a12s12's picture
Super-squash branch 'main' using huggingface_hub
f065e53
|
Raw
History Blame Contribute Delete
1.17 kB
---
license: other
tags:
- spatial-diffuse-slot
- semanticist
- migration-backup
---
# SpatialDiffuseSlot-L β€” migration backup (2026-07-16)
Semanticist DiT-L/ViT (tok_L pretrained init) + our 85 multi-res attn-pool
tokens + our spatial-align attention mask. DiT trunk frozen (Phase-1 warmup).
**Migration backup** taken at server deadline (7/16). Latest checkpoint + recon
samples so the run can be resumed / inspected off the original server.
## Contents
- `step150000/` β€” accelerate checkpoint (model.safetensors, optimizer.bin, EMA,
random states). Resume with `ckpt_path: .../models/step150000`.
- `samples/` β€” reconstruction grids (left=GT, right=recon) at several steps.
## Status at backup
- step 150000, epoch 15/400
- diff_loss 0.35 (step0) β†’ **0.17** (running avg), repa_loss βˆ’0.21 β†’ βˆ’0.23
- recon quality: recognizable structure/color/layout β€” spatial alignment working
- recipe: eff batch 256, blr 2.5e-5, warmup 100ep, cosine, EMA (== Semanticist)
## Resume
Code lives in the `semanticist` repo (SpatialDiffuseSlot). See
`SPATIAL_RUN_README.md` for full setup. Phase-2 = unfreeze trunk with
`freeze_dit: false` + `dit_lr_scale: 0.1`.