Skip to content

README.md

Latest commit

 

History

History
39 lines (28 loc) · 1.53 KB

README.md

File metadata and controls

39 lines (28 loc) · 1.53 KB

MDGPT (PrePrompt + DownPrompt)

Layout

  • Pretrain: multi-domain node contrastive learning (PrePromptModel).
  • Downstream: frozen GCN + trainable input prompt + prototype matching (DownPromptModel / DownPromptGraphModel).
  • Data convention: datasets/mdgpt/ (or GFM_DATA_ROOT).

Install

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

Run experiment scripts from the repository root so relative paths resolve.

Experiments

Step Script YAML template
Pretrain python scripts/mdgpt/pretrain.py configs/mdgpt/pretrain.yaml
Few-shot splits python scripts/mdgpt/generate_downstream.py few_shot configs/mdgpt/generate_downstream.yaml
Graph-level splits python scripts/mdgpt/generate_downstream.py graph_batch same (set mode)
Node k-shot finetune python scripts/mdgpt/finetune.py configs/mdgpt/finetune.yaml
Graph k-shot finetune python scripts/mdgpt/finetune_graph.py configs/mdgpt/finetune_graph.yaml
100-task sweep python scripts/mdgpt/run_1shot_100task.py configs/mdgpt/run_1shot_100task.yaml

Export merged defaults:

python scripts/mdgpt/pretrain.py --export-default-yaml configs/mdgpt/_defaults.yaml
python scripts/mdgpt/pretrain.py --export-run-yaml configs/mdgpt/_run.yaml -c configs/mdgpt/pretrain.yaml

Notes

  • Leave-one-out pretrain: --target Cora (excludes that dataset from source graphs).
  • Checkpoints default under ckpts/mdgpt/...; aligners saved as aligners.pkl when joblib is available.