How deepgraphlearning/confgf works
Project type and purpose
• This is a research-oriented Python library and command-line toolkit for molecular conformation generation using graph neural networks (the ConfGF model).
• It provides data preprocessing, model definition, training, generation, evaluation, and visualization scripts.
File structure
Top-level layout
• assets/ → static images
• config/ → experiment YAMLs for different datasets (qm9, drugs, iso17)
• confgf/ → core library package
• dataset/ → Dataset classes and loading logic
• layers/ → Graph network layer implementations (common, GIN)
• models/ → Model definitions (ScoreNet)
• runner/ → Orchestration of training and generation loops
• utils/ → Utility code (chemistry, torch helpers, metrics)
• script/ → CLI entry points (train.py, gen.py, processing scripts, evaluation scripts)
• setup.py, env.yml → installation and environment setup
README for additional insights
• ConfGF uses PyTorch and PyTorch Geometric for GNNs, RDKit for chemistry operations.
• Workflow stages: data preprocessing → model training → conformation generation → evaluation → visualization.
a. Main components
1. Data Ingestion & Preprocessing
• script/process_*.py
• confgf.dataset
• Input: raw GEOM/ISO17 files
• Output: pickled dataset files
2. Configuration Loader
• config/*.yml
• easydict or custom loader in runner
3. Training Runner
• script/train.py → confgf.runner.default_runner
• Uses confgf.models.ScoreNet and confgf.layers
• Reads preprocessed data via confgf.dataset
• Writes model checkpoints
4. Generation Runner
• script/gen.py → confgf.runner.default_runner (generation mode)
• Loads checkpoints, generates conformations
5. Evaluation
• script/get_task1_results.py, get_task2_results.py
• confgf.utils.evaluation computes COV/MAT/MMD
6. Visualization
• scripts or user code using PyMol and RDKit via confgf.utils.chem
b. Relationships & interactions
• Runners invoke dataset loaders, model builders, training loops
• Trainer outputs checkpoints → Generator inputs checkpoints
• Generator produces output conformations → Evaluator measures quality
c. Architectural patterns
• Modular Python package dividing concerns (dataset, model, runner, utils)
• Configuration-driven design (YAML)
• CLI scripts as thin wrappers invoking core library
d. Key technologies
• Python, PyTorch, PyTorch Geometric, RDKit
• YAML configs, Conda environment
Essential architectural elements to capture
• Separation between core library (confgf/) and CLI scripts (script/)
• Configuration driving behavior (config/*.yml)
• Data flow from raw files through preprocessing to model input
• Control flow in runner coordinating model, data, and training/generation loops
• Output artifacts: checkpoints, generated conformations, evaluation metrics