How shawnzhang0828/polynormer-reimplementation works
Polynormer-Reimplementation trains and evaluates a graph node-classification model. The principal workflow loads a supported dataset and its splits, trains local layers during warm-up, then trains with global attention enabled, evaluates the selected metric, and saves the best checkpoints. The model combines graph-local attention with linear global attention. The README also identifies a notebook experiment runner; its implementation was not sampled, so its internal wiring is not shown.
Main entry points: train.py CLI trainer · data_loaders.py Dataset loader · polynormer.py Polynormer · io.py Checkpoint I/O
Components 16
Experiment entry
- CLI trainer
train.py - Notebook runner
run_experiments.ipynb - Default config
config.py - Seed control
utils/seed.py
Graph data
- Dataset loader
utils/data_loaders.py - Dataset providers
- Graph data
Polynormer model
- Polynormer
models/polynormer.py - Local attention
models/local_attention.py - Global attention
models/global_attention.py
Training results
- Local warm-up
train.py - Local-global training
train.py - Metric calculation
utils/metrics.py - Checkpoint I/O
utils/io.py - Checkpoint files
Other components
- Experimenter
Connections 20
- Experimenter → CLI trainer: starts training
- CLI trainer → Default config: loads defaults
- CLI trainer → Seed control: sets seeds
- CLI trainer → Dataset loader: loads dataset
- Dataset loader → Dataset providers: requests dataset
- Dataset providers → Dataset loader: returns graph
- Dataset loader → Graph data: prepares splits
- CLI trainer → Graph data: reads features
- CLI trainer → Local warm-up: runs first
- Local warm-up → Polynormer: trains frozen-global
- CLI trainer → Checkpoint I/O: saves best warm-up
- CLI trainer → Checkpoint I/O: loads best warm-up
- CLI trainer → Local-global training: runs second
- Local-global training → Polynormer: trains full model
- Polynormer → Local attention: calls local layers
- Polynormer → Global attention: calls global layers
- CLI trainer → Metric calculation: computes metrics
- CLI trainer → Checkpoint I/O: saves best model
- Checkpoint I/O → Checkpoint files: writes and reads
- Experimenter → Notebook runner: runs experiments