lstm3d-pretrained-checkpoints

Four pretrained 3D-LSTM (ObsActionChunkLSTMImage3D) checkpoints used as warmup-stage models for the diffusion-policy "warmup-then-finetune" pipeline in the memory_diffusion_policy project. Each subfolder is a self-contained training output with best_model.pt, config.yaml, normalizer.pt, and TensorBoard logs/. The visualization-dumps (eval_viz/) are intentionally excluded to keep the bundle small.

Contents

Task	Subfolder	Hidden	Horizon	Best ckpt size
Lego-stacking	lego_h64_sub1_h600	64	600	~130 MB
Lightbulb	lightbulb_h128_sub1_h600	128	600	~134 MB
Peg-insert	peg_insert_h128_sub1_h1000	128	1000	~134 MB
Fruit-swap	fruit_swap_h64_sub1_h700	64	700	~130 MB

Downloading

All checkpoints at once

pip install huggingface_hub
huggingface-cli download harrywang01/lstm3d-pretrained-checkpoints \
    --local-dir ./lstm3d_checkpoints

or in Python:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="harrywang01/lstm3d-pretrained-checkpoints",
    local_dir="./lstm3d_checkpoints",
)

Just one checkpoint

huggingface-cli download harrywang01/lstm3d-pretrained-checkpoints \
    --include "lego_h64_sub1_h600/*" \
    --local-dir ./lstm3d_checkpoints

or in Python:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="harrywang01/lstm3d-pretrained-checkpoints",
    allow_patterns=["lego_h64_sub1_h600/*"],
    local_dir="./lstm3d_checkpoints",
)

Replace lego_h64_sub1_h600 with any other subfolder name from the table above.