─'*50}")
+    for mn,c in sorted(wc.items(),key=lambda x:-x[1]):
+        print(f"    {mn:<8} {c} {'█'*c*4}")
+    
+    sv={'tasks':{},'ood':{},'phi':PH}
+    for tn,t in R.items():
+        sv['tasks'][tn]={mn:{'mean':float(r['mean']),'std':float(r.get('std',0)),
+            'scores':[float(s) for s in r.get('scores',[r['mean']])],
+            'params':r.get('params',0),'hidden':r.get('hidden',0)} for mn,r in t.items()}
+    sv['ood']={mn:{k:float(v) if isinstance(v,(float,np.floating)) else v for k,v in r.items()} for mn,r in OD.items()}
+    with open('/app/results_v12.json','w') as f: json.dump(sv,f,indent=2,default=str)
+    print(f"\n  Saved.")
+
+if __name__=="__main__": main()