How jiamingkong/meshgraphnets_paddlepaddle works
Project Type and Purpose
• This is a machine-learning research reproduction project implemented in PaddlePaddle.
• It implements “Learning Mesh-Based Simulation with Graph Networks” for the cylinder_flow dataset.
• Core functionality: data preparation, graph-based model definition, training, inference (rollout), and visualization.
High-Level Architecture
a. Data Ingestion & Preparation
• Raw TFRecord files (DeepMind’s cylinder_flow)
• parse_tfrecord.py → converts TFRecords → HDF5/PKL format
b. Dataset & Pipeline
• dataset/
• data.py – Graph Data object and batching (Collator)
• dataloader.py – wraps Data into Paddle’s DataLoader
• fpc.py – (if used) feature preprocessing or additional I/O
• transforms/ – Graph transforms: FaceToEdge, Cartesian, Distance, Scatter
c. Model Definition
• model/
• blocks.py – reusable GNN blocks (e.g., edge-update, node-update)
• model.py – full MeshGraphNet architecture
• simulator.py – rollout logic for multi-step prediction
d. Training Component
• train.py
• uses PaddlePaddle optimizers, loss functions
• writes model checkpoints under checkpoint/
e. Inference (Rollout)
• rollout.py – loads checkpoint, performs forward rollouts on test graphs
• outputs PKL files under result/
f. Visualization
• visualize_cylinder_flow.py – renders PKL results → MP4 videos
g. Utilities
• utils/ – metrics (average_meter.py), noise injection, normalization, general helpers
Relationships and Data Flow
1. User downloads raw dataset → data/cylinder_flow/*.tfrecord
2. parse_tfrecord.py reads TFRecord → writes data/cylinder_flow/datapkls/
3. train.py
• Loads PKLs via dataset module
• Applies transforms → batched graph data
• Feeds into model → computes loss → optimizer step
• On intervals, saves parameters to checkpoint/
4. rollout.py
• Loads saved params → feeds initial graph(s) → uses simulator.py to unroll predictions → dumps PKL results
5. visualize_cylinder_flow.py
• Reads PKL results → Matplotlib/other tools to generate video
Architectural Patterns and Design Principles
• Modular separation: data, transforms, model, training, inference, visualization
• Graph-centric abstraction: custom Data and Collator mimic PyG behavior
• Pipeline pattern: clear stages (preprocess → dataset → transforms → model → postprocess)
• Checkpointing for reproducibility
• Single-process monolithic service rather than microservices
Technologies and Frameworks
• PaddlePaddle core for tensors, training, GPU support
• H5py or Pickle for on-disk graph data
• NumPy for transforms and batching
• Standard Python scripts for orchestration