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 20260703-175712_qlr2e4_s0, git 21e25f65632c6fe58f4ff5d8acf18dc90df0e464, seed 0, 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:
| step | val CE | val mIoU % |
|---|---|---|
| t0 | 0.7649 | 39.60 |
| t1 | 0.7143 | 42.51 |
| t2 | 0.6885 | 43.90 |
| t3 | 0.6741 | 44.60 |
| t4 | 0.6657 | 44.81 |
from canvit_pytorch_rl.q.net import ViewpointQNet
net = ViewpointQNet.from_pretrained("canvit/qpolicy-ade20k-c64-t5-2026-07-03").eval()
Drive it through canvit_pytorch_rl.q.policy.GreedyQPolicy + canvit_eval.run_episode for
protocol-exact rollouts; canvit_pytorch_rl.q.eval is the standalone evaluator.
Predates the 2026-07-04 qband; superseded as flagship by canvit/qpolicy-ade20k-c64-t5-qband-2026-07-04-s2 (all 8 band seeds are published).