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
|
"""
Shallow baseline for ViT-Mini: train BP and DFA on a 0-block ViT (just patch_embed
+ cls + pos + out_ln + out_head), to test whether the DFA accuracy on the full
ViT is just exploiting the patch embedder + head.
This is the codex-round-5 control for the "DFA actually trains the transformer
blocks" claim. If shallow DFA acc ≈ 24% (matching the 4-block ViT-Mini DFA acc),
then the blocks are passengers and the claim is too strong. If shallow DFA acc
is much lower, then the blocks are doing real work.
Usage:
CUDA_VISIBLE_DEVICES=2 python experiments/vit_shallow_baseline.py
"""
import sys, os
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
import torchvision
import torchvision.transforms as transforms
import numpy as np
from models.vit_mini import ViTMini
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, y = x.to(dev), y.to(dev)
preds = model(x).argmax(-1)
c += (preds == y).sum().item()
n += x.size(0)
return c / n
def train_bp_shallow(train_loader, test_loader, dev, epochs=30, seed=42, lr=1e-3, wd=0.05):
torch.manual_seed(seed); np.random.seed(seed); torch.cuda.manual_seed_all(seed)
m = ViTMini(num_blocks=0, d_model=128, n_heads=4).to(dev)
print(f"BP-shallow: n_params={sum(p.numel() for p in m.parameters())}", flush=True)
opt = optim.AdamW(m.parameters(), lr=lr, weight_decay=wd)
sch = optim.lr_scheduler.CosineAnnealingLR(opt, T_max=epochs)
for ep in range(1, epochs + 1):
m.train()
for x, y in train_loader:
x = x.to(dev); y = y.to(dev)
loss = F.cross_entropy(m(x), y)
opt.zero_grad(); loss.backward(); opt.step()
sch.step()
if ep % 5 == 0 or ep == 1 or ep == epochs:
acc = evaluate(m, test_loader, dev)
print(f" BP-shallow ep {ep}: test_acc={acc:.4f}", flush=True)
return m
def train_dfa_shallow(train_loader, test_loader, dev, epochs=30, seed=42, lr=1e-3, wd=0.05):
"""0-block ViT trained DFA-style: head with true CE on cls token,
embed (patch_embed + cls + pos) with random feedback `e_T @ B^T` from the head."""
torch.manual_seed(seed); np.random.seed(seed); torch.cuda.manual_seed_all(seed)
m = ViTMini(num_blocks=0, d_model=128, n_heads=4).to(dev)
print(f"DFA-shallow: n_params={sum(p.numel() for p in m.parameters())}", flush=True)
d_model, C = 128, 10
B0 = torch.randn(d_model, C, device=dev) / np.sqrt(C)
embed_opt = optim.AdamW(
list(m.patch_embed.parameters()) + [m.cls_token, m.pos_embed],
lr=lr, weight_decay=wd
)
head_opt = optim.AdamW(
list(m.out_head.parameters()) + list(m.out_ln.parameters()),
lr=lr, weight_decay=wd
)
sch1 = optim.lr_scheduler.CosineAnnealingLR(embed_opt, T_max=epochs)
sch2 = optim.lr_scheduler.CosineAnnealingLR(head_opt, T_max=epochs)
for ep in range(1, epochs + 1):
m.train()
for x, y in train_loader:
x = x.to(dev); y = y.to(dev)
with torch.no_grad():
logits, hi = m(x, return_hidden=True)
e_T = logits.softmax(-1); e_T[torch.arange(x.size(0)), y] -= 1
hL_det = hi[-1].detach()
# Head update via true CE on cls token
h_cls = m.out_ln(hL_det[:, 0])
head_opt.zero_grad()
F.cross_entropy(m.out_head(h_cls), y).backward()
head_opt.step()
# Embed update via DFA-style local loss
a0 = (e_T @ B0.T).detach()
rms = (a0 ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
h0 = m.embed(x) # (B, 65, d_model)
a0_b = a0.unsqueeze(1).expand_as(h0)
embed_loss = (h0 * (a0_b / rms.unsqueeze(1))).sum(-1).mean()
embed_opt.zero_grad()
embed_loss.backward()
embed_opt.step()
sch1.step(); sch2.step()
if ep % 5 == 0 or ep == 1 or ep == epochs:
acc = evaluate(m, test_loader, dev)
print(f" DFA-shallow ep {ep}: test_acc={acc:.4f}", flush=True)
return m
def main():
dev = torch.device('cuda:0')
print(f"Device: {dev}", flush=True)
train_loader, test_loader = get_loaders(batch_size=128)
print("\n=== BP shallow baseline (ViT-Mini num_blocks=0) ===", flush=True)
mb = train_bp_shallow(train_loader, test_loader, dev, epochs=30, seed=42)
bp_acc = evaluate(mb, test_loader, dev)
print(f"FINAL BP-shallow acc: {bp_acc:.4f}", flush=True)
print("\n=== DFA shallow baseline (ViT-Mini num_blocks=0) ===", flush=True)
md = train_dfa_shallow(train_loader, test_loader, dev, epochs=30, seed=42)
dfa_acc = evaluate(md, test_loader, dev)
print(f"FINAL DFA-shallow acc: {dfa_acc:.4f}", flush=True)
print(f"\n=== Summary ===")
print(f"BP-shallow: {bp_acc:.4f} (chance=0.10)")
print(f"DFA-shallow: {dfa_acc:.4f}")
print(f"Compare to ViT-Mini 4-block (3-seed avg): BP=0.792, DFA=0.237")
if __name__ == '__main__':
main()
|
