Viewpoint policy-gradient actor — ADE20K, c64, T=5
Glimpse-placement policy for a frozen CanViT, over a discrete grid of candidate viewpoints (2 scales x 16x16 centres): scores per candidate are policy logits trained on-policy by score-function credit (loss -(z * log pi) - entropy_bonus * H, K=1 fractional-CE rewards); deploy = full-scene glimpse at t0, argmax afterwards — the same deploy rule as the Q-regression policies. Method + code: CanViT-PyTorch-RL.
Provenance: run 20260707-144704_pgfloor_s5, git d63026b8589a6fc035f724fd491b3049837b5681, seed 5, step 6000
(deploy ckpt = best mean(t1-t4) validation CE over the run's evals). Backbone canvit/canvitb16-add-vpe-pretrain-g128px-s512px-in21k-dv3b16-2026-02-02;
probe resolved for canvas grid 64. Full-val metrics at this step:
| glimpse step t | val CE | val mIoU % |
|---|---|---|
| t0 | 0.7649 | 39.60 |
| t1 | 0.7161 | 42.59 |
| t2 | 0.6886 | 43.73 |
| t3 | 0.6758 | 44.21 |
| t4 | 0.6661 | 44.83 |
from canvit_pytorch_rl.policy.net import ViewpointScorer
net = ViewpointScorer.from_pretrained("canvit/pgpolicy-ade20k-c64-t5-floorband-2026-07-07-s5").eval()
Drive it through canvit_pytorch_rl.policy.deploy.ArgmaxPolicy + canvit_eval.run_episode for
protocol-exact rollouts; canvit_pytorch_rl.policy.eval is the standalone evaluator.
Seed 5 of the 8-seed pg entropy-floor band (floor 1.0 nat), band objective 0.6865±0.0012 vs qband 0.6853±0.0007; see docs/sessions/2026-07-05-e2e-diff-single-image-probe.md.