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
211
|
"""
Frozen-blocks and shallow baselines for the 4-block d=256 ResidualMLP on CIFAR-10.
This is the codex-round-8 control for ResMLP, parallel to the ViT-Mini frozen-blocks
experiment that walked back the "DFA trains a 4-block ViT" claim.
Conditions (4 per seed):
- BP shallow (num_blocks=0, just embed -> out_ln -> out_head)
- BP frozen-blocks (num_blocks=4, blocks frozen at random init, only embed/LN/head trainable)
- DFA shallow (num_blocks=0)
- DFA frozen-blocks (num_blocks=4, blocks frozen)
If frozen ≈ trainable for DFA: DFA-on-ResMLP also has the same "blocks are passengers"
problem as ViT-Mini, and the strongest remaining DFA performance result in the paper
falls. If frozen << trainable: DFA on ResMLP IS doing meaningful block training, and
the contrast with ViT becomes the most interesting result.
Usage:
CUDA_VISIBLE_DEVICES=2 python experiments/resmlp_frozen_blocks_baseline.py --seed 42
"""
import sys, os, argparse
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.residual_mlp import ResidualMLP
def get_loaders(batch_size=128, dataset='cifar10'):
if dataset == 'cifar100':
mean, std = (0.5071, 0.4867, 0.4408), (0.2675, 0.2565, 0.2761)
DatasetClass = torchvision.datasets.CIFAR100
else:
mean, std = (0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)
DatasetClass = torchvision.datasets.CIFAR10
tv_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean, std),
])
tv = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean, std),
])
tr = DatasetClass('./data', True, download=True, transform=tv_train)
te = DatasetClass('./data', False, download=True, transform=tv)
num_classes = 100 if dataset == 'cifar100' else 10
return (
DataLoader(tr, batch_size=batch_size, shuffle=True, num_workers=2),
DataLoader(te, batch_size=batch_size, shuffle=False, num_workers=2),
), num_classes
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 freeze_blocks(model):
for p in model.blocks.parameters():
p.requires_grad_(False)
def train_bp(model, train_loader, test_loader, dev, epochs, lr, wd, label):
"""Standard BP. Filters optimizer to requires_grad params."""
opt = optim.AdamW(filter(lambda p: p.requires_grad, 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 train_loader:
x = x.view(x.size(0), -1).to(dev); y = y.to(dev)
loss = F.cross_entropy(model(x), y)
opt.zero_grad(); loss.backward(); opt.step()
sch.step()
if ep % 10 == 0 or ep == 1 or ep == epochs:
acc = evaluate(model, test_loader, dev)
print(f" [{label}] ep {ep}: test_acc={acc:.4f}", flush=True)
return model
def train_dfa(model, train_loader, test_loader, dev, epochs, lr, wd, label, num_classes=10):
"""DFA-style: head with true CE, embed (and unfrozen blocks if any) with random feedback.
For frozen-blocks: blocks are skipped. For trainable blocks not used here.
For num_blocks=0 (shallow): only embed/head are updated.
"""
d_hidden = model.d_hidden
L = model.num_blocks
C = num_classes
Bs = [torch.randn(d_hidden, C, device=dev) / np.sqrt(C) for _ in range(max(L, 1))]
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
)
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):
model.train()
for x, y in train_loader:
x = x.view(x.size(0), -1).to(dev); y = y.to(dev)
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
hL_det = hiddens[-1].detach()
# Head update via true CE
logits_out = model.out_head(model.out_ln(hL_det))
head_opt.zero_grad()
F.cross_entropy(logits_out, y).backward()
head_opt.step()
# Embed update via DFA feedback
a0 = (e_T @ Bs[0].T).detach()
rms = (a0 ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
h0 = model.embed(x)
embed_loss = (h0 * (a0 / rms)).sum(-1).mean()
embed_opt.zero_grad()
embed_loss.backward()
embed_opt.step()
sch1.step(); sch2.step()
if ep % 10 == 0 or ep == 1 or ep == epochs:
acc = evaluate(model, test_loader, dev)
print(f" [{label}] ep {ep}: test_acc={acc:.4f}", flush=True)
return model
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--seed', type=int, default=42)
parser.add_argument('--epochs', type=int, default=100)
parser.add_argument('--lr', type=float, default=1e-3)
parser.add_argument('--wd', type=float, default=0.01)
parser.add_argument('--d_hidden', type=int, default=256)
parser.add_argument('--num_blocks', type=int, default=4)
parser.add_argument('--dataset', type=str, default='cifar10', choices=['cifar10', 'cifar100'])
args = parser.parse_args()
dev = torch.device('cuda:0')
print(f"Device: {dev}, seed={args.seed}, epochs={args.epochs}, dataset={args.dataset}", flush=True)
(train_loader, test_loader), C = get_loaders(batch_size=128, dataset=args.dataset)
results = {}
input_dim = 32 * 32 * 3
# Condition 1: BP shallow (num_blocks=0)
print(f"\n=== BP shallow (ResMLP num_blocks=0), seed={args.seed} ===", flush=True)
torch.manual_seed(args.seed); np.random.seed(args.seed); torch.cuda.manual_seed_all(args.seed)
m = ResidualMLP(input_dim, args.d_hidden, C, 0).to(dev)
print(f" n_params: {sum(p.numel() for p in m.parameters())} ({sum(p.numel() for p in m.parameters() if p.requires_grad)} trainable)", flush=True)
train_bp(m, train_loader, test_loader, dev, args.epochs, args.lr, args.wd, 'BP-shallow')
results['bp_shallow'] = evaluate(m, test_loader, dev)
print(f"FINAL BP-shallow: {results['bp_shallow']:.4f}", flush=True)
# Condition 2: BP frozen-blocks (blocks frozen at random init)
L = args.num_blocks
print(f"\n=== BP frozen-blocks (ResMLP num_blocks={L}, blocks frozen), seed={args.seed} ===", flush=True)
torch.manual_seed(args.seed); np.random.seed(args.seed); torch.cuda.manual_seed_all(args.seed)
m = ResidualMLP(input_dim, args.d_hidden, C, L).to(dev)
freeze_blocks(m)
print(f" n_params: {sum(p.numel() for p in m.parameters())} ({sum(p.numel() for p in m.parameters() if p.requires_grad)} trainable)", flush=True)
train_bp(m, train_loader, test_loader, dev, args.epochs, args.lr, args.wd, 'BP-frozen')
results['bp_frozen'] = evaluate(m, test_loader, dev)
print(f"FINAL BP-frozen-blocks: {results['bp_frozen']:.4f}", flush=True)
# Condition 3: DFA shallow (num_blocks=0)
print(f"\n=== DFA shallow (ResMLP num_blocks=0), seed={args.seed} ===", flush=True)
torch.manual_seed(args.seed); np.random.seed(args.seed); torch.cuda.manual_seed_all(args.seed)
m = ResidualMLP(input_dim, args.d_hidden, C, 0).to(dev)
print(f" n_params: {sum(p.numel() for p in m.parameters())} ({sum(p.numel() for p in m.parameters() if p.requires_grad)} trainable)", flush=True)
train_dfa(m, train_loader, test_loader, dev, args.epochs, args.lr, args.wd, 'DFA-shallow', num_classes=C)
results['dfa_shallow'] = evaluate(m, test_loader, dev)
print(f"FINAL DFA-shallow: {results['dfa_shallow']:.4f}", flush=True)
# Condition 4: DFA frozen-blocks (blocks frozen at random init)
print(f"\n=== DFA frozen-blocks (ResMLP num_blocks={L}, blocks frozen), seed={args.seed} ===", flush=True)
torch.manual_seed(args.seed); np.random.seed(args.seed); torch.cuda.manual_seed_all(args.seed)
m = ResidualMLP(input_dim, args.d_hidden, C, L).to(dev)
freeze_blocks(m)
print(f" n_params: {sum(p.numel() for p in m.parameters())} ({sum(p.numel() for p in m.parameters() if p.requires_grad)} trainable)", flush=True)
train_dfa(m, train_loader, test_loader, dev, args.epochs, args.lr, args.wd, 'DFA-frozen', num_classes=C)
results['dfa_frozen'] = evaluate(m, test_loader, dev)
print(f"FINAL DFA-frozen-blocks: {results['dfa_frozen']:.4f}", flush=True)
print(f"\n=== ResMLP frozen/shallow baseline summary, seed={args.seed} ===")
print(f" BP-shallow: {results['bp_shallow']:.4f}")
print(f" BP-frozen: {results['bp_frozen']:.4f}")
print(f" DFA-shallow: {results['dfa_shallow']:.4f}")
print(f" DFA-frozen: {results['dfa_frozen']:.4f}")
print(f"")
print(f"Compare to trainable 4-block ResMLP (3-seed): BP=0.6147 100ep / 0.585 30ep, DFA=0.306 100ep / 0.301 30ep")
print(f"")
print(f"Interpretation:")
print(f" If DFA-frozen ≈ DFA-trainable: blocks are passengers, walk-back parallels ViT")
print(f" If DFA-frozen << DFA-trainable: ResMLP DFA actually trains the blocks (interesting contrast with ViT)")
if __name__ == '__main__':
main()
|
