GraspNet dex_models — rebuilt for the vendored toolkit
Do not redistribute. This repository is a private machine-to-machine transfer artifact. Its contents are derived from GraspNet-1Billion (object meshes + SDFs), which is licensed for research use. The repository is private and must stay private.
Why this exists
The dex_models/*.pkl shipped in the upstream GraspNet distribution were pickled
by the pip graspnetAPI package against an older trimesh. Loading them in
an environment with trimesh==4.4.9 — the version this project's
requirements.txt pins — fails:
File graspformer/data/graspnet_api/graspnet_eval.py:88, in get_scene_models_dict
dexmodel = pickle.load(f)
AttributeError: 'DataStore' object has no attribute 'fast_hash'
trimesh.caching.DataStore no longer has fast_hash (only __hash__), so the
unpickle dies mid-reconstruction. Because get_scene_models_dict is called by the
validation harness, this kills a training run at its first validation rather
than at startup — with val_check_interval: 1600 batches and
accumulate_grad_batches: 16, that is exactly optimizer step 99.
These caches are re-pickled in a current environment so they load cleanly.
How they were produced
For each object id, the toolkit's own no-cache path was used and the result
re-pickled — i.e. exactly the object graspnet_eval.py would have built had the
cache been absent:
from graspformer.data.graspnet_api.utils.eval_utils import load_dexnet_model
m = load_dexnet_model(f"{GRASPNET_ROOT}/models/{obj_id:03d}/textured")
pickle.dump(m, open(f"dex_models/{obj_id:03d}.pkl", "wb"))
No project code was modified. The consuming code path
(if os.path.exists(dex_cache_path): pickle.load(...) else: load_dexnet_model(...))
is untouched; only the cache content changed.
Equivalence was verified on a round-trip: sdf.data and mesh.vertices are
np.array_equal to the freshly-built object. File sizes differ from upstream by
~1 KB (pickle metadata), not content — e.g. 000.pkl is 190,866,344 bytes here
vs 190,867,452 upstream.
Build environment
| python | 3.11.15 |
| trimesh | 4.4.9 |
| numpy | 1.26.4 |
| autolab-core | 1.1.1 |
| open3d | 0.19.0 |
| built on | sruk-wsc-2513 (RTX PRO 6000 Blackwell) |
Rebuild cost is ~4 s per object (88 objects, 6 workers); a rebuilt cache reloads
in ~0.07 s, versus ~4 s for the uncached load_dexnet_model path.
Contents
| File | Contents |
|---|---|
dex_models.tar.zst |
the dex_models/ directory — 88 NNN.pkl (GraspableObject3D), members stored as dex_models/NNN.pkl |
manifest.json |
per-file byte size + sha256, plus the archive's own sha256 |
Usage
hf download lipeize/graspnet-dex-models-rebuilt dex_models.tar.zst --local-dir /tmp/dex
tar -I zstd -xf /tmp/dex/dex_models.tar.zst -C "$GRASPNET_ROOT" # -> $GRASPNET_ROOT/dex_models/
The validation harness picks them up automatically — graspnet_eval.py looks for
$GRASPNET_ROOT/dex_models/%03d.pkl.
Note. If a box has no
dex_models/directory at all, the toolkit silently falls back toload_dexnet_modeland works correctly, just ~4 s per object instead of ~0.07 s. The upstream caches are only a problem when present and unreadable.