gsjang's picture
release_20261102: Estaillades and Portland 100-epoch SliceGAN runs (finished 2026-11-02), card note
ee555d4 verified
|
Raw History Blame Contribute Delete
6.68 kB
---
license: cc-by-4.0
pretty_name: Sweep world models and stereo-sections artifacts (2D-to-3D microstructure generation)
---
# Stereo-sections artifacts
Large artifacts of the project behind "Sweep World Models: Identifiability and Calibrated Transition Testing for 2D-to-3D
Microstructure Generation" (CVPR 2027 submission), and of its earlier stages ("What can sections say about volumes?":
slice-to-volume generation of 3D two-phase microstructure from 2D sections, identifiability analysis, StereoCal calibration,
few-section protocol E6). Code, paper
sources, the complete registered method log and all result JSONs are in the companion GitHub repository (private during
review: rayjang/stereo-sections).
Layout (tar files are plain `tar cf` archives of the directories named; untar in place next to the code):
- `ckpt/` -- every checkpoint (`ckpt_*.pt`, `ckpt_*.npz`): cross-material 3D EDM priors (`ckpt_blibC_3d_cond3long.pt`,
`ckpt_blibM_3d_cond3long.pt` with their descriptor normalisation `.npz`), per-target 2D section / pair models, the
StereoCal policies (`ckpt_{T}_3d_cal_*.pt`) and every earlier variant.
- `results/` -- per-run result files (`cal_*.json`, `calscore_*.json`, `maintab/`, tables, logs), `maintab/` -- metric caches.
- `local/data.tar` -- the preprocessed binary crops of all targets (`data/{T}_train.npy` = 400 random 32^3 crops of the
sectioned half, `data/{T}_test.npy` = non-overlapping cubes of the held-out half, bands, fraction laws, the few-section
budget files `data/{T}_k{k}_d{d}.npz` and `data/{T}_trainvol.npy`).
- `local/SliceGAN.tar`, `remote/SliceGAN.tar` -- official SliceGAN trainings (generators, logs, example tifs).
- `local/microlad_runs.tar` -- MicroLad baseline runs.
- `local/samples_{prefix}.tar`, `remote/samples_{prefix}.tar` -- every generated pool (`samples/{T}_{stem}.npy`, float16,
+1 pore / -1 solid); `remote/` holds the pools and checkpoints produced on the second machine (cathode targets).
Targets: carbonates bfKetton, bfIndiana, bfPortland, bfEstaillades, bfKME90583Cb; cathodes beNMCsg, beMJ1cath,
beTangcath, beN1u, beN1c, beN2u, beN2c, beN3u, beN3c, beN4c; development sandstones bcBerea, bcBuffBerea, bcClashach.
beNMCsg, the NMC example volume distributed with SliceGAN, is a 252x252x253 crop of the NREL cathode behind beN2u with the
carbon-binder domain counted as solid (beN2u counts it with the pores), so the fifteen paper targets come from fourteen scans.
Source scans are public datasets (see the paper's references); the 10-13 release above stores only derived crops and
generated volumes. The 10-25 release below also mirrors source data and public pretrained weights.
## Release 2026-10-25 (`release_20261025/`)
Every artifact that the GitHub repository does not hold and that was produced or changed after the 10-13 release, as raw
files at their workspace paths (the paper at that date).
- `release_20261025/local/` -- first machine: `kikuchi/samples/` (generated pools, 128 volumes of 64^3 per stem: official
SliceGAN `*_sgoff10T*`, pair critic `*_slab2_*`, fixed-target and projected-critic routes `*_swm_*` / `*_pwm_*`, the
verifier-RL arms, the fragmented control `*_official_s0_bk_*`, and earlier pools), `kikuchi/SliceGAN/Trained_Generators/`
(all trained generators, with epoch-8/9 snapshots where saved), checkpoints and caches at the top of `kikuchi/`,
`kikuchi/data/`, `kikuchi/real/` (the source X-ray CT scans as downloaded), `baselines/`, `microlad_runs/`, `ckpt/`,
`results/`, `data/`.
- `release_20261025/remote/` -- second machine, same layout. Directories holding many small files are packed as one plain tar
each (`baselines/DimExDAM_gb10.tar`, `baselines/TPDM.tar`, `microlad_runs/{target}.tar`); each tar's member list is the
`.tar.txt` next to it.
- The four official 100-epoch SliceGAN runs on bfEstaillades, bfPortland, beN2c and beN4c were stopped at epochs 73 to 85
(method log, 2026-10-25 17:34); their snapshots and last generator are included but are not 100-epoch results.
Third-party material: `kikuchi/real/` and `data/` contain copies of public datasets (X-ray CT scans, materials databases,
CIFAR-10), and `ckpt/` and `baselines/` contain public pretrained weights, mirrored as they were used in the experiments.
Their original licenses and citations apply (see the paper's references). The CC-BY-4.0 license above covers the files
produced by this project. The four Imperial carbonate scans (figshare 1230044 to 1230047, CC BY 4.0) are courtesy
of Qatar Carbonates and Carbon Storage Research Centre, Imperial College London.
## Release 2026-11-02 (`release_20261102/`)
Everything that the GitHub repository does not hold and that was still missing here on 2026-11-02, from both machines, as raw
files at their workspace paths. A file already present in an earlier release with the same path and size is not repeated.
- `release_20261102/local/` -- first machine: `kikuchi/samples/` (generated pools of the final paper: SliceGAN over seeds and
epoch snapshots `*_sgoff10_s{1,2}_ep{8,9,10}T*`, SliceGAN at its published 100-epoch stop `*_sgoff100*`, the pair critic,
MicroLad, the Gaussian random field `*_grf*`, the fragmented control and the route pools), `kikuchi/SliceGAN/` (trained
generators not in earlier releases), world-model and encoder checkpoints at the top of `kikuchi/`, `kikuchi/data/`,
`kikuchi/data64tmp/`, `kikuchi/kbudget/` arrays (perturbation twins and score caches), `kikuchi/real/` (source scans; the
Bentheim slice stack as `kikuchi/real/bentheim.tar`) and `microlad_runs/`.
- `release_20261102/remote/` -- second machine, same layout: generated pools, SliceGAN generators, MicroLad runs (each
target's training slices as `microlad_runs/{target}/train.tar`, the default-schedule runs as `microlad_runs/{target}_pub/`),
and baseline material (`baselines/DimExDAM_gb10/data/bthr/{x,y,z}.tar`,
`baselines/TPDM/dataset_sample/AAPM_256_CUBE_SCLIP_XY_L067.tar`).
- Every tar is a plain `tar cf` archive named after the folder it replaces; its member list (size and path) is the `.tar.txt`
next to it. Untar inside the folder that contains the tar.
- Added on 2026-11-03: the 100-epoch SliceGAN runs on bfEstaillades and bfPortland, which finished on 2026-11-02 (generators
with the epoch-98 and epoch-99 copies, training logs, and the T-protocol pools `*_sgoff100_s0_ep{98,99,100}T*`).
- Not included: LaTeX build products.
Pools are stored as `int8` or `float16` arrays of shape (volumes, 64, 64, 64) with +1 pore and -1 solid. The selection rule
of each release, the registered experiments that produced every file and all result tables are in `method_log.md` of the
GitHub repository.