GiLDM-open

Lightweight public repository for GiLDM, organized into two parts:

• Method code: gildm/, scripts/, configs/, preprocessing/
• Result processing: evaluation/ (scripts only, no large experiment snapshots)

Repository Layout

• gildm/: core model, training, and sampling code
• scripts/train_gildm.py: unified training entry (stage1, stage2, or all)
• configs/: selected v6 and ablation configs
• preprocessing/: dataset preprocessing utilities
• evaluation/: evaluation and plotting scripts
• data/sample_TIs/: small sample training images

Quick Start

1. Install dependencies

pip install -r requirements.txt

2. Prepare data

• Sample files are provided in data/sample_TIs/.
• Full datasets (.npy) should be placed in dataset/.

3. Train

python scripts/train_gildm.py --config configs/ldm_v6_channel.yaml --stage all

Resume stage 2 only:

python scripts/train_gildm.py --config configs/ldm_v6_channel.yaml --stage stage2 --resume

4. Evaluate

python evaluation/evaluate_full.py --config configs/ldm_v6_channel.yaml --ckpt stage2-10.ckpt --out-dir evaluation/full_eval_channel_v6
python evaluation/evaluate_full_delta.py --config configs/ldm_v6_delta.yaml --ckpt stage2-10.ckpt --out-dir evaluation/full_eval_delta_v6

Data and Checkpoints

Large datasets and checkpoints are distributed externally via Zenodo.

• Code DOI (GitHub release archive): https://doi.org/10.5281/zenodo.19141791
• Assets record (datasets + checkpoints): https://zenodo.org/records/19142031
• Assets DOI: https://doi.org/10.5281/zenodo.19142031

Released files in the assets record:

• gildm_checkpoints_v1.0.0.tar.gz
• gildm_datasets_v1.0.0.tar.gz
• SHA256SUMS.txt

Integrity (SHA256)

• gildm_checkpoints_v1.0.0.tar.gz: dfc90983963e0fa842dcc27ba9f50bcf2fe1549c2a4640e8d52468803fbc84ae
• gildm_datasets_v1.0.0.tar.gz: b3fcb29e6f5e6bc9143358f6bbb42949835a1a1c05a3a3f990afab4849518390

Verify and extract:

shasum -a 256 -c SHA256SUMS.txt
tar -xzf gildm_checkpoints_v1.0.0.tar.gz -C .
tar -xzf gildm_datasets_v1.0.0.tar.gz -C .

Notes

• The codebase uses the gildm namespace.
• .gitignore excludes large artifacts such as *.ckpt, *.npy, wandb/, and *.raylink.downloading.