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
|
"""
Minimal worked example showing how a future FA paper author would use the
diagnostic protocol on their own model. Trains a fresh tiny ResMLP with DFA
on CIFAR-10 for 5 epochs (so the script runs in <2 minutes on CPU), applies
the protocol, and prints the verdict.
This is the "API tutorial" version of the protocol. Real applications would
train for 100 epochs and use a real test set; the structure is identical.
Run:
python -m protocol.examples.minimal_worked_example
"""
import os
import sys
import time
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
REPO_ROOT = os.path.dirname(
os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
)
sys.path.insert(0, REPO_ROOT)
from models.residual_mlp import ResidualMLP # noqa: E402
from protocol import diagnose # noqa: E402
def get_loaders(batch_size=128):
tv = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)),
])
tr = torchvision.datasets.CIFAR10("./data", train=True, download=True, transform=tv)
te = torchvision.datasets.CIFAR10("./data", train=False, download=True, transform=tv)
return (
DataLoader(tr, batch_size=batch_size, shuffle=True, num_workers=0),
DataLoader(te, batch_size=batch_size, shuffle=False, num_workers=0),
)
def evaluate(model, loader, device):
model.eval()
correct = total = 0
with torch.no_grad():
for x, y in loader:
x = x.view(x.size(0), -1).to(device)
y = y.to(device)
preds = model(x).argmax(-1)
correct += (preds == y).sum().item()
total += x.size(0)
return correct / total
def train_dfa_one_epoch(model, train_loader, Bs, device, lr=1e-3):
L = model.num_blocks
opts = [optim.AdamW(b.parameters(), lr=lr) for b in model.blocks]
embed_opt = optim.AdamW(model.embed.parameters(), lr=lr)
head_opt = optim.AdamW(
list(model.out_head.parameters()) + list(model.out_ln.parameters()), lr=lr
)
model.train()
for x, y in train_loader:
x = x.view(x.size(0), -1).to(device); y = y.to(device)
with torch.no_grad():
logits, hiddens = model(x, return_hidden=True)
e_T = logits.softmax(-1); e_T[torch.arange(x.size(0)), y] -= 1
head_opt.zero_grad()
F.cross_entropy(model.out_head(model.out_ln(hiddens[-1].detach())), y).backward()
head_opt.step()
for l in range(L):
h_l = hiddens[l].detach()
a = (e_T @ Bs[l].T).detach()
rms = (a ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
f = model.blocks[l](h_l)
loss = (f * (a / rms)).sum(-1).mean()
opts[l].zero_grad(); loss.backward(); opts[l].step()
a0 = (e_T @ Bs[0].T).detach()
rms0 = (a0 ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
h0 = model.embed(x)
embed_opt.zero_grad()
(h0 * (a0 / rms0)).sum(-1).mean().backward()
embed_opt.step()
def main():
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Device: {device}")
print()
print("Step 1: train a tiny 4-block d=128 ResMLP with DFA on CIFAR-10")
print("(5 epochs, just for the example — real runs would use ~100 epochs)")
print()
train_loader, test_loader = get_loaders(batch_size=128)
torch.manual_seed(42); np.random.seed(42)
model = ResidualMLP(input_dim=3072, d_hidden=128, num_classes=10, num_blocks=4).to(device)
Bs = [torch.randn(128, 10, device=device) / np.sqrt(10) for _ in range(4)]
t0 = time.time()
for ep in range(1, 6):
train_dfa_one_epoch(model, train_loader, Bs, device)
acc = evaluate(model, test_loader, device)
print(f" epoch {ep}: test_acc = {acc:.4f} ({time.time()-t0:.0f}s elapsed)")
print()
print("Step 2: build the eval batches the protocol needs (8 batches × 128 samples)")
eval_batches = []
tv = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)),
])
te = torchvision.datasets.CIFAR10("./data", train=False, download=True, transform=tv)
sub_loader = DataLoader(te, batch_size=128, shuffle=False, num_workers=0)
for x, y in sub_loader:
x = x.view(x.size(0), -1).to(device); y = y.to(device)
eval_batches.append((x, y))
if len(eval_batches) >= 8:
break
print()
print("Step 3: apply the protocol")
print()
final_acc = evaluate(model, test_loader, device)
report = diagnose(
model=model,
eval_batches=eval_batches,
headline_acc=final_acc,
# In a real paper, you'd train an architecture-matched random-blocks
# baseline and pass its accuracy here. For this example we use the
# 3-seed mean from our paper (4-block d=256 ResMLP DFA-shallow).
# The width is different (d=128 vs d=256) but the diagnostic
# interpretation is the same.
frozen_baseline_acc=0.349,
method_name="DFA (5-epoch demo)",
notes="4-block d=128 ResMLP, CIFAR-10, seed 42, 5 epochs",
)
print(report)
print()
print("Step 4: interpret")
print()
if report.verdict == "trustworthy":
print(" The protocol gave a trustworthy verdict, meaning the network is")
print(" in the meaningful measurement regime. You can report headline")
print(" accuracy and Γ alignment confidently.")
else:
print(" The protocol flagged this run for walk-back. Specifically:")
print(f" {report.verdict}")
print()
print(" In a real paper using this protocol, you would either:")
print(" (1) Walk back the headline claim and report the failure mode, OR")
print(" (2) Modify the training (e.g., add a residual-stream penalty)")
print(" to bring the diagnostics into the healthy regime, then re-")
print(" apply the protocol to the modified network.")
if __name__ == "__main__":
main()
|
