============================================================
POST-HOC FINE-TUNING Experiment
Job ID: 15261462 | Node: gpub032
Start: Sat Jan  3 09:41:02 CST 2026
============================================================
NVIDIA A40, 46068 MiB
============================================================
================================================================================
POST-HOC LYAPUNOV FINE-TUNING EXPERIMENT
================================================================================
Dataset: cifar100
Depths: [4, 8, 12, 16]
Pretrain: 100 epochs (vanilla, lr=0.001)
Finetune: 50 epochs (Lyapunov, lr=0.0001)
Lyapunov: reg_type=extreme, λ_reg=0.1, threshold=2.0
================================================================================

============================================================
POST-HOC FINE-TUNING: Depth = 4
Pretrain: 100 epochs (vanilla)
Finetune: 50 epochs (Lyapunov, reg_type=extreme)
============================================================
Parameters: 1,756,836

--- Phase 1: Vanilla Pre-training (100 epochs) ---
  Epoch  10: train=0.494 test=0.429
  Epoch  20: train=0.629 test=0.503
  Epoch  30: train=0.707 test=0.521
  Epoch  40: train=0.763 test=0.561
  Epoch  50: train=0.815 test=0.562
  Epoch  60: train=0.851 test=0.567
  Epoch  70: train=0.884 test=0.588
  Epoch  80: train=0.901 test=0.601
  Epoch  90: train=0.912 test=0.601
  Epoch 100: train=0.916 test=0.603
  Best pretrain acc: 0.610

--- Phase 2: Lyapunov Fine-tuning (50 epochs) ---
  reg_type=extreme, lambda_reg=0.1, threshold=2.0
  Epoch 110: train=0.911 test=0.606 λ=1.995
  Epoch 120: train=0.921 test=0.602 λ=1.995
  Epoch 130: train=0.925 test=0.611 λ=1.999
  Epoch 140: train=0.926 test=0.609 λ=1.996
  Epoch 150: train=0.929 test=0.608 λ=1.999
  Best finetune acc: 0.615
  Final λ: 1.999

============================================================
POST-HOC FINE-TUNING: Depth = 8
Pretrain: 100 epochs (vanilla)
Finetune: 50 epochs (Lyapunov, reg_type=extreme)
============================================================
Parameters: 4,892,196

--- Phase 1: Vanilla Pre-training (100 epochs) ---
  Epoch  10: train=0.392 test=0.358
  Epoch  20: train=0.545 test=0.423
  Epoch  30: train=0.642 test=0.483
  Epoch  40: train=0.716 test=0.504
  Epoch  50: train=0.779 test=0.502
  Epoch  60: train=0.830 test=0.529
  Epoch  70: train=0.870 test=0.532
  Epoch  80: train=0.898 test=0.538
  Epoch  90: train=0.911 test=0.540
  Epoch 100: train=0.913 test=0.532
  Best pretrain acc: 0.545

--- Phase 2: Lyapunov Fine-tuning (50 epochs) ---
  reg_type=extreme, lambda_reg=0.1, threshold=2.0
  Epoch 110: train=0.062 test=0.038 λ=2.002
  Epoch 120: train=0.085 test=0.016 λ=1.950
  Epoch 130: train=0.099 test=0.016 λ=1.923
  Epoch 140: train=0.104 test=0.016 λ=1.911
  Epoch 150: train=0.106 test=0.014 λ=1.907
  Best finetune acc: 0.516
  Final λ: 1.907

============================================================
POST-HOC FINE-TUNING: Depth = 12
Pretrain: 100 epochs (vanilla)
Finetune: 50 epochs (Lyapunov, reg_type=extreme)
============================================================
Parameters: 8,027,556

--- Phase 1: Vanilla Pre-training (100 epochs) ---
  Epoch  10: train=0.213 test=0.087
  Epoch  20: train=0.289 test=0.069
  Epoch  30: train=0.346 test=0.083
  Epoch  40: train=0.388 test=0.076
  Epoch  50: train=0.430 test=0.076
  Epoch  60: train=0.467 test=0.087
  Epoch  70: train=0.502 test=0.101