Publish code/reproduction.patch

code/reproduction.patch

@@ -853,10 +853,126 @@ index dee49f8..6252056 100644
 +The [original upload report](../../reports/slat_targets_huggingface_20261015.md) records their coverage and codec identities.
 +Exact checkpoint resume requires the checkpoint's original data and feature identities.
 diff --git a/preprocess/slat/README.md b/preprocess/slat/README.md
-index 46e4436..663732a 100644
+index 46e4436..c5f93e5 100644
 --- a/preprocess/slat/README.md
 +++ b/preprocess/slat/README.md
-@@ -181,42 +181,31 @@ Training resume also checks cache provenance and latent target checksums.
+@@ -1,7 +1,7 @@
+ # SLAT preparation
+ 
+ This stage prepares reusable training inputs for multi-object SLAT shape finetuning.
+-It reads the saved 3D-FRONT and SAGE-10k extent datasets after capture and voxel export.
++It reads saved extent datasets from 3D-FRONT, SAGE-10k, and FoundationPose after capture and voxel export.
+ It preserves each saved cube, object identity, and room or layout split.
+ 
+ The generation implementation and its configurations reside in this package.
+@@ -11,6 +11,95 @@ The generation implementation and its configurations reside in this package.
+ | `targets.py` | Select posed meshes, preserve the saved cube, and write object and scene targets |
+ | `encode.py` | Convert meshes to dual grids and apply the frozen pretrained shape encoder |
+ | `prepare.py` | Coordinate dataset phases and extract frozen conditioning features |
++| `scene_targets.py` | Export one scene token per reference voxel with an explicit latent validity flag |
++
++## Aligned scene targets at 32³
++
++The scene export combines complete meshes for all prompted objects before one frozen shape encoder processes them.
++It preserves each training view's saved cube, calibration, object order, and dataset split.
++Different views of one room can use different cubes.
++The export therefore retains each view ID.
++
++Reference voxels are the union of complete occupancy for prompted objects.
++The shared `prompt_object_cells` function selects these voxels from saved 64³ targets and pools them to 32³.
++Unprompted objects do not add reference voxels.
++Older feature caches can contain extra scene voxels from unprompted objects.
++Those extra rows do not define the reference occupancy for this export.
++
++Each reference voxel has exactly one scene token, including voxels that contain multiple objects.
++The native scene encoder can omit a small number of reference voxels because its rasterization method differs.
++The export retains those coordinates with zero feature storage and `latent_valid=false`.
++Training must exclude invalid rows from the latent loss.
++The export does not average object latents or interpolate missing values.
++
++| Array | Shape | Meaning |
++| --- | --- | --- |
++| `coords` | `[N, 3]` | Unique reference coordinates in sorted order, stored as `uint8` |
++| `feats` | `[N, 32]` | Physical scene latents stored as `float16`, with zeros at invalid rows |
++| `latent_valid` | `[N]` | Boolean flag for an exact coordinate match with native scene encoding |
++| `memberships` | `[N, K]` | Independent per-object voxel occupancy for the same reference rows |
++| `object_ids` | `[K]` | Prompt identities in the saved order |
++| `model_to_world` | `[4, 4]` | Original transform from the normalized cube to world coordinates |
++
++Multiple columns of `memberships` can be positive at one voxel.
++They do not create additional scene tokens.
++
++Run the combined export from the repository root:
++
++```bash
++python -m preprocess.slat.scene_targets \
++  --config preprocess/slat/configs/scenes_32.json
++```
++
++The configuration selects training views only and uses GPUs 0, 1, and 2.
++Sixteen CPU processes prepare meshes for six persistent GPU workers.
++Each GPU hosts two encoder processes.
++Each GPU worker uses four CPU threads to prepare native dual grids while the GPU encodes earlier grids.
++The `mesh_workers` and `raster_workers` settings control these CPU counts.
++The latter count applies separately to each GPU worker.
++The `mesh_executor` setting selects CPU processes or threads.
++CPU processes avoid Python thread contention during mesh preparation.
++They transfer geometry and reference metadata without RGB or depth tensors.
++Prefetch retains a bounded number of grids in memory.
++
++Mesh clipping omits plane cuts when all vertices lie strictly inside the plane.
++Vertices near a clipping plane retain the original Trimesh procedure.
++These changes preserve the cube, rasterization settings, encoder weights, and reference voxels.
++
++Its `encoder_config` selects the existing codec weights, dual-grid settings, and native Python environment.
++Each dataset saves its targets under `datasets/<dataset>/slat/32/scene_targets/`.
++The format identifier is `slat-scene-targets-v1`.
++Logs and combined progress reside under `outputs/slat_scene/32/preparation/`.
++
++The export reuses completed native scene encodings after it checks their dataset, cube, prompts, codec, and file checksums.
++Existing object caches retain their original files.
++Per-view completion records support resume after an interruption.
++Resume retries failed encodings and replaces corrupt output files.
++
++Add `--limit-per-dataset 2` for a short GPU check.
++Repeat the command without that argument to complete every training view.
++The command finishes active jobs before it honors SIGINT or SIGTERM.
++The default storage reserve stops preparation if free space falls below 20 GiB.
++
++`dataloader.slat_scene.SceneTargetDataset` reads one or more exported caches without loading meshes.
++It checks exact reference coordinates, occupancy labels, object identities, and cube transforms against the saved extent views.
++It normalizes valid latents with the pretrained flow statistics and keeps invalid rows zero.
++
++### Test native latent equivalence
++
++Run the regression with the actual pretrained encoder:
++
++```bash
++PARTPROBE3D_NATIVE_SLAT_TEST=1 python -m pytest \
++  preprocess/tests/test_slat_native_equivalence.py -q
++```
++
++This test requires the native CUDA environment and existing codec checkpoints.
++Set `PARTPROBE3D_NATIVE_SLAT_DEVICE` to select a GPU other than `cuda:0`.
++The test compares original serial preparation with optimized clipping and four CPU threads for rasterization.
++It covers inside meshes, clipped meshes, and overlapping objects encoded together.
++It requires identical coordinates and identical stored `float16` latents with zero tolerance.
++The normal test suite skips this GPU test unless the environment flag enables it.
+ 
+ ## Prepared data
+ 
+@@ -84,8 +173,8 @@ Scene encoding preserves object failures and includes their selected meshes in t
+ Failed objects still require successful encoding before feature extraction or finetuning.
+ 
+ Resume preserves completed object and scene latents.
+-The cached training configurations supervise objects and do not require scene targets.
+-Configurations with `require_scene_targets=true` retain scene supervision in their feature caches.
++All shape training configurations supervise object latents only.
++Feature caches exclude scene supervision, while target exports retain scene latents for diagnostics.
+ 
+ 1. Prepare targets and features for both training datasets.
+ 
+@@ -181,42 +270,31 @@ Training resume also checks cache provenance and latent target checksums.
  
  ## Hugging Face snapshots