Skip to content

README.md

Latest commit

 

History

History
52 lines (36 loc) · 1.9 KB

README.md

File metadata and controls

52 lines (36 loc) · 1.9 KB

BRIDGE

Same directory conventions as MDGPT (datasets/bridge, ckpts/bridge, downstream_data/bridge).

Install

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

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

Model (pygfm.baseline_models.bridge)

Module Role
BridgePrePromptModel Domain mask + 3-layer GCN + contrastive + variance regularizer
BridgeDownPromptModel Frozen backbone + MoE mask + prototypes + spectral reg + routing entropy (node)
BridgeDownPromptGraphModel Above + subgraph mean-pool / disjoint batch graph forward

Data layout (bridge name tag)

  • Raw graphs: datasets/bridge (can symlink to datasets/mdgpt)
  • Pretrain ckpt: ckpts/bridge/{dataset}/preprompt_{dataset}.pth
  • Downstream splits: downstream_data/bridge/{dataset}/{1|5}shot/splits.pt and {k}shot_graph_batch/splits.pt

Commands

# 1) Leave-one-out pretrain (exclude Cora)
python scripts/bridge/pretrain.py --target Cora --row_norm

# 2) Few-shot splits
python scripts/bridge/generate_downstream.py few_shot --dataset Cora --k_shot 1 --data_root datasets/bridge

# 3) Node classification finetune
python scripts/bridge/finetune.py --dataset Cora --k_shot 1 \
  --ckpt ckpts/bridge/cora/preprompt_cora.pth --row_norm

# 4) Graph-level few-shot (generate graph_batch first)
python scripts/bridge/generate_downstream.py graph_batch --dataset Cora --k_shot 1 --data_root datasets/bridge
python scripts/bridge/finetune_graph.py --dataset Cora --k_shot 1 \
  --ckpt ckpts/bridge/cora/preprompt_cora.pth

# 5) Batch 1-shot × 100 tasks
python scripts/bridge/run_1shot_100task.py

Note: --row_norm matches common BRIDGE setups (row-normalize then PCA). Drop it to align with MDGPT-style PCA-only.

YAML template: copy configs/_templates/gfm_preprompt_pretrain.yaml to configs/bridge/pretrain.yaml with save_dir: ckpts/bridge and data_root: datasets/bridge.