tower-hanoi-resnet18-reward-v2

Tower of Hanoi ResNet Reward Model V2

Input Format

The model expects a Wan-aligned RGB multiview image tensor:

shape: [3, 544, 320]
layout: CHW
view order: cam_high, cam_left_wrist, cam_right_wrist

Preprocessing contract:

Each camera image: 320x180 RGB
Vertical stack: 540x320
Resize: 540x320 -> 544x320 with bilinear interpolation
Final tensor: [3, 544, 320]

The view order and resize are part of the checkpoint contract. Do not use the old horizontal [3, 224, 672] format or a bottom-padded [3, 544, 320] input.

Output Format

The model outputs a scalar binary reward/logit. The evaluation uses:

prob = sigmoid(logit)
pred = prob > 0.5

Interpretation:

prob close to 1: state looks like a successful/completed tower state
prob close to 0: state looks like an incomplete or failed state

Files

full_weights.pt          model weights
README.md                model card
eval_results.json        validation and hard-validation metrics
dataset_metadata.json    processed dataset metadata
manifest.json            file sizes and SHA-256 checksums

Test Results

val.pt:
  count: 1926
  positive: 963
  negative: 963
  loss: 0.0159152
  accuracy: 0.9968848
  auc: 0.9993875
  prob_mean: 0.5008704
  prob_pos_mean: 0.9956364
  prob_neg_mean: 0.0061045

hard_val.pt:
  count: 842
  positive: 421
  negative: 421
  loss: 0.0166353
  accuracy: 0.9964371
  auc: 0.9998984
  prob_mean: 0.5043597
  prob_pos_mean: 0.9975614
  prob_neg_mean: 0.0111578