Add clean voxel arm F3+realism EMA 70k
Browse files- 07_clean_voxel_f3r/README.md +15 -0
07_clean_voxel_f3r/README.md
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Clean voxel arm: F3 + realism, EMA at step 70000
|
| 2 |
+
|
| 3 |
+
`ema_070000.pt` - best voxel arm of the clean line (outputs/clean/f3r_main/ema_070000.pt).
|
| 4 |
+
|
| 5 |
+
- Data: official RE10K train only, online VGGT-Omega teacher (vggt_omega_1b_512), fresh 11-frame clip per visit.
|
| 6 |
+
- Start: 3D-Belief trunk, decoder warm start retrained on RE10K train (free_completion_20260927 ray/step002000), fresh voxel head.
|
| 7 |
+
- Recipe: F3 + realism (--voxel-head new --voxel-fuser --voxel-gen-realism --stream-p 0.3 --voxel-stream), 8 H100 x 1 clip,
|
| 8 |
+
warmup-stable-decay lr held constant (70k is the constant-lr snapshot; a 74k->82k cooldown gave no gain). Seed 1.
|
| 9 |
+
- Full command: command.json. Code: branch clean-re10k-online-teacher (commit f1252bf at upload; merge f1252bf).
|
| 10 |
+
- Held-out (40 clean RE10K episodes, unified 4, blind, T=1.0, voxel memory rule, cuda renderer): eval_summary_40ep.json.
|
| 11 |
+
Visible LPIPS W1/W3 0.211/0.206, revisit-A PSNR after W3 25.15 dB, reveal LPIPS W2 0.213, whole KID 0.0011,
|
| 12 |
+
patch KID occluded/out-of-view 0.122/0.049.
|
| 13 |
+
- Load as an evaluate_stream_amodal.py --snapshot (dict with 'ema', 'step', 'args').
|
| 14 |
+
|
| 15 |
+
sha256 ema_070000.pt: 8a933eff05bbda72c17383b48fbeb917fb1b470cc115a419a505522166762015
|