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
|
"""
Train vanilla DFA (no penalty) on the standard 4-block d=256 ResMLP and
save checkpoints at the early epochs (1, 2, 3) BEFORE ‖g_L‖ has
collapsed to the numerical floor.
Codex round 19's #3 priority experiment to disambiguate:
- Hypothesis A: deep-layer alignment was always present in vanilla DFA but
hidden by the post-collapse measurement degeneracy. Penalty just made
the measurement interpretable.
- Hypothesis B: deep-layer alignment was created by the penalty
intervention. Vanilla DFA at any epoch has zero deep alignment.
Test: measure deep-layer cos at vanilla checkpoints from ep 1, 2, 3 (when
‖g_L‖ should still be in the meaningful regime).
Run:
CUDA_VISIBLE_DEVICES=2 python experiments/vanilla_dfa_early_ckpt.py --seed 42
"""
import os
import sys
import argparse
import json
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
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from models.residual_mlp import ResidualMLP
def get_loaders(batch_size=128):
tv_train = transforms.Compose([
transforms.RandomCrop(32, padding=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)),
])
tr = torchvision.datasets.CIFAR10('./data', True, download=True, transform=tv_train)
te = torchvision.datasets.CIFAR10('./data', False, download=True, transform=tv)
return (
DataLoader(tr, batch_size=batch_size, shuffle=True, num_workers=2),
DataLoader(te, batch_size=batch_size, shuffle=False, num_workers=2),
)
def evaluate(model, loader, dev):
model.eval()
n = c = 0
with torch.no_grad():
for x, y in loader:
x = x.view(x.size(0), -1).to(dev); y = y.to(dev)
preds = model(x).argmax(-1)
c += (preds == y).sum().item()
n += x.size(0)
return c / n
def diagnose_norms(model, x_eval, y_eval, dev):
model.eval()
with torch.no_grad():
_, hi = model(x_eval, return_hidden=True)
h_norms = [h.norm(dim=-1).median().item() for h in hi]
h0 = model.embed(x_eval.detach())
hs = [h0.clone().requires_grad_(True)]
for b in model.blocks:
hs.append(hs[-1] + b(hs[-1]))
lo = model.out_head(model.out_ln(hs[-1]))
loss = F.cross_entropy(lo, y_eval)
gs = torch.autograd.grad(loss, hs)
g_norms = [g.norm(dim=-1).median().item() for g in gs]
return h_norms, g_norms
def train_vanilla_dfa(model, train_loader, dev, max_epoch, lr, wd, Bs, x_eval, y_eval, save_at, output_dir, seed):
L = model.num_blocks
block_opts = [optim.AdamW(b.parameters(), lr=lr, weight_decay=wd) for b in model.blocks]
embed_opt = optim.AdamW(model.embed.parameters(), lr=lr, weight_decay=wd)
head_opt = optim.AdamW(
list(model.out_head.parameters()) + list(model.out_ln.parameters()),
lr=lr, weight_decay=wd
)
log = []
h0_norms, g0_norms = diagnose_norms(model, x_eval, y_eval, dev)
log.append({"epoch": 0, "h_norms": h0_norms, "g_norms": g0_norms})
print(f" ep 0: h_norms={[f'{h:.2e}' for h in h0_norms]}, g_norms={[f'{g:.2e}' for g in g0_norms]}", flush=True)
for ep in range(1, max_epoch + 1):
model.train()
for x, y in train_loader:
x = x.view(x.size(0), -1).to(dev); y = y.to(dev)
batch = x.size(0)
with torch.no_grad():
logits, hiddens = model(x, return_hidden=True)
e_T = logits.softmax(-1); e_T[torch.arange(batch), y] -= 1
hL_det = hiddens[-1].detach()
head_opt.zero_grad()
F.cross_entropy(model.out_head(model.out_ln(hL_det)), 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()
block_opts[l].zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.blocks[l].parameters(), 1.0)
block_opts[l].step()
a0 = (e_T @ Bs[0].T).detach()
rms0 = (a0 ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
h0_emb = model.embed(x)
embed_opt.zero_grad()
(h0_emb * (a0 / rms0)).sum(-1).mean().backward()
embed_opt.step()
h_norms, g_norms = diagnose_norms(model, x_eval, y_eval, dev)
log.append({"epoch": ep, "h_norms": h_norms, "g_norms": g_norms})
print(f" ep {ep}: h_norms={[f'{h:.2e}' for h in h_norms]}, g_norms={[f'{g:.2e}' for g in g_norms]}", flush=True)
if ep in save_at:
ckpt_path = os.path.join(output_dir, f"vanilla_dfa_s{seed}_ep{ep}.pt")
torch.save({
"state_dict": model.state_dict(),
"Bs": [b.cpu() for b in Bs],
"epoch": ep,
"h_norms": h_norms,
"g_norms": g_norms,
}, ckpt_path)
print(f" saved {ckpt_path}", flush=True)
return log
def main():
p = argparse.ArgumentParser()
p.add_argument("--seed", type=int, default=42)
p.add_argument("--max_epoch", type=int, default=5)
p.add_argument("--lr", type=float, default=1e-3)
p.add_argument("--wd", type=float, default=0.01)
p.add_argument("--save_at", type=int, nargs="+", default=[1, 2, 3, 4, 5])
p.add_argument("--output_dir", type=str, default="results/vanilla_dfa_early_ckpts")
args = p.parse_args()
os.makedirs(args.output_dir, exist_ok=True)
dev = torch.device("cuda:0")
print(f"Vanilla DFA early-epoch checkpoint sweep: seed={args.seed}, max_epoch={args.max_epoch}", flush=True)
train_loader, test_loader = get_loaders(batch_size=128)
# Eval batch
xs, ys = [], []
for x, y in test_loader:
xs.append(x.view(x.size(0), -1)); ys.append(y)
if sum(xb.size(0) for xb in xs) >= 1024:
break
x_eval = torch.cat(xs)[:1024].to(dev)
y_eval = torch.cat(ys)[:1024].to(dev)
L, d, C = 4, 256, 10
torch.manual_seed(args.seed); np.random.seed(args.seed); torch.cuda.manual_seed_all(args.seed)
m = ResidualMLP(3072, d, C, L).to(dev)
Bs = [torch.randn(d, C, device=dev) / np.sqrt(C) for _ in range(L)]
log = train_vanilla_dfa(m, train_loader, dev, args.max_epoch, args.lr, args.wd, Bs, x_eval, y_eval, args.save_at, args.output_dir, args.seed)
out = {"config": vars(args), "log": log}
out_path = os.path.join(args.output_dir, f"vanilla_dfa_s{args.seed}_log.json")
with open(out_path, "w") as f:
json.dump(out, f, indent=2)
print(f"Saved {out_path}")
if __name__ == "__main__":
main()
|
