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
|
=== CIFAR-100: d=256 L=4 ===
Start: Sun Apr 26 09:13:20 AM CDT 2026
--- Frozen baseline ---
frozen s=42 (Sun Apr 26 09:13:20 AM CDT 2026)
Device: cuda:0, seed=42, epochs=100, dataset=cifar100
=== BP shallow (ResMLP num_blocks=0), seed=42 ===
n_params: 812900 (812900 trainable)
[BP-shallow] ep 1: test_acc=0.1119
[BP-shallow] ep 10: test_acc=0.1320
[BP-shallow] ep 20: test_acc=0.1286
[BP-shallow] ep 30: test_acc=0.1324
[BP-shallow] ep 40: test_acc=0.1407
[BP-shallow] ep 50: test_acc=0.1555
[BP-shallow] ep 60: test_acc=0.1599
[BP-shallow] ep 70: test_acc=0.1706
[BP-shallow] ep 80: test_acc=0.1738
[BP-shallow] ep 90: test_acc=0.1780
[BP-shallow] ep 100: test_acc=0.1787
FINAL BP-shallow: 0.1787
=== BP frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=42 ===
n_params: 1341284 (812900 trainable)
[BP-frozen] ep 1: test_acc=0.1109
[BP-frozen] ep 10: test_acc=0.1313
[BP-frozen] ep 20: test_acc=0.1252
[BP-frozen] ep 30: test_acc=0.1271
[BP-frozen] ep 40: test_acc=0.1338
[BP-frozen] ep 50: test_acc=0.1557
[BP-frozen] ep 60: test_acc=0.1613
[BP-frozen] ep 70: test_acc=0.1713
[BP-frozen] ep 80: test_acc=0.1751
[BP-frozen] ep 90: test_acc=0.1764
[BP-frozen] ep 100: test_acc=0.1770
FINAL BP-frozen-blocks: 0.1770
=== DFA shallow (ResMLP num_blocks=0), seed=42 ===
n_params: 812900 (812900 trainable)
[DFA-shallow] ep 1: test_acc=0.0914
[DFA-shallow] ep 10: test_acc=0.1120
[DFA-shallow] ep 20: test_acc=0.1130
[DFA-shallow] ep 30: test_acc=0.1198
[DFA-shallow] ep 40: test_acc=0.1170
[DFA-shallow] ep 50: test_acc=0.1211
[DFA-shallow] ep 60: test_acc=0.1248
[DFA-shallow] ep 70: test_acc=0.1203
[DFA-shallow] ep 80: test_acc=0.1248
[DFA-shallow] ep 90: test_acc=0.1254
[DFA-shallow] ep 100: test_acc=0.1255
FINAL DFA-shallow: 0.1255
=== DFA frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=42 ===
n_params: 1341284 (812900 trainable)
[DFA-frozen] ep 1: test_acc=0.0920
[DFA-frozen] ep 10: test_acc=0.1004
[DFA-frozen] ep 20: test_acc=0.1171
[DFA-frozen] ep 30: test_acc=0.1141
[DFA-frozen] ep 40: test_acc=0.1207
[DFA-frozen] ep 50: test_acc=0.1208
[DFA-frozen] ep 60: test_acc=0.1204
[DFA-frozen] ep 70: test_acc=0.1235
[DFA-frozen] ep 80: test_acc=0.1243
[DFA-frozen] ep 90: test_acc=0.1262
[DFA-frozen] ep 100: test_acc=0.1256
FINAL DFA-frozen-blocks: 0.1256
=== ResMLP frozen/shallow baseline summary, seed=42 ===
BP-shallow: 0.1787
BP-frozen: 0.1770
DFA-shallow: 0.1255
DFA-frozen: 0.1256
Compare to trainable 4-block ResMLP (3-seed): BP=0.6147 100ep / 0.585 30ep, DFA=0.306 100ep / 0.301 30ep
Interpretation:
If DFA-frozen ≈ DFA-trainable: blocks are passengers, walk-back parallels ViT
If DFA-frozen << DFA-trainable: ResMLP DFA actually trains the blocks (interesting contrast with ViT)
frozen s=123 (Sun Apr 26 09:54:18 AM CDT 2026)
Device: cuda:0, seed=123, epochs=100, dataset=cifar100
=== BP shallow (ResMLP num_blocks=0), seed=123 ===
n_params: 812900 (812900 trainable)
[BP-shallow] ep 1: test_acc=0.1098
[BP-shallow] ep 10: test_acc=0.1309
[BP-shallow] ep 20: test_acc=0.1203
[BP-shallow] ep 30: test_acc=0.1262
[BP-shallow] ep 40: test_acc=0.1415
[BP-shallow] ep 50: test_acc=0.1532
[BP-shallow] ep 60: test_acc=0.1622
[BP-shallow] ep 70: test_acc=0.1725
[BP-shallow] ep 80: test_acc=0.1751
[BP-shallow] ep 90: test_acc=0.1745
[BP-shallow] ep 100: test_acc=0.1756
FINAL BP-shallow: 0.1756
=== BP frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=123 ===
n_params: 1341284 (812900 trainable)
[BP-frozen] ep 1: test_acc=0.1100
[BP-frozen] ep 10: test_acc=0.1328
[BP-frozen] ep 20: test_acc=0.1256
[BP-frozen] ep 30: test_acc=0.1333
[BP-frozen] ep 40: test_acc=0.1411
[BP-frozen] ep 50: test_acc=0.1596
[BP-frozen] ep 60: test_acc=0.1638
[BP-frozen] ep 70: test_acc=0.1720
[BP-frozen] ep 80: test_acc=0.1737
[BP-frozen] ep 90: test_acc=0.1769
[BP-frozen] ep 100: test_acc=0.1777
FINAL BP-frozen-blocks: 0.1777
=== DFA shallow (ResMLP num_blocks=0), seed=123 ===
n_params: 812900 (812900 trainable)
[DFA-shallow] ep 1: test_acc=0.0928
[DFA-shallow] ep 10: test_acc=0.1025
st=0.2812
[BP] Epoch 30: loss=2.5874, train=0.3376, test=0.3021
[BP] Epoch 40: loss=2.4113, train=0.3699, test=0.3104
[BP] Epoch 50: loss=2.2468, train=0.4084, test=0.3160
[BP] Epoch 60: loss=2.1034, train=0.4373, test=0.3209
[BP] Epoch 70: loss=1.9664, train=0.4711, test=0.3212
[BP] Epoch 80: loss=1.8659, train=0.4913, test=0.3208
[BP] Epoch 90: loss=1.8143, train=0.5098, test=0.3201
[BP] Epoch 100: loss=1.7758, train=0.5158, test=0.3218
Final test acc: 0.3218
--- DFA ---
[DFA] Epoch 1: loss=4.1790, train=0.0644, test=0.0808
[DFA] Epoch 10: loss=4.1013, train=0.0738, test=0.0764
[DFA] Epoch 20: loss=4.0720, train=0.0808, test=0.0803
[DFA] Epoch 30: loss=4.0493, train=0.0865, test=0.0845
[DFA] Epoch 40: loss=4.0403, train=0.0866, test=0.0855
[DFA] Epoch 50: loss=4.0321, train=0.0897, test=0.0852
[DFA] Epoch 60: loss=4.0243, train=0.0921, test=0.0856
[DFA] Epoch 70: loss=4.0213, train=0.0924, test=0.0868
[DFA] Epoch 80: loss=4.0207, train=0.0933, test=0.0867
[DFA] Epoch 90: loss=4.0178, train=0.0948, test=0.0875
[DFA] Epoch 100: loss=4.0181, train=0.0932, test=0.0872
Final test acc: 0.0872
--- FA ---
[FA] Epoch 1: loss=4.1971, train=0.0632, test=0.0708
[FA] Epoch 10: loss=4.0477, train=0.0854, test=0.0847
[FA] Epoch 20: loss=3.9867, train=0.0968, test=0.0997
[FA] Epoch 30: loss=3.9504, train=0.1036, test=0.1037
[FA] Epoch 40: loss=3.9204, train=0.1070, test=0.1068
[FA] Epoch 50: loss=3.8915, train=0.1107, test=0.1091
[FA] Epoch 60: loss=3.8680, train=0.1147, test=0.1135
[FA] Epoch 70: loss=3.8517, train=0.1166, test=0.1156
[FA] Epoch 80: loss=3.8433, train=0.1188, test=0.1182
[FA] Epoch 90: loss=3.8342, train=0.1202, test=0.1215
[FA] Epoch 100: loss=3.8330, train=0.1228, test=0.1208
Final test acc: 0.1208
All results saved to results/cifar100_d256_L4/results_cifar100.json
methods s=456 (Sun Apr 26 09:51:41 AM CDT 2026)
Using device: cuda:0
============================================================
Seed 456
============================================================
--- BP ---
[BP] Epoch 1: loss=3.9722, train=0.0978, test=0.1436
[BP] Epoch 10: loss=3.0679, train=0.2433, test=0.2496
[BP] Epoch 20: loss=2.7902, train=0.2983, test=0.2857
[BP] Epoch 30: loss=2.5920, train=0.3374, test=0.3018
[BP] Epoch 40: loss=2.4046, train=0.3747, test=0.3166
[BP] Epoch 50: loss=2.2421, train=0.4090, test=0.3165
[BP] Epoch 60: loss=2.0908, train=0.4420, test=0.3204
[BP] Epoch 70: loss=1.9548, train=0.4750, test=0.3202
[BP] Epoch 80: loss=1.8580, train=0.4973, test=0.3177
[BP] Epoch 90: loss=1.8029, train=0.5128, test=0.3217
[BP] Epoch 100: loss=1.7769, train=0.5179, test=0.3219
Final test acc: 0.3219
--- DFA ---
[DFA] Epoch 1: loss=4.1619, train=0.0684, test=0.0832
[DFA] Epoch 10: loss=4.0780, train=0.0790, test=0.0777
[DFA] Epoch 20: loss=4.0602, train=0.0848, test=0.0813
[DFA] Epoch 30: loss=4.0430, train=0.0885, test=0.0878
[DFA] Epoch 40: loss=4.0391, train=0.0893, test=0.0872
[DFA] Epoch 50: loss=4.0372, train=0.0914, test=0.0834
[DFA] Epoch 60: loss=4.0358, train=0.0919, test=0.0884
[DFA] Epoch 70: loss=4.0340, train=0.0928, test=0.0906
[DFA] Epoch 80: loss=4.0334, train=0.0926, test=0.0879
[DFA] Epoch 90: loss=4.0325, train=0.0935, test=0.0898
[DFA] Epoch 100: loss=4.0329, train=0.0929, test=0.0894
Final test acc: 0.0894
--- FA ---
[FA] Epoch 1: loss=4.2178, train=0.0611, test=0.0534
[FA] Epoch 10: loss=3.9339, train=0.1008, test=0.0999
[FA] Epoch 20: loss=3.8903, train=0.1079, test=0.1125
[FA] Epoch 30: loss=3.8439, train=0.1169, test=0.1138
[FA] Epoch 40: loss=3.8094, train=0.1220, test=0.1228
[FA] Epoch 50: loss=3.7933, train=0.1252, test=0.1240
[FA] Epoch 60: loss=3.7808, train=0.1273, test=0.1275
[FA] Epoch 70: loss=3.7675, train=0.1281, test=0.1252
[FA] Epoch 80: loss=3.7592, train=0.1312, test=0.1307
[FA] Epoch 90: loss=3.7554, train=0.1333, test=0.1311
[FA] Epoch 100: loss=3.7508, train=0.1319, test=0.1310
Final test acc: 0.1310
All results saved to results/cifar100_d256_L4/results_cifar100.json
=== CIFAR-100 SCAN DONE (Sun Apr 26 10:11:18 AM CDT 2026) ===
|
