============================================================
Teacher-Mark Binary Classifier Training
============================================================
  Data dir:    outlook/data_correction
  Device:      cuda
  Epochs:      100
  Batch size:  1
  LR:          0.001
  Latent dim:  64
  MP steps:    3
  Checkpoint:  outlook/ckpt/correction.pt
  Log:         /public/home/dxw/Codes/afem/outlook/ckpt/correction_train_log.json

Loading dataset...
  Found 1998 files across 100 samples
  Train samples: 80  Val samples: 20
  Train: 1598 graphs, 2203757 nodes, 65355 positive (3.0%)
  Val:   400 graphs, 513740 nodes, 15208 positive (3.0%)
  Loaded in 5.7s

Training...
  Epoch   0 | train=0.9876 val=0.9234 | auc=0.8342 topk=0.216 phys=0.120 lr=1.00e-03
  Epoch  10 | train=0.7189 val=0.7505 | auc=0.8856 topk=0.307 phys=0.120 lr=1.00e-03
  Epoch  20 | train=0.6477 val=0.7192 | auc=0.8954 topk=0.312 phys=0.120 lr=1.00e-03
  Epoch  30 | train=0.6108 val=0.7108 | auc=0.9025 topk=0.342 phys=0.120 lr=1.00e-03
  Epoch  40 | train=0.5436 val=0.7289 | auc=0.9107 topk=0.357 phys=0.120 lr=5.00e-04