Skip to content

README.md

Latest commit

 

History

History
56 lines (37 loc) · 2.34 KB

README.md

File metadata and controls

56 lines (37 loc) · 2.34 KB

GraphKeeper

GraphKeeper-style chain prediction pretraining and LoRA, wired into the same script/data layout as MDGPT (PyG; no DGL domain-incremental expert loop in this port).

Install

cd /path/to/repo
pip install -e .

Run python scripts/graphkeeper/*.py from the repository root.

Differences from the original (read this)

Original new_model/GraphKeeper This port
DGL, NET.observe multi-task incremental Multi-domain single-graph PrePrompt (same as MDGPT)
GCN_LoRA per-layer LoRA EmbeddingLoRA on node embeddings (low-rank residual)
Regression + analytic linear layer TaskHead prototype matching (same as MDGPT DownPrompt)

Full Domain-IL / multi-expert / DBSCAN prototypes still belong to the official repo; this port targets unified GFM pretraining + few-shot downstream.

Models

  • GraphKeeperPrePromptModel: NodeLevelPrompt per domain + shared GCN + chain-prediction BCE (+ optional contrastive loss).
  • GraphKeeperDownPromptModel / GraphKeeperDownPromptGraphModel: frozen GCN, train prompt + LoRA.

Paths

  • Data: datasets/graphkeeper (can be the same tree as datasets/mdgpt)
  • Checkpoints: ckpts/graphkeeper/{datasets}/preprompt.pth, or leave-one-domain preprompt_{target}.pth (--datasets names the output subdir; multi-domain list uses --domains)
  • Downstream: downstream_data/graphkeeper/...

Example commands

# Pretrain (leave-one-domain Cora, save under ckpts/graphkeeper/cora/, or pass --datasets my_run)
python scripts/graphkeeper/pretrain.py --target Cora --no_swanlab

# Multi-domain sequential incremental (example)
python scripts/graphkeeper/pretrain.py --domains Cora,Citeseer,Pubmed,Photo,Computers --datasets incremental_run --no_swanlab

# Add MDGPT-style contrastive (weight 0.5)
python scripts/graphkeeper/pretrain.py --target Cora --contrastive_weight 0.5 --no_swanlab

python scripts/graphkeeper/generate_downstream.py few_shot --dataset Cora --k_shot 1 --data_root datasets/graphkeeper

python scripts/graphkeeper/finetune.py --dataset Cora --k_shot 1 \
  --ckpt ckpts/graphkeeper/cora/preprompt_cora.pth --no_swanlab
# If you pretrained with --datasets my_run, use --ckpt ckpts/graphkeeper/my_run/preprompt.pth

Citation (original paper)

See BibTeX in new_model/GraphKeeper/README.md.