Skip to content

methods.md

Latest commit

 

History

History
49 lines (38 loc) · 2.63 KB

methods.md

File metadata and controls

49 lines (38 loc) · 2.63 KB
Raw

Methods

Every method — SCCD and the baselines — is a registered stem2crystal.methods.Method, so they generate and evaluate through the same harness.

from stem2crystal.methods import list_methods
list_methods()      # ['diffcsp', 'mattergen', 'microscopygpt', 'sccd']  (+ any you register)
Key Method Conditioning Upstream stack (not in this repo) Weights
sccd STEM2Crystal CoDiffusion (ours) image + composition DiffCSP + EMDiffuse (set $SCCD_HOME) download_models.py --model sccd
diffcsp DiffCSP (Jiao et al.) composition only diffcsp package ($DIFFCSP_HOME) --model diffcsp
microscopygpt MicroscopyGPT (Choudhary et al.) image + composition unsloth + 11B VLM --model microscopygpt
mattergen MatterGen (Zeni et al.) de-novo (chemical system) mattergen package ($MATTERGEN_HOME) auto by mattergen

No weights or outputs are committed. Fetch weights with scripts/download_models.py into a git-ignored outputs/. Baselines also need their upstream package installed (they are large, independent codebases); the adapter raises a clear, actionable error if the package or checkpoint is missing.

Run any method

python scripts/download_models.py --model diffcsp
python scripts/generate.py --method diffcsp --benchmark synthetic --noise low
python scripts/evaluate.py  --method diffcsp --benchmark synthetic --noise low

The composition-only and de-novo baselines ignore the image argument; image-conditioned methods (sccd, microscopygpt) use it.

Setup per baseline

DiffCSP — pip install its repo (or set $DIFFCSP_HOME); export DIFFCSP_CKPT=outputs/diffcsp/...ckpt.

MicroscopyGPT — pip install unsloth; download_models.py --model microscopygpt fetches the LoRA adapter; export MICROSCOPYGPT_ADAPTER=outputs/microscopygpt. The base 11B VLM is auto-downloaded by unsloth (≈12 GB GPU).

MatterGen — pip install mattergen (or set $MATTERGEN_HOME); the chemical_system_energy_above_hull checkpoint is fetched by the package on first use.

SCCD — export SCCD_HOME=<training-repo> (the dir with the sccd reference scripts + DiffCSP/EMDiffuse); download_models.py --model sccd; export SCCD_PHASE3=outputs/sccd/phase3_joint/best.pt. For the canonical batched 4-GPU run use the reference scripts in stem2crystal/methods/_sccd_impl/.

Add your own

See add_a_method.md — implement generate(image, composition, k), register, done. The standalone evaluator scores it with every paper metric, exactly like the methods above.