Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# fastwam-egodex-uncond-1cam320-30k (v1, cosine lr)
|
| 2 |
+
|
| 3 |
+
FastWAM on the EgoDex human-hands corpus: 26 task dirs, 44,507 episodes / 12,230,294 frames, lerobot v2.1, AV1 180x320 @ 30 fps, 44-D action/state.
|
| 4 |
+
|
| 5 |
+
4 GPU x batch 16 x grad-accum 3 = global batch 192, 30k steps, **lr 1e-4 with cosine decay** (to 1e-6 by the end).
|
| 6 |
+
|
| 7 |
+
Trained at **192x320**, not the 176x320 the recipe originally specified: the wan22 VAE compresses 16x, so a 176-tall canvas becomes an 11-tall latent and the DiT patch (2, 2) rejects the odd dimension. The real constraint is a multiple of 32. 192 is the nearest one above the native 180.
|
| 8 |
+
|
| 9 |
+
Weights only; `checkpoints/state/` is not published.
|