Knowing/rlsplat_grpo_23c_det3000

RLSplat — GRPO allocation policy, run sweep_23c_fromrandom_det_only_3000

Best GRPO checkpoint of the 2026-09-10 sweep: the random-allocation pretraining checkpoint (b200_random_up504, 40k steps, 8x B200) finetuned for 3000 optimizer steps with the softmax-proportional allocation head and the GRPO rank surrogate (deterministic main render, no entropy term).

Files

file	what
epoch_0-step_3000.ckpt	full Lightning checkpoint (encoder incl. DVLT backbone, allocation head, optimizer, LR schedule). Load with checkpointing.load=... (resume) or model.encoder.pretrained_weights=... (weights only).
config.yaml / overrides.yaml	the composed Hydra config and the command-line overrides of the run
sweep_23c.env	the sweep queue entry (scripts/sweep_handoff/) that launched it
test_comparison_2026-09-10.txt	mode=test (align_pose=false) numbers below, all 12 cells

Recipe

• Base: +training=rlsplat_softmax_grpo, pretrained weights = random 40k ckpt
• optimizer.lr=1e-5, gaussian_allocation_head_lr_multiplier=100
• train.gaussian_allocation_rank_loss=1.0, gaussian_allocation_entropy_weight=0
• model.encoder.gaussian_allocation_main_render=deterministic, min_gaussians_per_anchor=0, gaussian_allocation_remainder_jitter=true
• 4x B200, accumulate_grad_batches=2, 3000 steps, 24 context views per GPU (mixed RE10K + DL3DV, dynamic view counts)
• Code: RLSplat repo at commit 3360aa6 (2026-09-09)

Test results (mode=test, align_pose=false)

PSNR / LPIPS / SSIM; "uniform 3000" is the ablation control (same base, same steps, uniform allocation instead of the policy); "random 40k" is the base checkpoint under random allocation.

views	budget/view	dataset	random 40k	uniform 3000	GRPO 23c
2	8192	re10k	22.362 / 0.2109 / 0.7460	22.687 / 0.2038 / 0.7582	22.867 / 0.2002 / 0.7640
2	8192	acid	22.772 / 0.2748 / 0.6445	22.827 / 0.2690 / 0.6490	22.983 / 0.2646 / 0.6535
2	1024	re10k	20.832 / 0.3101 / 0.6893	21.222 / 0.2899 / 0.7078	21.791 / 0.2756 / 0.7223
2	1024	acid	21.804 / 0.3939 / 0.6101	21.953 / 0.3779 / 0.6196	22.370 / 0.3658 / 0.6289
8	8192	re10k	24.300 / 0.1478 / 0.8167	24.984 / 0.1399 / 0.8393	25.236 / 0.1371 / 0.8471
8	8192	acid	23.752 / 0.2326 / 0.6813	23.951 / 0.2285 / 0.6937	24.155 / 0.2236 / 0.7003
8	1024	re10k	23.407 / 0.1967 / 0.7867	23.949 / 0.1879 / 0.8088	24.480 / 0.1783 / 0.8218
8	1024	acid	23.188 / 0.2976 / 0.6577	23.263 / 0.2977 / 0.6685	23.683 / 0.2861 / 0.6785
24	8192	re10k	25.064 / 0.1287 / 0.8408	25.747 / 0.1218 / 0.8618	26.101 / 0.1191 / 0.8717
24	8192	acid	24.492 / 0.2098 / 0.7214	24.722 / 0.2062 / 0.7330	24.982 / 0.2018 / 0.7418
24	1024	re10k	24.585 / 0.1515 / 0.8258	25.043 / 0.1510 / 0.8433	25.560 / 0.1445 / 0.8547
24	1024	acid	24.175 / 0.2403 / 0.7051	24.104 / 0.2515 / 0.7116	24.569 / 0.2404 / 0.7227

GRPO > uniform > random in every cell. Policy gain over the uniform control: +0.42..+0.57 dB at 1024/view, +0.16..+0.35 dB at 8192/view (largest on ACID).

Validation yardstick (in-run, 24 RE10K scenes, 2-view, budgets 1024..16384 x7)

Last round (step 3000) LPIPS per budget: 0.2632 / 0.2116 / 0.1956 / 0.1917 / 0.1909 / 0.1907 / 0.1910 (AUC 0.1985).