09_voxel_update_f1_harm: voxel update head f1 (do no harm) 6k, head of the TSDF-decided best arm (orbit benchmark)

09_voxel_update_f1_harm/README.md

@@ -0,0 +1,57 @@
+# Voxel update head f1 (do no harm), step 6000: the head of the best orbit-benchmark arm
+
+`head_006000.pt` - the trained voxel update head (outputs/voxupd4/f1_harm/head_006000.pt) behind the **TSDF-decided write**,
+the best arm on the DL3DV orbit (revisit) benchmark as of 2026-10-30. The file holds the voxel head only (7.9 M parameters);
+the trunk, proposal head and surface decoder come from the clean voxel arm in `07_clean_voxel_f3r/`.
+
+- **Needs, besides this file** (run from the repo root):
+  - `07_clean_voxel_f3r/ema_070000.pt` at `outputs/clean_voxel_f3r/ema_070000.pt` (the base path this checkpoint records);
+  - the decoder construction files `free_completion_20260927/checkpoints/ray/step002000` from `03_checkpoints_streaming.zip`,
+    at `outputs/clean/free_completion_20260927/checkpoints/ray/step002000` (the path F3 records; the zip restores it under
+    `outputs/free_completion_20260927/`, so move or link it). F3's weights are loaded over it;
+  - the 3D-Belief prior `3d_belief_re10k_256.pt` at `$FLC_BELIEF_CKPT`, read when the trunk is built (F3's weights are loaded
+    over it). Not in this repo;
+  - the orbit caches `data/orbit_cache/dl3dv_val33` / `dl3dv_test33` (DL3DV orbits with VGGT-Omega geometry,
+    `scripts/data/build_orbit_cache.py`). Not in this repo.
+- **Training:** `flc/unified/train_voxel_update.py`, 6000 steps, one B200, about 1.6 h, from F3 70k with everything but the
+  voxel head frozen. Data: 400 DL3DV training orbits (`data/orbit_cache/dl3dv_train`), rollouts over each stream with the write
+  in the loop: A4 write (association along the ray 8 voxels, evidence weights, depth test 8, prune .01), evidence / age input.
+  Loss: L1 + LPIPS + 0.1 depth at window frames and at 5 other frames that see the window's region (up to 150 degrees away,
+  weight .7), plus the do-no-harm term (weight 3): a hinged penalty on any rise of the other frames' 16 px patch L1 / frame
+  LPIPS over the memory before the update. Full command: command.json.
+- **Code:** branch amodal-completion. Trained on 2026-10-29 from the working tree; the code is in commit 9edc5dd (2026-10-30),
+  whose later additions to the trainer are off by default (not re-run at that commit).
+- **Load** as an `eval_long_stream.py --snapshot` (dict with 'voxel_update', 'base', 'voxel_args', 'ema_head', 'args', 'step';
+  `flc.unified.train_root_ae.load_teacher`).
+
+## The best arm: TSDF-decided write
+
+Voxel memory written by this head. A wide-band TSDF (band 15 voxels and at least the merge range) is kept beside it, fed
+each window's own prediction and the observed colours. A window may change the memory only in its holes and in 16 px patches
+where the memory disagrees with the TSDF (log-depth > .4 on most covered pixels, or patch-mean colour > .3), about 21-24 % of
+its pixels; elsewhere the memory is kept as stored.
+
+    python scripts/unified/eval_long_stream.py --snapshot 09_voxel_update_f1_harm/head_006000.pt \
+      --data data/orbit_cache/dl3dv_test33 --clips 33 --loops 2 --async-period 0 --diag-drift --arm none --mem-rule voxel \
+      --vox-prune .01 --vox-assoc 8 --vox-wcap 1e9 --vox-evmax --vox-gate 8 \
+      --vox-patch 0 --vox-patch-ref tsdf --vox-patch-hole pixel --vox-patch-dd .4 --vox-patch-dc .3 \
+      --compose-tsdf 15 --compose-3d --out <dir>
+
+(from commit 9edc5dd, `--preset tsdf_decided` sets the write flags and `--compose-tsdf 15`.)
+
+dl3dv_test33 (33 DL3DV test orbits), the stream walked twice (6 orbits), the first orbit's held-out frames rendered after every
+window, all pixels (holes count). One seed, B200. Runs are not bit-reproducible (GPU nondeterminism flips borderline patch
+decisions, up to 0.09 dB per window on one scene).
+
+| test33, 6 orbits, all pixels | PSNR orbit 1 / 3 / 6 | change at 6 | LPIPS at 6 |
+|---|---|---|---|
+| headline A: memory + the TSDF's low frequencies (box-15 blur of TSDF - memory, TSDF in holes; `track_*_all_c15`) | 15.25 / 15.61 / 15.84 | +0.59 | .561 |
+| headline B: 3D recolouring (each root's Gaussians shifted by the mean TSDF - Gaussian colour; `track_*_all_c3r`) | 16.01 / 15.96 / 15.98 | -0.03 | .552 |
+| the memory alone (`track_*_all`) | 15.71 / 15.58 / 15.60 | -0.11 | .554 |
+| wide-band TSDF (baseline, same window predictions) | 14.98 / 15.36 / 15.56 | +0.58 | .622 |
+| this head, every write re-fuses (no TSDF decision) | 15.87 / 15.22 / 14.90 | -0.97 | .578 |
+
+Curves after every window: `best_arm_test33_6orbits.png` (numbers in `best_arm_test33_6orbits.json`). Which headline to report
+is still open.
+
+sha256 head_006000.pt: af728224d497462e686ae8275d6bc051a6adb7330e636808522a1ba100b38969

09_voxel_update_f1_harm/best_arm_test33_6orbits.json

@@ -0,0 +1,72 @@
+{
+ "Headline A: memory + TSDF lows (c15)": {
+  "psnr": {
+   "n": 33,
+   "start": 11.810133934020996,
+   "mid": 15.608428868380459,
+   "end": 15.839989719968854
+  },
+  "lpips": {
+   "n": 33,
+   "start": 0.5832957029342651,
+   "mid": 0.563083928642851,
+   "end": 0.5604310893651211
+  }
+ },
+ "Headline B: 3D recolouring (c3r)": {
+  "psnr": {
+   "n": 33,
+   "start": 11.464308420817057,
+   "mid": 15.959378589283336,
+   "end": 15.981326305505002
+  },
+  "lpips": {
+   "n": 33,
+   "start": 0.5789229273796082,
+   "mid": 0.5521636569138729,
+   "end": 0.5524708339662263
+  }
+ },
+ "TSDF-decided memory alone": {
+  "psnr": {
+   "n": 33,
+   "start": 11.507584571838379,
+   "mid": 15.584838838288277,
+   "end": 15.585186756018436
+  },
+  "lpips": {
+   "n": 33,
+   "start": 0.5779265562693278,
+   "mid": 0.553575286359498,
+   "end": 0.5541323242765485
+  }
+ },
+ "Wide-band TSDF (baseline)": {
+  "psnr": {
+   "n": 33,
+   "start": 10.84955883026123,
+   "mid": 15.356085892879602,
+   "end": 15.56182676373106
+  },
+  "lpips": {
+   "n": 33,
+   "start": 0.6679924329121908,
+   "mid": 0.6317940372409243,
+   "end": 0.6224375312978571
+  }
+ },
+ "f1 head, every write re-fuses": {
+  "psnr": {
+   "n": 33,
+   "start": 11.507584571838379,
+   "mid": 15.215329430320047,
+   "end": 14.902457555135092
+  },
+  "lpips": {
+   "n": 33,
+   "start": 0.5779262781143188,
+   "mid": 0.5684382572318568,
+   "end": 0.5781347408439174
+  }
+ }
+}

09_voxel_update_f1_harm/command.json

@@ -0,0 +1,75 @@
+{
+ "argv": [
+  "flc/unified/train_voxel_update.py",
+  "--out",
+  "outputs/voxupd4/f1_harm",
+  "--snapshot",
+  "outputs/clean_voxel_f3r/ema_070000.pt",
+  "--data",
+  "data/orbit_cache/dl3dv_train",
+  "--steps",
+  "6000",
+  "--save-every",
+  "1500",
+  "--evidence",
+  "--prune",
+  ".01",
+  "--max-ang",
+  "150",
+  "--see-region",
+  "--assoc",
+  "8",
+  "--evw",
+  "--gate",
+  "8",
+  "--w-harm",
+  "3"
+ ],
+ "args": {
+  "out": "outputs/voxupd4/f1_harm",
+  "snapshot": "outputs/clean_voxel_f3r/ema_070000.pt",
+  "data": "data/orbit_cache/dl3dv_train",
+  "overfit": "",
+  "exclude": "",
+  "steps": "6000",
+  "rollouts": "2",
+  "lr": "5e-05",
+  "warmup": "100",
+  "ema": "0.999",
+  "n_win": "3",
+  "n_other": "5",
+  "far": "12",
+  "max_ang": "150.0",
+  "see_region": "True",
+  "w_other": "0.7",
+  "lpips_w": "1.0",
+  "w_harm": "3.0",
+  "harm_patch": "16",
+  "w_depth": "0.1",
+  "dep_clamp": "0.3",
+  "hold": "False",
+  "wcap": "0.0",
+  "evidence": "True",
+  "gate": "8.0",
+  "band": "False",
+  "evw": "True",
+  "assoc": "8.0",
+  "assoc_lat": "1.0",
+  "accum": "",
+  "accum_range": "8.0",
+  "angle": "False",
+  "train_trunk": "False",
+  "lr_trunk": "2e-05",
+  "w_self": "0.2",
+  "prune": "0.01",
+  "max_roots_render": "200000",
+  "log_every": "20",
+  "save_every": "1500",
+  "ot_w": "0.0",
+  "ot_geo_w": "0.0",
+  "ot_n": "2",
+  "ot_frac": "0.3",
+  "adv_start": "300",
+  "seed": "0"
+ }
+}

09_voxel_update_f1_harm/head_006000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af728224d497462e686ae8275d6bc051a6adb7330e636808522a1ba100b38969
+size 31681221