DFA + residual-branch penalty test: seed=123, lam=0.01
  ep 0: ||h_L||=8.791e+00 ||g_2||=9.848e-04 acc=0.0957
  ep 1: ||h_L||=8.047e+02 ||g_2||=1.064e-05 eval_acc=0.3320 test_acc=0.3279
  ep 10: ||h_L||=8.563e+03 ||g_2||=1.780e-06 eval_acc=0.3555 test_acc=0.3425
  ep 20: ||h_L||=1.665e+04 ||g_2||=1.229e-06 eval_acc=0.3691 test_acc=0.3600
  ep 30: ||h_L||=2.371e+04 ||g_2||=1.049e-06 eval_acc=0.3525 test_acc=0.3479
  ep 40: ||h_L||=2.960e+04 ||g_2||=9.341e-07 eval_acc=0.3584 test_acc=0.3509
  ep 50: ||h_L||=3.417e+04 ||g_2||=8.874e-07 eval_acc=0.3486 test_acc=0.3436
  ep 60: ||h_L||=3.745e+04 ||g_2||=8.382e-07 eval_acc=0.3477 test_acc=0.3521
  ep 70: ||h_L||=3.959e+04 ||g_2||=8.236e-07 eval_acc=0.3623 test_acc=0.3573
  ep 80: ||h_L||=4.074e+04 ||g_2||=8.102e-07 eval_acc=0.3662 test_acc=0.3579
  ep 90: ||h_L||=4.119e+04 ||g_2||=8.103e-07 eval_acc=0.3613 test_acc=0.3595
  ep 100: ||h_L||=4.126e+04 ||g_2||=8.094e-07 eval_acc=0.3643 test_acc=0.3623

FINAL test acc: 0.3623
Compare to:
  DFA-vanilla (3-seed mean):    0.308
  DFA-shallow (3-seed mean):    0.349
  DFA-frozen (3-seed mean):     0.349
  BP-trainable (3-seed mean):   0.609
Saved results/dfa_residual_penalty/dfa_pen_lam0.01_s123.json