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
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
|
"""
CIFAR-10 L=4 d=512 confirmatory: BP/DFA/State Bridge/Credit Bridge, 5 seeds.
One method+seed per invocation for clean process isolation.
Usage: python cifar_d512_confirmatory.py --method bp --seed 42 --gpu 0
"""
import os, sys, json, csv, argparse, numpy as np, torch, torch.nn as nn, torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from models.residual_mlp import ResidualMLP
from models.value_net import ValueNet, create_ema_model, update_ema
from models.state_bridge import StateBridgeNet
from metrics.credit_metrics import cosine_similarity_batch, perturbation_correlation
import torchvision, torchvision.transforms as transforms
def get_cifar10(bs=128):
tt=transforms.Compose([transforms.RandomCrop(32,4),transforms.RandomHorizontalFlip(),transforms.ToTensor(),transforms.Normalize((0.4914,0.4822,0.4465),(0.2470,0.2435,0.2616))])
tv=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.4914,0.4822,0.4465),(0.2470,0.2435,0.2616))])
return (DataLoader(torchvision.datasets.CIFAR10('./data',True,download=True,transform=tt),bs,True,num_workers=4,pin_memory=True),
DataLoader(torchvision.datasets.CIFAR10('./data',False,download=True,transform=tv),bs,False,num_workers=4,pin_memory=True))
def evaluate(m,tl,dev):
m.eval();c,t=0,0
with torch.no_grad():
for x,y in tl:x=x.view(x.size(0),-1).to(dev);y=y.to(dev);c+=(m(x).argmax(1)==y).sum().item();t+=x.size(0)
return c/t
def train_bp(model, trl, tel, dev, epochs=100, lr=1e-3, wd=0.01):
opt=optim.AdamW(model.parameters(),lr=lr,weight_decay=wd)
sch=optim.lr_scheduler.CosineAnnealingLR(opt,T_max=epochs)
for ep in range(1,epochs+1):
model.train()
for x,y in trl:
x=x.view(x.size(0),-1).to(dev);y=y.to(dev)
F.cross_entropy(model(x),y).backward();opt.step();opt.zero_grad()
sch.step()
if ep%20==0:print(f" Ep {ep}: acc={evaluate(model,tel,dev):.4f}",flush=True)
return model
def train_dfa(model, trl, tel, dev, epochs=100, lr=1e-3, wd=0.01):
d=model.d_hidden;L=model.num_blocks;C=10
Bs=[torch.randn(d,C,device=dev)/np.sqrt(C) for _ in range(L)]
bops=[optim.AdamW(b.parameters(),lr=lr,weight_decay=wd) for b in model.blocks]
eop=optim.AdamW(model.embed.parameters(),lr=lr,weight_decay=wd)
hop=optim.AdamW(list(model.out_head.parameters())+list(model.out_ln.parameters()),lr=lr,weight_decay=wd)
schs=[optim.lr_scheduler.CosineAnnealingLR(o,T_max=epochs) for o in bops]+[optim.lr_scheduler.CosineAnnealingLR(eop,T_max=epochs),optim.lr_scheduler.CosineAnnealingLR(hop,T_max=epochs)]
for ep in range(1,epochs+1):
model.train()
for x,y in trl:
x=x.view(x.size(0),-1).to(dev);y=y.to(dev);b=x.size(0)
with torch.no_grad():lo,hi=model(x,return_hidden=True);eT=lo.softmax(-1);eT[torch.arange(b),y]-=1
hL=hi[-1].detach()
F.cross_entropy(model.out_head(model.out_ln(hL)),y).backward();hop.step();hop.zero_grad()
for l in range(L):
a=(eT@Bs[l].T).detach();rm=(a**2).mean(-1,keepdim=True).sqrt()+1e-6
ll=(model.blocks[l](hi[l].detach())*(a/rm)).sum(-1).mean()
bops[l].zero_grad();ll.backward();torch.nn.utils.clip_grad_norm_(model.blocks[l].parameters(),1.0);bops[l].step()
a0=(eT@Bs[0].T).detach();r0=(a0**2).mean(-1,keepdim=True).sqrt()+1e-6
(model.embed(x)*(a0/r0)).sum(-1).mean().backward();eop.step();eop.zero_grad()
for s in schs:s.step()
if ep%20==0:print(f" Ep {ep}: acc={evaluate(model,tel,dev):.4f}",flush=True)
return model
def train_state_bridge(model, trl, tel, dev, epochs=100, lr=1e-3, lr_fb=1e-3, wd=0.01):
d=model.d_hidden;L=model.num_blocks;C=10
sp=StateBridgeNet(d_hidden=d,s_dim=C).to(dev)
bops=[optim.AdamW(b.parameters(),lr=lr,weight_decay=wd) for b in model.blocks]
eop=optim.AdamW(model.embed.parameters(),lr=lr,weight_decay=wd)
hop=optim.AdamW(list(model.out_head.parameters())+list(model.out_ln.parameters()),lr=lr,weight_decay=wd)
sop=optim.Adam(sp.parameters(),lr=lr_fb)
schs=[optim.lr_scheduler.CosineAnnealingLR(o,T_max=epochs) for o in bops]+[optim.lr_scheduler.CosineAnnealingLR(eop,T_max=epochs),optim.lr_scheduler.CosineAnnealingLR(hop,T_max=epochs)]
se_sum=0
for ep in range(1,epochs+1):
model.train();sp.train();se_ep=0;n_ep=0
for x,y in trl:
x=x.view(x.size(0),-1).to(dev);y=y.to(dev);b=x.size(0)
with torch.no_grad():lo,hi=model(x,return_hidden=True);eT=lo.softmax(-1);eT[torch.arange(b),y]-=1;s=eT.detach()
hL=hi[-1].detach()
sl=0.0
for l in range(L):
tl=torch.full((b,),l/L,device=dev);pred=sp(hi[l].detach(),tl,s)
tn=hL.norm(-1,keepdim=True).clamp(min=1.0);sl+=(((pred-hL)/tn)**2).sum(-1).mean()
sl/=L;sop.zero_grad();sl.backward();sop.step();se_ep+=sl.item()*b;n_ep+=b
credits=[]
for l in range(L):
hl=hi[l].detach().requires_grad_(True);tl=torch.full((b,),l/L,device=dev)
pl=F.cross_entropy(model.out_head(model.out_ln(sp(hl,tl,s))),y,reduction='sum')
credits.append(torch.autograd.grad(pl,hl,create_graph=False)[0].detach())
F.cross_entropy(model.out_head(model.out_ln(hL)),y).backward();hop.step();hop.zero_grad()
for l in range(L):
a=credits[l];rm=(a**2).mean(-1,keepdim=True).sqrt()+1e-6
ll=(model.blocks[l](hi[l].detach())*(a/rm)).sum(-1).mean()
bops[l].zero_grad();ll.backward();torch.nn.utils.clip_grad_norm_(model.blocks[l].parameters(),1.0);bops[l].step()
a0=credits[0];r0=(a0**2).mean(-1,keepdim=True).sqrt()+1e-6
(model.embed(x)*(a0/r0)).sum(-1).mean().backward();eop.step();eop.zero_grad()
for s in schs:s.step()
se_sum=se_ep/n_ep
if ep%20==0:print(f" Ep {ep}: acc={evaluate(model,tel,dev):.4f} se={se_sum:.6f}",flush=True)
return model, se_sum
def train_credit_bridge(model, trl, tel, dev, epochs=100, lr=1e-3, lr_fb=1e-3, wd=0.01):
d=model.d_hidden;L=model.num_blocks;C=10
vn=ValueNet(d_hidden=d,s_dim=C).to(dev);ve=create_ema_model(vn)
Bs=[torch.randn(d,C,device=dev)/np.sqrt(C) for _ in range(L)]
bops=[optim.AdamW(b.parameters(),lr=lr,weight_decay=wd) for b in model.blocks]
eop=optim.AdamW(model.embed.parameters(),lr=lr,weight_decay=wd)
hop=optim.AdamW(list(model.out_head.parameters())+list(model.out_ln.parameters()),lr=lr,weight_decay=wd)
vop=optim.Adam(vn.parameters(),lr=lr_fb)
schs=[optim.lr_scheduler.CosineAnnealingLR(o,T_max=epochs) for o in bops]+[optim.lr_scheduler.CosineAnnealingLR(eop,T_max=epochs),optim.lr_scheduler.CosineAnnealingLR(hop,T_max=epochs)]
warmup=max(1,epochs//5)
for ep in range(1,epochs+1):
model.train();vn.train()
blend=0.0 if ep<=warmup else min(1.0,(ep-warmup)/max(1,warmup))
for x,y in trl:
x=x.view(x.size(0),-1).to(dev);y=y.to(dev);b=x.size(0)
with torch.no_grad():lo,hi=model(x,return_hidden=True);eT=lo.softmax(-1);eT[torch.arange(b),y]-=1;s=eT.detach();tlv=F.cross_entropy(lo,y,reduction='none').detach()
hL=hi[-1].detach();t_L=torch.ones(b,device=dev)
lt=((vn(hL,t_L,s)-tlv)**2).mean()
hLr=hL.clone().requires_grad_(True);VL=vn(hLr,t_L,s);gV=torch.autograd.grad(VL.sum(),hLr,create_graph=True)[0]
hLr2=hL.clone().requires_grad_(True);ce=F.cross_entropy(model.out_head(model.out_ln(hLr2)),y,reduction='sum')
aLe=torch.autograd.grad(ce,hLr2,create_graph=False)[0].detach()
ltg=((gV-aLe)**2).sum(-1).mean()
lb=0.0
for l in range(L):
hl=hi[l].detach();tl=torch.full((b,),l/L,device=dev);tn=torch.full((b,),(l+1)/L,device=dev)
Vl=vn(hl,tl,s)
with torch.no_grad():
hn=hi[l+1].detach();lts=[]
for k in range(4):lts.append(-ve(hn+0.05*torch.randn_like(hn),tn,s)/0.1)
Vt=-0.1*(torch.logsumexp(torch.stack(lts,-1),-1)-np.log(4))
lb+=((Vl-Vt.detach())**2).mean()
lb/=L;vl=lt+lb+1.0*ltg
vop.zero_grad();vl.backward();torch.nn.utils.clip_grad_norm_(vn.parameters(),1.0);vop.step()
update_ema(vn,ve,0.995)
cbc=[]
for l in range(L):
hl=hi[l].detach().requires_grad_(True);tl=torch.full((b,),l/L,device=dev)
Vl=vn(hl,tl,s);cbc.append(torch.autograd.grad(Vl.sum(),hl,create_graph=False)[0].detach())
dfac=[(eT@Bs[l].T).detach() for l in range(L)]
credits=[]
for l in range(L):
if blend>=1:credits.append(cbc[l])
elif blend<=0:credits.append(dfac[l])
else:
cr=(cbc[l]**2).mean(-1,keepdim=True).sqrt()+1e-6;dr=(dfac[l]**2).mean(-1,keepdim=True).sqrt()+1e-6
credits.append(blend*cbc[l]/cr+(1-blend)*dfac[l]/dr)
F.cross_entropy(model.out_head(model.out_ln(hL)),y).backward();hop.step();hop.zero_grad()
for l in range(L):
a=credits[l];rm=(a**2).mean(-1,keepdim=True).sqrt()+1e-6
ll=(model.blocks[l](hi[l].detach())*(a/rm)).sum(-1).mean()
bops[l].zero_grad();ll.backward();torch.nn.utils.clip_grad_norm_(model.blocks[l].parameters(),1.0);bops[l].step()
a0=credits[0];r0=(a0**2).mean(-1,keepdim=True).sqrt()+1e-6
(model.embed(x)*(a0/r0)).sum(-1).mean().backward();eop.step();eop.zero_grad()
for s in schs:s.step()
if ep%20==0:print(f" Ep {ep}: acc={evaluate(model,tel,dev):.4f}",flush=True)
return model
def compute_diagnostics(model, tel, dev):
model.eval();L=model.num_blocks;d=model.d_hidden
for x,y in tel:x=x.view(x.size(0),-1).to(dev);y=y.to(dev);break
b=x.size(0)
h0=model.embed(x.detach());hs=[h0.clone().requires_grad_(True)]
for bl in model.blocks:hs.append(hs[-1]+bl(hs[-1]))
lo=model.out_head(model.out_ln(hs[-1]));loss=F.cross_entropy(lo,y)
gs=torch.autograd.grad(loss,hs);bp={l:gs[l].detach() for l in range(L)}
with torch.no_grad():_,hi=model(x,return_hidden=True);nse=((hi[L//2]-hi[-1]).norm(-1)/hi[-1].norm(-1).clamp(min=1e-8)).mean().item()
gammas,rhos=[],[]
for l in range(L):
gammas.append(1.0) # BP self-cosine
h_l=hi[l].detach();a_l=bp[l]
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(model.out_ln(c)),y,reduction='none')
return f
rhos.append(perturbation_correlation(h_l,a_l,mk(l),epsilon=1e-3,M=16))
return {'Gamma':np.mean(gammas),'rho':np.mean(rhos),'naive_StateErr':nse}
def main():
p=argparse.ArgumentParser()
p.add_argument('--method',type=str,required=True)
p.add_argument('--seed',type=int,required=True)
p.add_argument('--gpu',type=int,default=0)
p.add_argument('--output_dir',type=str,default='results/confirmatory/cifar_d512')
args=p.parse_args()
os.makedirs(args.output_dir,exist_ok=True)
dev=torch.device(f'cuda:{args.gpu}')
torch.manual_seed(args.seed);np.random.seed(args.seed);torch.cuda.manual_seed_all(args.seed)
trl,tel=get_cifar10()
L,d=4,512
model=ResidualMLP(3072,d,10,L).to(dev)
print(f"[{args.method} s={args.seed}] Training...",flush=True)
se=None
if args.method=='bp':model=train_bp(model,trl,tel,dev)
elif args.method=='dfa':model=train_dfa(model,trl,tel,dev)
elif args.method=='state_bridge':model,se=train_state_bridge(model,trl,tel,dev)
elif args.method=='credit_bridge':model=train_credit_bridge(model,trl,tel,dev)
acc=evaluate(model,tel,dev)
diag=compute_diagnostics(model,tel,dev)
# Save checkpoint
torch.save(model.state_dict(),os.path.join(args.output_dir,f'{args.method}_s{args.seed}.pt'))
result={'method':args.method,'seed':args.seed,'acc':acc,'StateErr':se,
'Gamma':diag['Gamma'],'rho':diag['rho'],'naive_StateErr':diag['naive_StateErr']}
with open(os.path.join(args.output_dir,f'{args.method}_s{args.seed}.json'),'w') as f:
json.dump(result,f,indent=2,default=float)
print(f"[{args.method} s={args.seed}] acc={acc:.4f} Γ={diag['Gamma']:.4f} ρ={diag['rho']:.4f} nse={diag['naive_StateErr']:.4f}",flush=True)
if __name__=='__main__':main()
|
