rasulkhanbayov/ec-tttlrm-archive

Error-corrective fast-weight updates for long-stream tttLRM: research archive

This is the complete research record of a project that fine-tunes the autoregressive tttLRM (tttLRM-AR) with an error-corrective update of its fast weights. It holds every fine-tuned checkpoint, including runs that failed or were dropped, and a full copy of the code, the evaluation logs, the per-scene result CSVs, the SLURM scripts with their job logs, and the decision notes. The notes record the pass/fail criteria that were fixed before each result came in. provenance/MANIFEST_sha256.tsv gives the sha256 of every file in this repository.

Layout

Path	What
checkpoints/<round>/<run>/	final.pt (model state dict), args.json (exact training arguments), train_log.csv
checkpoints/warmstart/depth_head.pt	depth-head warm start shared by all *_s1_n192 runs
project/	code (02_code), paper sources (01_paper), results and notes (04_results), SLURM scripts and logs (07_slurm_jobs)
provenance/lineage.tsv	for every run: the checkpoint it was initialised from, its scene list, resolution, views, steps and seed
provenance/third_party_commits.tsv	upstream repositories and the exact commits used (tttLRM, Long-LRM, DL3DV-10K tools, FSM, ReCoSplat)
provenance/env_fwf27.yml, pip_freeze_fwf27.txt	software environment
provenance/released_tttlrm/	sha256 and Hugging Face download metadata of the released tttLRM weights we started from
provenance/data_listing/*.tsv.gz	path and size of every file in each dataset tree we used (not the data)

Models in the paper's tables

Each main model is trained in three stages. provenance/lineage.tsv has the exact arguments for each one.

1. Start from the released dl3dv_ar.pt and train the first stage: r2/{ctl,ec05,l2}_s0_n256 and r3/{ecdir,fsm}_s0_n256.
2. Train at higher resolution: r3/hr_*.
3. Train at native resolution: r3/nat_*, the models that are evaluated.

Table row	Checkpoint
EC (ours)	r3/nat_ec05_s0 (seed 1: r3/nat_ec05_s1)
Matched fine-tune (control)	r3/nat_ctl_s0 (seed 1: r3/nat_ctl_s1)
EC, direction only	r3/nat_ecdir_s0
Elastic consolidation (FSM rule)	r3/nat_fsm_s0
L2 inner loss	r3/nat_l2_s0
Ablation (tune30): control, EC, direction only, plain step, delta rule	r2/ctl_s1_n192, r3/{ec05,ecdir,deltap2,delta4}_s1_n192

All other runs (r1, pilot, pilots, the remaining r2/r3 runs, smoke, debug) are explored or failed variants. Their outcomes are recorded in project/04_results/notes.md.

Reproducing

• Data, not included. DL3DV-10K / DL3DV-140 and ScanNet++ may not be redistributed. Obtain them from their owners under their terms. Then rebuild our exact inputs with the scene lists in project/04_results/scene_lists/ and the scripts in project/02_code/fwf_torch/data_prep/. Use provenance/data_listing/ to check the result.
• Released tttLRM weights, not included. They are public at their source; provenance/released_tttlrm/sha256.txt identifies the exact files. The download script is project/07_slurm_jobs/02_download_ckpts_dl3dv140_cpu.sh.
• Jobs. project/07_slurm_jobs/NN_*.sh are numbered in the order they ran. Each job's .out/.err log sits next to its script.
• Paper numbers. Every number in the paper is generated from the logs in project/04_results by project/02_code/fwf_torch/jobs/main_analyze.py and make_paper_tables.py, which write project/01_paper/ec_results.tex and project/01_paper/tables/.

License

The checkpoints are derivative works of tttLRM. They are distributed under the Adobe Research License in LICENSE_tttLRM_Adobe_Research.md and may be used for noncommercial research only.