Downloads · 30 days
7
29% of all-time downloads
arunos728/fastwam-dexjoco-uncond-3cam224-100k-egodexinit
fastwam-dexjoco-uncond-3cam224-100k-egodexinit is a machine learning model from arunos728. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Trained with FastWAM (arunos728/fastwamtest) on the dexjoco 11-task corpus, 4 GPU x batch 16 = global batch 64, 100k steps, lr 1e-4 cosine.
Downloads · 30 days
7
29% of all-time downloads
All-time downloads
24
Public
Repo size
12 GB
Likes
0
Public
Click a slice to open those files.
.pt12 GB · 100%
From the Hugging Face model README
Trained with FastWAM (arunos728/fastwam_test) on the dexjoco 11-task corpus,
4 GPU x batch 16 = global batch 64, 100k steps, lr 1e-4 cosine.
This run did not start from scratch. It was initialised from an EgoDex
human-hands checkpoint (30k steps, single base camera at 192x320, 26 task
dirs / 44,507 episodes / 12.2M frames).
Three tensor groups were dropped from that checkpoint before loading, because they are embodiment-specific -- EgoDex is human hands and dexjoco is a robot, so index i means a different joint in each:
| dropped | shape |
|---|---|
mixtures.action.action_encoder.{weight,bias} | (1024, 44) |
mixtures.action.head.{weight,bias} | (44, 1024) |
proprio_encoder.{weight,bias} | (4096, 44) |
1645 of 1649 mot tensors carried over -- the video/action expert transformer
stack, which is embodiment-agnostic. The dropped layers were reinitialised.
proprio_encoder had to go regardless: it loads with strict=True and EgoDex is
44-wide against dexjoco's 46.
Weights only: checkpoints/state/ (DeepSpeed ZeRO-1 optimizer state) is not
published, so this is for evaluation rather than resuming.
v1 takes its init from the EgoDex run trained with a cosine lr decay. Its sibling -egodexinit-v2 is identical except that its init came from an EgoDex run held at a constant lr after warmup, so the pair isolates the effect of the pretraining schedule.