How mariewelt/openchem works
Project Type and Purpose
• OpenChem is an open-source deep learning toolkit (library) for computational chemistry and drug design.
• It provides a unified, modular API on top of PyTorch to build, train, and evaluate sequence- and graph-based molecular models via configuration files.
Top-Level Architecture Components
a. Entry Points & Configuration
• run.py, launch.py, example_configs/…
• Configuration reader → builds model, data loader, criterion, optimizer, trainer
b. Data Processing Layer
• openchem/data/: various DataLayer classes (smiles, graph, siamese, protein)
• benchmark_datasets/…: CSVs and pickled datasets
• Preprocessing utilities (openchem/data/utils.py, graph_utils.py)
c. Model Definition Layer
• openchem/modules/: embeddings (onehot, basic, openchem), encoders (RNN, GCN, CNN, edge-attention), mlp
• openchem/layers/: building blocks (conv_bn_relu, gcn, stack_augmentation)
• openchem/models/: high-level Model classes (MoleculeProtein2Label, GenerativeRNN, MolecularRNN, SiameseModel, etc.)
d. Criterion & Optimization
• openchem/criterion/: loss functions (multitask_loss, policy_gradient_loss)
• openchem/optimizer/: scheduler and optimizer wrappers (openchem_optimizer, openchem_lr_scheduler)
e. Training Engine & Utilities
• Training loop orchestration and multi-GPU support (openchem/utils/comm.py)
• Logging & monitoring (openchem/utils/logger.py, textlogger.py → Tensorboard integration)
• Metrics & RL utilities (metrics.py, rl_utils.py, sa_score/)
f. Infrastructure & Dev-Ops
• Dockerfile, .travis.yml, devtools/: CI scripts, conda environments, build recipes
• docs/: Sphinx-generated HTML and source; example usage and tutorials
• setup.py, requirements.txt: packaging and dependencies
Component Interactions and Data Flow
• The user invokes run.py/launch.py with a config → Configuration Loader
↓
• Config instantiates DataLayer → loads benchmark or user data, applies tokenization/graph conversion
↓
• Data fed to Model instance → composed from embeddings → encoder → output head
↓
• Output and ground truth passed to Criterion → loss computed
↓
• Optimizer updates model parameters via Training Engine (supports multi-GPU)
↓
• Training Engine calls Logger → writes metrics to Tensorboard and logs/…
• At checkpoints, model snapshots saved in logs/…/checkpoint
Architectural Patterns & Design Principles
• Modular plugin architecture: each layer (data, modules, models, criterion, optimizer) implements a base interface for interchangeability.
• Configuration-driven instantiation: no hard-coded wiring—everything wired via YAML/Python config.
• Separation of concerns: data handling, model definition, training loop, and logging live in distinct packages.
• Extensibility: new DataLayer, Encoder, or Model classes drop-in under openchem/… and referenced in config.
Essential Details to Capture
• Entry points and config reader that wire everything together.
• Data preprocessing pipelines converting SMILES → tensors/graphs.
• Modular breakdown: embeddings → encoders → model head → loss → optimizer.
• Multi-GPU training orchestration and logging outputs.
• Extensibility hooks (where new modules plug in).