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
|
"""
Clean full sparsity analysis — one method+seed per invocation.
Usage: python clean_sparsity_full.py --dataset cifar --method bp --seed 42 --gpu 0
python clean_sparsity_full.py --dataset synth --method bp --seed 42 --alpha 1.0 --depth 4 --gpu 0
"""
import os, sys, json, argparse, numpy as np, torch, torch.nn.functional as F
from torch.utils.data import DataLoader, TensorDataset
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from models.residual_mlp import ResidualMLP
import torchvision, torchvision.transforms as transforms
class StudentBlock(torch.nn.Module):
def __init__(self, d, alpha=1.0):
super().__init__()
self.ln=torch.nn.LayerNorm(d);self.w=torch.nn.Linear(d,d,bias=False)
torch.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(torch.nn.Module):
def __init__(self, d, C, L, alpha=1.0):
super().__init__()
self.blocks=torch.nn.ModuleList([StudentBlock(d,alpha) for _ in range(L)])
self.out_head=torch.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(torch.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=torch.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(torch.nn.Parameter(U@torch.diag(S.clamp(max=0.3))@Vh,requires_grad=False))
self.U=torch.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 main():
p = argparse.ArgumentParser()
p.add_argument('--dataset', type=str, required=True, choices=['cifar','synth'])
p.add_argument('--method', type=str, required=True)
p.add_argument('--seed', type=int, required=True)
p.add_argument('--alpha', type=float, default=1.0)
p.add_argument('--depth', type=int, default=4)
p.add_argument('--gpu', type=int, default=0)
p.add_argument('--output_dir', type=str, default='results/confirmatory/clean_sparsity')
args = p.parse_args()
os.makedirs(args.output_dir, exist_ok=True)
device = torch.device(f'cuda:{args.gpu}')
thresholds = [1e-8, 1e-7, 1e-6, 1e-5, 1e-4]
if args.dataset == 'cifar':
L, d, C = 4, 256, 10
tv = transforms.Compose([transforms.ToTensor(),
transforms.Normalize((0.4914,0.4822,0.4465),(0.2470,0.2435,0.2616))])
tel = DataLoader(torchvision.datasets.CIFAR10('./data', False, download=True, transform=tv),
256, False, num_workers=0)
for x, y in tel: x = x.view(x.size(0),-1).to(device); y = y.to(device); break
ckpt = f'results/confirmatory/checkpoints_A2/{args.method}_s{args.seed}.pt'
model = ResidualMLP(3072, d, C, L).to(device)
model.load_state_dict(torch.load(ckpt, map_location=device), strict=True)
model.eval()
is_cifar = True
else:
L, d, C = args.depth, 128, 10
teacher = TeacherNet(d, C, L, args.alpha, seed=0).to(device)
torch.manual_seed(args.seed + 10000)
x = torch.randn(512, d, device=device)
with torch.no_grad(): y = teacher(x).argmax(-1)
ckpt = f'results/confirmatory/checkpoints_A1/a{args.alpha}_L{L}_{args.method}_s{args.seed}.pt'
torch.manual_seed(args.seed)
model = StudentNet(d, C, L, args.alpha).to(device)
model.load_state_dict(torch.load(ckpt, map_location=device), strict=True)
model.eval()
is_cifar = False
batch = x.size(0)
Lm = model.num_blocks
dm = d if not is_cifar else 256
# BP gradients
if is_cifar:
h0 = model.embed(x.detach())
else:
h0 = x.detach()
hs = [h0.clone().requires_grad_(True)]
for b in model.blocks: hs.append(hs[-1] + b(hs[-1]))
if is_cifar:
lo = model.out_head(model.out_ln(hs[-1]))
else:
lo = model.out_head(hs[-1])
loss = F.cross_entropy(lo, y)
acc = (lo.argmax(1) == y).float().mean().item()
gs = torch.autograd.grad(loss, hs)
bp = {l: gs[l].detach() for l in range(Lm)}
# DFA Bs (for Gamma)
if is_cifar:
torch.manual_seed(args.seed); _ = ResidualMLP(3072, dm, C, Lm)
else:
torch.manual_seed(args.seed); _ = StudentNet(d, C, Lm, args.alpha)
dfa_Bs = [torch.randn(dm, C, device=device)/np.sqrt(C) for _ in range(Lm)]
with torch.no_grad():
if is_cifar:
logits = model(x)
else:
logits = model(x)
e_T = logits.softmax(-1); e_T[torch.arange(batch), y] -= 1
result = {
'dataset': args.dataset, 'method': args.method, 'seed': args.seed,
'alpha': args.alpha if args.dataset == 'synth' else None,
'depth': L, 'batch': batch, 'loss': loss.item(), 'acc': acc,
'per_layer': []
}
for l in range(Lm):
g = bp[l]
norms = g.norm(dim=-1) # (batch,)
log_norms = torch.log10(norms.clamp(min=1e-30)).cpu().numpy()
# Support fractions
support = {}
for tau in thresholds:
support[str(tau)] = (norms > tau).float().mean().item()
# Element-wise concentration
ninf = g.abs().max(dim=-1).values
n2 = norms.clamp(min=1e-30)
n4 = (g.abs()**4).sum(-1)**(1/4)
n1 = g.abs().sum(-1)
r_inf = (ninf / n2)
pr = (n2**4 / (n4**4).clamp(min=1e-60)) / dm
hoyer = (n1 / (n2 * dm**0.5).clamp(min=1e-30))**2
eff_dim = n1**2 / (n.pow(2).sum(-1) * dm).clamp(min=1e-60) if False else n1**2 / ((g**2).sum(-1) * dm).clamp(min=1e-60)
gsq = g**2; te = gsq.sum(-1, keepdim=True).clamp(min=1e-60)
ssq, _ = gsq.sort(dim=-1, descending=True); cs = ssq.cumsum(-1)
topk = {}
for k in [1, 5, 10, 25]:
idx = max(1, int(dm * k / 100)) - 1
topk[str(k)] = (cs[:, idx:idx+1] / te).squeeze(-1).mean().item()
# Gamma (DFA vs BP) — active subset
a_dfa = (e_T @ dfa_Bs[l].T).detach()
cos_all = F.cosine_similarity(a_dfa, g, dim=-1)
gamma_raw = cos_all.mean().item()
gamma_active = {}; gamma_ew = {}
for tau in thresholds:
mask = norms > tau
gamma_active[str(tau)] = cos_all[mask].mean().item() if mask.sum() > 0 else None
w = norms**2
gamma_ew[str(tau)] = (cos_all * w).sum().item() / (w.sum().item() + 1e-20)
layer_data = {
'layer': l,
'mean_norm': norms.mean().item(),
'median_norm': norms.median().item(),
'max_norm': norms.max().item(),
'min_norm': norms.min().item(),
'support': support,
'log_norms_percentiles': {str(p): float(np.percentile(log_norms, p)) for p in [1,5,10,25,50,75,90,95,99]},
'log_norms_histogram': np.histogram(log_norms, bins=50)[0].tolist(),
'log_norms_bin_edges': np.histogram(log_norms, bins=50)[1].tolist(),
'r_inf_mean': r_inf.mean().item(), 'r_inf_median': r_inf.median().item(),
'pr_mean': pr.mean().item(), 'pr_median': pr.median().item(),
'hoyer_mean': hoyer.mean().item(),
'eff_dim_mean': eff_dim.mean().item(),
'topk_energy': topk,
'gamma_raw': gamma_raw,
'gamma_active': gamma_active,
'gamma_energy_weighted': gamma_ew,
}
result['per_layer'].append(layer_data)
# Summary print
tag = f"{args.dataset}_{args.method}_s{args.seed}"
if args.dataset == 'synth': tag += f"_a{args.alpha}_L{L}"
print(f"[{tag}] acc={acc:.4f} loss={loss.item():.4f}", flush=True)
for ld in result['per_layer']:
l = ld['layer']
print(f" L{l}: norm={ld['mean_norm']:.2e} s(1e-6)={ld['support']['1e-06']:.4f} "
f"r_inf={ld['r_inf_mean']:.4f} PR={ld['pr_mean']:.4f} "
f"top1%={ld['topk_energy']['1']:.4f} Gr={ld['gamma_raw']:.4f}", flush=True)
out = os.path.join(args.output_dir, f'{tag}.json')
with open(out, 'w') as f:
json.dump(result, f, indent=2, default=float)
print(f" -> {out}", flush=True)
if __name__ == '__main__':
main()
|
