diff options
Diffstat (limited to 'results/cifar100_scan.log')
| -rw-r--r-- | results/cifar100_scan.log | 179 |
1 files changed, 176 insertions, 3 deletions
diff --git a/results/cifar100_scan.log b/results/cifar100_scan.log index 637893f..15d1d46 100644 --- a/results/cifar100_scan.log +++ b/results/cifar100_scan.log @@ -113,7 +113,180 @@ FINAL BP-frozen-blocks: 0.1777 n_params: 812900 (812900 trainable) [DFA-shallow] ep 1: test_acc=0.0928 [DFA-shallow] ep 10: test_acc=0.1025 -st=0.2812 + [DFA-shallow] ep 20: test_acc=0.1146 + [DFA-shallow] ep 30: test_acc=0.1180 + [DFA-shallow] ep 40: test_acc=0.1239 + [DFA-shallow] ep 50: test_acc=0.1283 + [DFA-shallow] ep 60: test_acc=0.1204 + [DFA-shallow] ep 70: test_acc=0.1237 + [DFA-shallow] ep 80: test_acc=0.1261 + [DFA-shallow] ep 90: test_acc=0.1261 + [DFA-shallow] ep 100: test_acc=0.1269 +FINAL DFA-shallow: 0.1269 + +=== DFA frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=123 === + n_params: 1341284 (812900 trainable) + [DFA-frozen] ep 1: test_acc=0.0916 + [DFA-frozen] ep 10: test_acc=0.1060 + [DFA-frozen] ep 20: test_acc=0.1167 + [DFA-frozen] ep 30: test_acc=0.1125 + [DFA-frozen] ep 40: test_acc=0.1153 + [DFA-frozen] ep 50: test_acc=0.1237 + [DFA-frozen] ep 60: test_acc=0.1217 + [DFA-frozen] ep 70: test_acc=0.1254 + [DFA-frozen] ep 80: test_acc=0.1239 + [DFA-frozen] ep 90: test_acc=0.1254 + [DFA-frozen] ep 100: test_acc=0.1257 +FINAL DFA-frozen-blocks: 0.1257 + +=== ResMLP frozen/shallow baseline summary, seed=123 === + BP-shallow: 0.1756 + BP-frozen: 0.1777 + DFA-shallow: 0.1269 + DFA-frozen: 0.1257 + +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=456 (Sun Apr 26 10:35:07 AM CDT 2026) +Device: cuda:0, seed=456, epochs=100, dataset=cifar100 + +=== BP shallow (ResMLP num_blocks=0), seed=456 === + n_params: 812900 (812900 trainable) + [BP-shallow] ep 1: test_acc=0.1073 + [BP-shallow] ep 10: test_acc=0.1327 + [BP-shallow] ep 20: test_acc=0.1250 + [BP-shallow] ep 30: test_acc=0.1303 + [BP-shallow] ep 40: test_acc=0.1411 + [BP-shallow] ep 50: test_acc=0.1529 + [BP-shallow] ep 60: test_acc=0.1651 + [BP-shallow] ep 70: test_acc=0.1724 + [BP-shallow] ep 80: test_acc=0.1743 + [BP-shallow] ep 90: test_acc=0.1757 + [BP-shallow] ep 100: test_acc=0.1776 +FINAL BP-shallow: 0.1776 + +=== BP frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=456 === + n_params: 1341284 (812900 trainable) + [BP-frozen] ep 1: test_acc=0.1073 + [BP-frozen] ep 10: test_acc=0.1326 + [BP-frozen] ep 20: test_acc=0.1226 + [BP-frozen] ep 30: test_acc=0.1276 + [BP-frozen] ep 40: test_acc=0.1495 + [BP-frozen] ep 50: test_acc=0.1535 + [BP-frozen] ep 60: test_acc=0.1645 + [BP-frozen] ep 70: test_acc=0.1685 + [BP-frozen] ep 80: test_acc=0.1773 + [BP-frozen] ep 90: test_acc=0.1777 + [BP-frozen] ep 100: test_acc=0.1794 +FINAL BP-frozen-blocks: 0.1794 + +=== DFA shallow (ResMLP num_blocks=0), seed=456 === + n_params: 812900 (812900 trainable) + [DFA-shallow] ep 1: test_acc=0.0913 + [DFA-shallow] ep 10: test_acc=0.1150 + [DFA-shallow] ep 20: test_acc=0.1153 + [DFA-shallow] ep 30: test_acc=0.1161 + [DFA-shallow] ep 40: test_acc=0.1140 + [DFA-shallow] ep 50: test_acc=0.1202 + [DFA-shallow] ep 60: test_acc=0.1229 + [DFA-shallow] ep 70: test_acc=0.1243 + [DFA-shallow] ep 80: test_acc=0.1240 + [DFA-shallow] ep 90: test_acc=0.1237 + [DFA-shallow] ep 100: test_acc=0.1235 +FINAL DFA-shallow: 0.1235 + +=== DFA frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=456 === + n_params: 1341284 (812900 trainable) + [DFA-frozen] ep 1: test_acc=0.0862 + [DFA-frozen] ep 10: test_acc=0.1059 + [DFA-frozen] ep 20: test_acc=0.1130 + [DFA-frozen] ep 30: test_acc=0.1200 + [DFA-frozen] ep 40: test_acc=0.1227 + [DFA-frozen] ep 50: test_acc=0.1177 + [DFA-frozen] ep 60: test_acc=0.1229 + [DFA-frozen] ep 70: test_acc=0.1232 + [DFA-frozen] ep 80: test_acc=0.1240 + [DFA-frozen] ep 90: test_acc=0.1237 + [DFA-frozen] ep 100: test_acc=0.1236 +FINAL DFA-frozen-blocks: 0.1236 + +=== ResMLP frozen/shallow baseline summary, seed=456 === + BP-shallow: 0.1776 + BP-frozen: 0.1794 + DFA-shallow: 0.1235 + DFA-frozen: 0.1236 + +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) + +--- FA + DFA + BP --- + s=42 (Sun Apr 26 11:16:00 AM CDT 2026) +Using device: cuda:0 + +============================================================ +Seed 42 +============================================================ + +--- BP --- + [BP] Epoch 1: loss=3.9752, train=0.0983, test=0.1432 + [BP] Epoch 10: loss=3.0848, train=0.2424, test=0.2492 + [BP] Epoch 20: loss=2.8041, train=0.2935, test=0.2883 + [BP] Epoch 30: loss=2.6019, train=0.3352, test=0.3078 + [BP] Epoch 40: loss=2.4193, train=0.3727, test=0.3158 + [BP] Epoch 50: loss=2.2631, train=0.4053, test=0.3160 + [BP] Epoch 60: loss=2.1134, train=0.4371, test=0.3223 + [BP] Epoch 70: loss=1.9686, train=0.4729, test=0.3207 + [BP] Epoch 80: loss=1.8724, train=0.4941, test=0.3197 + [BP] Epoch 90: loss=1.8161, train=0.5069, test=0.3197 + [BP] Epoch 100: loss=1.7897, train=0.5126, test=0.3192 + Final test acc: 0.3192 + +--- DFA --- + [DFA] Epoch 1: loss=4.1736, train=0.0679, test=0.0775 + [DFA] Epoch 10: loss=4.0844, train=0.0798, test=0.0819 + [DFA] Epoch 20: loss=4.0627, train=0.0840, test=0.0759 + [DFA] Epoch 30: loss=4.0466, train=0.0876, test=0.0840 + [DFA] Epoch 40: loss=4.0357, train=0.0896, test=0.0862 + [DFA] Epoch 50: loss=4.0347, train=0.0909, test=0.0879 + [DFA] Epoch 60: loss=4.0298, train=0.0933, test=0.0879 + [DFA] Epoch 70: loss=4.0244, train=0.0958, test=0.0883 + [DFA] Epoch 80: loss=4.0232, train=0.0939, test=0.0871 + [DFA] Epoch 90: loss=4.0219, train=0.0962, test=0.0870 + [DFA] Epoch 100: loss=4.0244, train=0.0949, test=0.0875 + Final test acc: 0.0875 + +--- FA --- + [FA] Epoch 1: loss=4.1842, train=0.0639, test=0.0598 + [FA] Epoch 10: loss=3.9551, train=0.0978, test=0.0949 + [FA] Epoch 20: loss=3.8745, train=0.1103, test=0.1101 + [FA] Epoch 30: loss=3.8457, train=0.1160, test=0.1212 + [FA] Epoch 40: loss=3.7975, train=0.1235, test=0.1247 + [FA] Epoch 50: loss=3.7623, train=0.1290, test=0.1332 + [FA] Epoch 60: loss=3.7338, train=0.1341, test=0.1397 + [FA] Epoch 70: loss=3.7109, train=0.1404, test=0.1400 + [FA] Epoch 80: loss=3.6910, train=0.1426, test=0.1457 + [FA] Epoch 90: loss=3.6844, train=0.1436, test=0.1455 + [FA] Epoch 100: loss=3.6859, train=0.1440, test=0.1464 + Final test acc: 0.1464 + +All results saved to results/cifar100_d256_L4/results_cifar100.json + s=123 (Sun Apr 26 11:35:15 AM CDT 2026) +Using device: cuda:0 + +============================================================ +Seed 123 +============================================================ + +--- BP --- + [BP] Epoch 1: loss=3.9679, train=0.0986, test=0.1439 + [BP] Epoch 10: loss=3.0754, train=0.2440, test=0.2501 + [BP] Epoch 20: loss=2.8025, train=0.2938, test=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 @@ -153,7 +326,7 @@ st=0.2812 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) + s=456 (Sun Apr 26 11:55:24 AM CDT 2026) Using device: cuda:0 ============================================================ @@ -204,4 +377,4 @@ Seed 456 All results saved to results/cifar100_d256_L4/results_cifar100.json -=== CIFAR-100 SCAN DONE (Sun Apr 26 10:11:18 AM CDT 2026) === +=== CIFAR-100 DONE (Sun Apr 26 12:15:42 PM CDT 2026) === |