#!/usr/bin/env python3
"""
v13: ALIGNED PHASE — phase along signal axes, not independent

v10: sin(ω·g + π·tanh(Wφ·x))      ← additive but independent → chaotic
v12: sin(ω · g·(1+0.2·φ))          ← frequency modulation → drift
v13: sin(ω·g + 0.1·g·tanh(Wφ·x))  ← additive phase, aligned to g → stable

The key: φ ∝ g. Phase only shifts where signal exists.
No frequency drift (ω stays fixed). No independent 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

# v10 (free phase) for comparison
class v10Layer(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.Wphi=nn.Linear(di,mid,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)
            nn.init.zeros_(self.Wphi.weight); nn.init.zeros_(self.Wphi.bias)
    def forward(self,x):
        phi=math.pi*torch.tanh(self.Wphi(x))
        return self.ln(self.Wo(torch.sin(self.w0*self.Wg(x)+phi)*self.Wv(x)))

class v10Net(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(v10Layer(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

# v12 (FM) for comparison
class v12Layer(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.Wphi=nn.Linear(di,mid,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)
            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))
        return self.ln(self.Wo(torch.sin(self.w0*g*(1.+0.2*phi))*self.Wv(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

# v13: ALIGNED PHASE
class v13Layer(nn.Module):
    """
    g = Wg·x
    φ = 0.1 · g · tanh(Wφ·x)     ← phase ALIGNED to signal
    core = sin(ω·g + φ)            ← additive, no freq drift
    y = LN(Wo(core ⊙ Wv·x))
    """
    def __init__(self,di,do,mid,w0=30.,alpha=0.1):
        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.a=alpha; 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 = self.a * g * torch.tanh(self.Wphi(x))  # aligned to signal
        core = torch.sin(self.w0 * g + phi)            # additive phase, no freq drift
        return self.ln(self.Wo(core * self.Wv(x)))
    
    def get_corr(self,x):
        """Measure correlation between g and phi"""
        with torch.no_grad():
            g=self.Wg(x); phi=self.a*g*torch.tanh(self.Wphi(x))
            # pearson correlation per neuron, averaged
            gf=g.flatten(); pf=phi.flatten()
            if gf.std()==0 or pf.std()==0: return 0.
            return ((gf-gf.mean())*(pf-pf.mean())).mean()/(gf.std()*pf.std()+1e-8)

class v13Net(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(v13Layer(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_corrs(self,x):
        cs=[]; h=x
        for l in self.layers:
            if isinstance(l,v13Layer): cs.append(l.get_corr(h).item()); h=l(h)
            else: h=l(h)
        return cs

# 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("  v13: ALIGNED PHASE  |  sin(ω·g + 0.1·g·tanh(Wφ·x))")
    print("  + corr(g,φ) analysis  |  vs SinGLU, v10(free), v12(FM)")
    print("="*80)
    
    N=3
    Ms={'Vanilla':(VanillaMLP,{}),'SinGLU':(SinGLUNet,{'w0':None}),
        'v10:free':(v10Net,{'w0':None}),'v12:FM':(v12Net,{'w0':None}),'v13':(v13Net,{'w0':None})}
    
    tasks=[
        ("Complex","r",d_cx,4,1,5000,300,1e-3,30.,750),
        ("Nested","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),
        ("HiFreq","r",d_hf,1,1,8000,300,1e-3,60.,700),
        ("Memorize","r",d_mm,8,4,5000,400,1e-3,10.,200),
    ]
    R={}; CORR={}
    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=='v13' and seed==SEEDS[-1]:
                    mdl.eval(); CORR[tn]=mdl.get_corrs(xe[:100])
            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':<10} {'H':>3} {'P':>6} {met:>24}")
        print(f"  {'─'*46}")
        for mn,r in tr.items():
            m=r['mean']; s=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:<10} {r['hidden']:>3} {r['params']:>6,} {ms:>24}{' ★' if mn==best else ''}")
        print(f"  → {best}")
        if tn in CORR: print(f"    v13 corr(g,φ) per layer: {['%.3f'%c for c in CORR[tn]]}")
        R[tn]=tr
    
    # OOD
    print(f"\n{'━'*80}\n  OOD\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':<10} {'ID':>12} {'OOD':>12} {'Deg':>7}")
    print(f"  {'─'*44}")
    for mn,r in OD.items(): print(f"  {mn:<10} {r['id']:>8.4f}±{r['is']:.3f} {r['ood']:>8.4f}±{r['os']:.3f} {r['deg']:>6.1f}x{' ★' if mn==bo else ''}")
    R['OOD']={mn:{'mean':r['ood'],'std':r['os']} for mn,r in OD.items()}
    
    # Summary
    print(f"\n{'='*80}\n  SUMMARY\n{'='*80}")
    wc={k:0 for k in Ms}
    print(f"\n  {'Task':<10}",end="")
    for mn in Ms: print(f" {mn:>10}",end="")
    print(f"  {'W':>8}")
    print(f"  {'─'*68}")
    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:<10}"
        for mn in Ms:
            s=sc[mn]
            if ic: row+=f" {s:>9.1%}"
            elif s<.001: row+=f" {s:>9.2e}"
            else: row+=f" {s:>9.4f}"
        row+=f"  ->{w}"; print(row)
    print(f"\n  {'─'*68}")
    for mn,c in sorted(wc.items(),key=lambda x:-x[1]):
        print(f"    {mn:<10} {c} {'█'*c*3}")
    
    print(f"\n  corr(g,φ) — does v13 phase align with signal?")
    for tn,cs in CORR.items():
        print(f"    {tn:<10} layers: {['%.3f'%c for c in cs]}  avg={np.mean(cs):.3f}")
    
    sv={'tasks':{},'ood':{},'corr':CORR}
    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_v13.json','w') as f: json.dump(sv,f,indent=2,default=str)
    print(f"\n  Saved.")

if __name__=="__main__": main()