1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
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
79
80
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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
|
"""EP on Synthetic ladder. Same EP algorithm as ep_baseline.py but for StudentNet."""
import os, sys, json, argparse, numpy as np, torch, torch.nn as nn, torch.nn.functional as F
import torch.optim as optim
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from metrics.credit_metrics import cosine_similarity_batch, perturbation_correlation
class StudentBlock(nn.Module):
def __init__(self, d, alpha=1.0):
super().__init__()
self.ln=nn.LayerNorm(d);self.w=nn.Linear(d,d,bias=False)
nn.init.normal_(self.w.weight,std=0.01);self.alpha=alpha
def forward(self, h):
return self.w(((1-self.alpha)*self.ln(h)+self.alpha*torch.tanh(self.ln(h))))
class StudentNet(nn.Module):
def __init__(self, d, C, L, alpha=1.0):
super().__init__()
self.blocks=nn.ModuleList([StudentBlock(d,alpha) for _ in range(L)])
self.out_head=nn.Linear(d,C);self.num_blocks=L;self.d_hidden=d
def forward(self, x, return_hidden=False):
h=x;hi=[h] if return_hidden else None
for b in self.blocks:
h=h+b(h)
if return_hidden:hi.append(h)
lo=self.out_head(h)
return (lo,hi) if return_hidden else lo
class TeacherNet(nn.Module):
def __init__(self, d, C, L, alpha=1.0, seed=0):
super().__init__()
self.alpha=alpha;rng=torch.Generator().manual_seed(seed)
self.Ws=nn.ParameterList()
for _ in range(L):
W=torch.randn(d,d,generator=rng)*0.3/(d**0.5)
U,S,Vh=torch.linalg.svd(W,full_matrices=False)
self.Ws.append(nn.Parameter(U@torch.diag(S.clamp(max=0.3))@Vh,requires_grad=False))
self.U=nn.Parameter(torch.randn(C,d,generator=rng)/(d**0.5),requires_grad=False)
def forward(self, x):
h=x
for W in self.Ws:h=h+((1-self.alpha)*h+self.alpha*torch.tanh(h))@W.T
return h@self.U.T
def train_ep_synth(model, teacher, dev, d, C, L, epochs=80, steps=50, bs=256,
lr=1e-3, wd=0.01, beta=0.5, T_nudge=20, alpha_nudge=0.1):
block_opts=[optim.AdamW(b.parameters(),lr=lr,weight_decay=wd) for b in model.blocks]
head_opt=optim.AdamW(model.out_head.parameters(),lr=lr,weight_decay=wd)
for ep in range(1,epochs+1):
model.train()
for _ in range(steps):
x=torch.randn(bs,d,device=dev)
with torch.no_grad():y=teacher(x).argmax(-1)
# Free phase
with torch.no_grad():_,hiddens_free=model(x,return_hidden=True)
h_free=[h.detach().clone() for h in hiddens_free]
# Nudged phase
h_nudge=[h.clone().requires_grad_(True) for h in h_free]
for t in range(T_nudge):
E=0.0
for l in range(L):
res=h_nudge[l+1]-(h_nudge[l].detach()+model.blocks[l](h_nudge[l].detach()))
E+=0.5*(res**2).sum(-1).mean()
lo=model.out_head(h_nudge[-1])
cost=F.cross_entropy(lo,y)
total=E+beta*cost
grads=torch.autograd.grad(total,[h_nudge[l] for l in range(1,L+1)],create_graph=False)
with torch.no_grad():
for l in range(L):
h_nudge[l+1]=h_nudge[l+1]-alpha_nudge*grads[l]
h_nudge[l+1].requires_grad_(True)
# Weight updates
for l in range(L):
res_free=h_free[l+1]-(h_free[l]+model.blocks[l](h_free[l]))
res_nudge=h_nudge[l+1].detach()-(h_nudge[l].detach()+model.blocks[l](h_nudge[l].detach()))
loss_free=(-res_free.detach()*model.blocks[l](h_free[l])).sum(-1).mean()
loss_nudge=(-res_nudge.detach()*model.blocks[l](h_nudge[l].detach())).sum(-1).mean()
block_loss=(loss_nudge-loss_free)/beta
block_opts[l].zero_grad();block_loss.backward()
torch.nn.utils.clip_grad_norm_(model.blocks[l].parameters(),1.0)
block_opts[l].step()
lo_n=model.out_head(h_nudge[-1].detach())
head_loss=F.cross_entropy(lo_n,y)
head_opt.zero_grad();head_loss.backward();head_opt.step()
if ep%20==0:
model.eval()
with torch.no_grad():
xt=torch.randn(512,d,device=dev);yt=teacher(xt).argmax(-1)
acc=(model(xt).argmax(1)==yt).float().mean().item()
print(f" Ep {ep}: acc={acc:.4f}",flush=True)
def compute_diagnostics(model, teacher, dev, d, C, L, beta=0.5, T_nudge=20, alpha_nudge=0.1):
model.eval()
x=torch.randn(512,d,device=dev)
with torch.no_grad():y=teacher(x).argmax(-1)
# BP grads
hs=[x.detach().requires_grad_(True)]
for b in model.blocks:hs.append(hs[-1]+b(hs[-1]))
lo=model.out_head(hs[-1]);loss=F.cross_entropy(lo,y)
gs=torch.autograd.grad(loss,hs);bp={l:gs[l].detach() for l in range(L+1)}
# EP credit
with torch.no_grad():_,hf=model(x,return_hidden=True)
h_free=[h.detach().clone() for h in hf]
h_nudge=[h.clone().requires_grad_(True) for h in h_free]
for t in range(T_nudge):
E=0.0
for l in range(L):
res=h_nudge[l+1]-(h_nudge[l].detach()+model.blocks[l](h_nudge[l].detach()))
E+=0.5*(res**2).sum(-1).mean()
lo2=model.out_head(h_nudge[-1]);cost=F.cross_entropy(lo2,y)
total=E+beta*cost
grads=torch.autograd.grad(total,[h_nudge[l] for l in range(1,L+1)],create_graph=False)
with torch.no_grad():
for l in range(L):
h_nudge[l+1]=h_nudge[l+1]-alpha_nudge*grads[l]
h_nudge[l+1].requires_grad_(True)
gammas,rhos=[],[]
with torch.no_grad():_,hi=model(x,return_hidden=True)
for l in range(L):
# EP nudge moves h toward lower loss, so (h_nudge - h_free) points opposite to BP grad.
# Negate to align with BP gradient convention (pointing toward loss increase).
a_ep=-(h_nudge[l+1].detach()-h_free[l+1])/beta
gammas.append(cosine_similarity_batch(a_ep,bp[l+1]))
def mk(sl):
def f(h):
with torch.no_grad():
c=h
for i in range(sl,L):c=c+model.blocks[i](c)
return F.cross_entropy(model.out_head(c),y,reduction='none')
return f
rhos.append(perturbation_correlation(hi[l+1].detach(),a_ep,mk(l+1),epsilon=1e-3,M=16))
acc=(model(x).argmax(1)==y).float().mean().item()
return {'Gamma':float(np.mean(gammas)),'rho':float(np.mean(rhos)),'acc':acc}
def main():
p=argparse.ArgumentParser()
p.add_argument('--alpha',type=float,required=True)
p.add_argument('--seed',type=int,required=True)
p.add_argument('--gpu',type=int,default=0)
p.add_argument('--depth',type=int,default=4)
p.add_argument('--output_dir',type=str,default='results/ep_synthetic')
args=p.parse_args()
os.makedirs(args.output_dir,exist_ok=True)
dev=torch.device(f'cuda:{args.gpu}')
d,C,L=128,10,args.depth
torch.manual_seed(args.seed);np.random.seed(args.seed);torch.cuda.manual_seed_all(args.seed)
teacher=TeacherNet(d,C,L,args.alpha,seed=0).to(dev)
model=StudentNet(d,C,L,args.alpha).to(dev)
print(f"[EP synth a={args.alpha} L={L} s={args.seed}] Training...",flush=True)
train_ep_synth(model,teacher,dev,d,C,L)
diag=compute_diagnostics(model,teacher,dev,d,C,L)
result={'method':'ep','alpha':args.alpha,'depth':L,'seed':args.seed,
'acc':diag['acc'],'Gamma':diag['Gamma'],'rho':diag['rho']}
torch.save(model.state_dict(),os.path.join(args.output_dir,f'ep_a{args.alpha}_L{L}_s{args.seed}.pt'))
out=os.path.join(args.output_dir,f'ep_a{args.alpha}_L{L}_s{args.seed}.json')
with open(out,'w') as f:json.dump(result,f,indent=2,default=float)
print(f" acc={diag['acc']:.4f} Gamma={diag['Gamma']:.4f} rho={diag['rho']:.4f}",flush=True)
if __name__=='__main__':main()
|
