Viewpoint-Q policy — ADE20K, c64, T=5
Glimpse-placement policy for a frozen CanViT: Q over a discrete grid of candidate viewpoints (2 scales x 16x16 centres); deploy = full-scene glimpse at t0, argmax afterwards. Dueling head (Q = V + mean-zero A); trained with K=1 fractional-CE rewards. Method + code: CanViT-PyTorch-RL.
Provenance: run 20260704-021003_qband_s5, git 007f7173e264d3ccdd6c735234ad246c102ed32c, seed 5, step 4000
(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 | 0.7149 | 42.60 |
| t2 | 0.6875 | 44.10 |
| t3 | 0.6736 | 44.61 |
| t4 | 0.6645 | 45.04 |
from canvit_pytorch_rl.q.net import ViewpointScorer
net = ViewpointScorer.from_pretrained("canvit/qpolicy-ade20k-c64-t5-qband-2026-07-04-s5").eval()
Drive it through canvit_pytorch_rl.q.policy.ArgmaxPolicy + canvit_eval.run_episode for
protocol-exact rollouts; canvit_pytorch_rl.policy.eval is the standalone evaluator.
One of the 8 published seeds of the 2026-07-04 qband (...-qband-2026-07-04-s{0..7}). Performance claims cite the BAND (mean(t1-t4) val CE 0.6853 +- 0.0007 over 8 seeds), never a single seed.