============================================================ ASYMMETRIC Lyapunov Regularization Job ID: 14632852 | Node: gpub079 Start: Wed Dec 31 10:16:37 CST 2025 ============================================================ NVIDIA A40, 46068 MiB ============================================================ ================================================================================ DEPTH SCALING BENCHMARK ================================================================================ Dataset: cifar100 Depths: [4, 8, 12, 16] Timesteps: 4 Epochs: 150 λ_reg: 0.3, λ_target: -0.1 Reg type: asymmetric, Warmup epochs: 20 Device: cuda ================================================================================ Loading cifar100... Classes: 100, Input: (3, 32, 32) Train: 50000, Test: 10000 Depth configurations: [(4, '4×1'), (8, '4×2'), (12, '4×3'), (16, '4×4')] Regularization type: asymmetric Warmup epochs: 20 ============================================================ Depth = 4 conv layers (4 stages × 1 blocks) ============================================================ Vanilla: depth=4, params=1,756,836 Epoch 10: train=0.498 test=0.438 σ=9.52e-01/3.57e-08 Epoch 20: train=0.631 test=0.506 σ=5.87e-01/2.45e-08 Epoch 30: train=0.706 test=0.562 σ=4.78e-01/1.99e-08 Epoch 40: train=0.756 test=0.554 σ=4.24e-01/1.76e-08 Epoch 50: train=0.800 test=0.583 σ=3.77e-01/1.55e-08 Epoch 60: train=0.833 test=0.593 σ=3.51e-01/1.39e-08 Epoch 70: train=0.863 test=0.591 σ=3.18e-01/1.26e-08 Epoch 80: train=0.883 test=0.596 σ=3.04e-01/1.21e-08 Epoch 90: train=0.907 test=0.601 σ=2.91e-01/1.12e-08 Epoch 100: train=0.922 test=0.615 σ=2.69e-01/1.04e-08 Epoch 110: train=0.933 test=0.614 σ=2.62e-01/9.69e-09 Epoch 120: train=0.942 test=0.615 σ=2.47e-01/9.10e-09 Epoch 130: train=0.950 test=0.620 σ=2.44e-01/8.88e-09 Epoch 140: train=0.951 test=0.623 σ=2.46e-01/8.88e-09 Epoch 150: train=0.952 test=0.618 σ=2.22e-01/8.47e-09 Best test acc: 0.624 Lyapunov: depth=4, params=1,756,836 Epoch 10: train=0.036 test=0.010 λ=1.488 σ=6.35e-01/1.45e-08 Epoch 20: train=0.009 test=0.010 λ=1.425 σ=1.08e-01/3.34e-15 Epoch 30: train=0.009 test=0.010 λ=1.451 σ=4.78e-02/9.92e-21 Epoch 40: train=0.010 test=0.010 λ=1.455 σ=1.57e-02/0.00e+00 Epoch 50: train=0.009 test=0.010 λ=1.476 Epoch 60: train=0.009 test=0.010 λ=1.473 Epoch 70: train=0.010 test=0.010 λ=1.476 σ=1.34e-03/0.00e+00 Epoch 80: train=0.009 test=0.010 λ=1.479 Epoch 90: train=0.009 test=0.010 λ=1.480 Epoch 100: train=0.009 test=0.010 λ=1.480 Epoch 110: train=0.009 test=0.010 λ=1.482 Epoch 120: train=0.009 test=0.010 λ=1.481 Epoch 130: train=0.009 test=0.010 λ=1.481 Epoch 140: train=0.009 test=0.010 λ=1.480 Epoch 150: train=0.010 test=0.010 λ=1.481 Best test acc: 0.089 ============================================================ Depth = 8 conv layers (4 stages × 2 blocks) ============================================================ Vanilla: depth=8, params=4,892,196 Epoch 10: train=0.393 test=0.335 σ=7.83e-01/3.02e-08 Epoch 20: train=0.547 test=0.453 σ=4.71e-01/2.14e-08 Epoch 30: train=0.629 test=0.462 σ=3.73e-01/1.78e-08 Epoch 40: train=0.701 test=0.500 σ=3.33e-01/1.55e-08 Epoch 50: train=0.752 test=0.517 σ=3.06e-01/1.40e-08 Epoch 60: train=0.798 test=0.528 σ=2.76e-01/1.25e-08 Epoch 70: train=0.835 test=0.530 σ=2.70e-01/1.18e-08 Epoch 80: train=0.872 test=0.534 σ=2.51e-01/1.07e-08 Epoch 90: train=0.897 test=0.538 σ=2.38e-01/9.81e-09 Epoch 100: train=0.920 test=0.536 σ=2.26e-01/9.08e-09 Epoch 110: train=0.935 test=0.544 σ=2.06e-01/8.19e-09 Epoch 120: train=0.947 test=0.551 σ=2.11e-01/7.87e-09 Epoch 130: train=0.954 test=0.551 σ=2.02e-01/7.42e-09 Epoch 140: train=0.956 test=0.553 σ=1.92e-01/7.38e-09 Epoch 150: train=0.956 test=0.546 σ=2.05e-01/7.55e-09 Best test acc: 0.554 Lyapunov: depth=8, params=4,892,196 Epoch 10: train=0.031 test=0.010 λ=1.573 σ=4.25e-01/1.02e-08 Epoch 20: train=0.019 test=0.010 λ=1.580 σ=2.80e-01/1.31e-09 Epoch 30: train=0.010 test=0.010 λ=1.519 σ=1.87e-01/8.73e-13 Epoch 40: train=0.009 test=0.010 λ=1.532 σ=7.89e-02/1.22e-26 Epoch 50: train=0.010 test=0.010 λ=1.536 σ=6.60e-02/4.34e-19 Epoch 60: train=0.009 test=0.010 λ=1.542 σ=5.19e-02/0.00e+00 Epoch 70: train=0.010 test=0.010 λ=1.540 σ=2.67e-02/0.00e+00 Epoch 80: train=0.009 test=0.010 λ=1.544 σ=6.28e-03/0.00e+00 Epoch 90: train=0.009 test=0.010 λ=1.542 σ=4.08e-02/0.00e+00 Epoch 100: train=0.010 test=0.010 λ=1.543 Epoch 110: train=0.009 test=0.010 λ=1.543 Epoch 120: train=0.010 test=0.010 λ=1.543 σ=3.14e-03/0.00e+00 Epoch 130: train=0.009 test=0.010 λ=1.543 Epoch 140: train=0.010 test=0.010 λ=1.545 Epoch 150: train=0.010 test=0.010 λ=1.542 Best test acc: 0.020 ============================================================ Depth = 12 conv layers (4 stages × 3 blocks) ============================================================ Vanilla: depth=12, params=8,027,556 Epoch 10: train=0.214 test=0.062 σ=6.26e-01/2.24e-08 Epoch 20: train=0.293 test=0.061 σ=3.27e-01/1.57e-08 Epoch 30: train=0.335 test=0.075 σ=2.66e-01/1.35e-08 Epoch 40: train=0.374 test=0.066 σ=2.32e-01/1.23e-08 Epoch 50: train=0.407 test=0.063 σ=2.29e-01/1.23e-08 Epoch 60: train=0.442 test=0.078 σ=2.18e-01/1.20e-08 Epoch 70: train=0.472 test=0.098 σ=2.30e-01/1.20e-08 Epoch 80: train=0.501 test=0.108 σ=2.21e-01/1.21e-08 Epoch 90: train=0.532 test=0.110 σ=2.24e-01/1.18e-08 Epoch 100: train=0.557 test=0.098 σ=2.25e-01/1.20e-08 Epoch 110: train=0.579 test=0.105 σ=2.26e-01/1.18e-08 Epoch 120: train=0.599 test=0.104 σ=2.31e-01/1.20e-08 Epoch 130: train=0.609 test=0.117 σ=2.28e-01/1.19e-08 Epoch 140: train=0.619 test=0.118 σ=2.23e-01/1.17e-08 Epoch 150: train=0.620 test=0.112 σ=2.30e-01/1.19e-08 Best test acc: 0.122 Lyapunov: depth=12, params=8,027,556 Epoch 10: train=0.017 test=0.010 λ=1.619 σ=3.10e-01/2.44e-12 Epoch 20: train=0.014 test=0.010 λ=1.620 σ=3.48e-01/1.80e-12 Epoch 30: train=0.010 test=0.010 λ=1.551 σ=1.95e-02/1.07e-16 Epoch 40: train=0.010 test=0.010 λ=1.556 σ=2.86e-02/0.00e+00 Epoch 50: train=0.010 test=0.010 λ=1.552 σ=1.02e-01/3.07e-15 Epoch 60: train=0.009 test=0.010 λ=1.560 σ=5.22e-02/0.00e+00 Epoch 70: train=0.009 test=0.010 λ=1.567 σ=3.97e-02/0.00e+00 Epoch 80: train=0.009 test=0.010 λ=1.564 Epoch 90: train=0.009 test=0.010 λ=1.570 σ=1.20e-02/0.00e+00 Epoch 100: train=0.009 test=0.010 λ=1.566 Epoch 110: train=0.009 test=0.010 λ=1.568 σ=1.35e-02/0.00e+00 Epoch 120: train=0.009 test=0.010 λ=1.566 Epoch 130: train=0.009 test=0.010 λ=1.568 Epoch 140: train=0.009 test=0.010 λ=1.566 Epoch 150: train=0.010 test=0.010 λ=1.565 Best test acc: 0.016 ============================================================ Depth = 16 conv layers (4 stages × 4 blocks) ============================================================ Vanilla: depth=16, params=11,162,916 Epoch 10: train=0.091 test=0.011 σ=4.40e-01/1.32e-08 Epoch 20: train=0.133 test=0.015 σ=2.83e-01/1.07e-08 Epoch 30: train=0.156 test=0.018 σ=2.23e-01/9.48e-09 Epoch 40: train=0.176 test=0.018 σ=1.95e-01/9.15e-09 Epoch 50: train=0.191 test=0.021 σ=1.80e-01/8.93e-09 Epoch 60: train=0.204 test=0.022 σ=1.71e-01/8.85e-09 Epoch 70: train=0.218 test=0.028 σ=1.62e-01/9.03e-09 Epoch 80: train=0.227 test=0.025 σ=1.64e-01/8.92e-09 Epoch 90: train=0.238 test=0.028 σ=1.60e-01/9.11e-09 Epoch 100: train=0.249 test=0.028 σ=1.63e-01/9.28e-09 Epoch 110: train=0.257 test=0.031 σ=1.60e-01/9.29e-09 Epoch 120: train=0.265 test=0.027 σ=1.65e-01/9.32e-09 Epoch 130: train=0.270 test=0.028 σ=1.62e-01/9.29e-09 Epoch 140: train=0.273 test=0.026 σ=1.66e-01/9.44e-09 Epoch 150: train=0.274 test=0.026 σ=1.70e-01/9.40e-09 Best test acc: 0.033 Lyapunov: depth=16, params=11,162,916 Epoch 10: train=0.011 test=0.010 λ=1.693 σ=2.40e-01/9.58e-25 Epoch 20: train=0.009 test=0.010 λ=1.575 σ=1.05e-01/7.85e-12 Epoch 30: train=0.010 test=0.010 λ=1.590 σ=2.47e-01/2.27e-12 Epoch 40: train=0.009 test=0.010 λ=1.600 σ=4.94e-02/6.93e-15 Epoch 50: train=0.009 test=0.010 λ=1.594 σ=7.66e-02/2.70e-17 Epoch 60: train=0.010 test=0.010 λ=1.592 σ=5.57e-02/0.00e+00 Epoch 70: train=0.010 test=0.010 λ=1.595 σ=5.96e-02/0.00e+00 Epoch 80: train=0.010 test=0.010 λ=1.592 σ=4.24e-02/0.00e+00 Epoch 90: train=0.010 test=0.010 λ=1.594 σ=3.31e-02/0.00e+00 Epoch 100: train=0.011 test=0.010 λ=1.597 σ=3.36e-02/0.00e+00 Epoch 110: train=0.010 test=0.010 λ=1.599 σ=2.33e-02/0.00e+00 Epoch 120: train=0.010 test=0.010 λ=1.593 Epoch 130: train=0.010 test=0.010 λ=1.595 σ=1.07e-02/0.00e+00 Epoch 140: train=0.010 test=0.010 λ=1.593 σ=6.22e-03/0.00e+00 Epoch 150: train=0.010 test=0.010 λ=1.590 Best test acc: 0.011 ==================================================================================================== DEPTH SCALING RESULTS: CIFAR100 ==================================================================================================== Depth Vanilla Acc Lyapunov Acc Δ Acc Lyap λ Van ∇norm Lyap ∇norm Van κ ---------------------------------------------------------------------------------------------------- 4 0.618 0.010 -0.608 1.481 4.60e-01 8.81e-02 1.2e+09 8 0.546 0.010 -0.536 1.542 3.79e-01 1.58e-01 4.7e+09 12 0.112 0.010 -0.102 1.565 6.41e-01 9.04e-02 3.8e+07 16 0.026 0.010 -0.016 1.590 5.09e-01 2.32e+00 3.0e+07 ==================================================================================================== GRADIENT HEALTH ANALYSIS: Depth 4: ⚠️ Vanilla has ill-conditioned gradients (κ > 1e6) Depth 8: ⚠️ Vanilla has ill-conditioned gradients (κ > 1e6) Depth 12: ⚠️ Vanilla has ill-conditioned gradients (κ > 1e6) Depth 16: ⚠️ Vanilla has ill-conditioned gradients (κ > 1e6) KEY OBSERVATIONS: Vanilla 4→16 layers: -0.592 accuracy change Lyapunov 4→16 layers: +0.000 accuracy change ✓ Lyapunov regularization enables better depth scaling! Results saved to runs/depth_scaling_asymm/cifar100_20260101-112330 ============================================================ Finished: Thu Jan 1 11:23:32 CST 2026 ============================================================