saifahmad123/Franka_GraspNet_Test),
trained with openpi.train.py wrote it.| Path | Size | Purpose |
|---|---|---|
params/ | 6.0 GB | Model weights — all that is needed for inference |
train_state/ | 3.0 GB | Optimizer / EMA state — only needed to resume training |
assets/saifahmad123/Franka_GraspNet_Test/norm_stats.json | 2 KB | Normalization statistics |
_CHECKPOINT_METADATA | — | Orbax metadata |
hf download saifahmad123/pi05-franka-graspnet-test --local-dir ./checkpoints/pi05_franka_graspnet_40001from openpi.policies import policy_config
2from openpi.training import config as _config
3
4cfg = _config.get_config("pi05_Franka_GraspNet_Test")
5policy = policy_config.create_trained_policy(
6 cfg, "./checkpoints/pi05_franka_graspnet_4000"
7)
8
9action_chunk = policy.infer(example)["actions"]create_trained_policy reads params/ and assets/; train_state/ is ignored at
inference time. To skip the 3 GB optimizer state entirely:1hf download saifahmad123/pi05-franka-graspnet-test \
2 --local-dir ./checkpoints/pi05_franka_graspnet_4000 \
3 --exclude "train_state/*"<exp_name>/<step> and pass
--resume to train.py, matching the config that produced it.