Paper Code

Paper Code

Implement experimental code following best practices for reproducibility and academic standards.

Project Structure

project/
├── configs/           # Hyperparameters, settings
├── data/              # Raw and processed data
├── models/            # Model implementations
├── scripts/           # Training and evaluation scripts
├── notebooks/        # Exploratory analysis
├── tests/             # Unit and integration tests
├── logs/              # Experiment tracking
└── README.md          # Documentation

Code Quality Standards

Reproducibility

• Fixed random seeds (torch.manual_seed, numpy.random.seed)
• Config files for all hyperparameters
• Exact command reproduction in logs
• Environment specification (requirements.txt, conda env)

Style

• Clear variable and function names
• Comprehensive docstrings (Google/NumPy style)
• Type hints for function signatures
• Modular, reusable components

Testing

• Test core functions individually
• Validate data preprocessing pipelines
• Check metric calculations match paper definitions

Scripts

code-scaffold.py

Generate project templates for common experiment types:

• Classification experiments
• Generation experiments
• RL training loops
• Baseline implementations

References

• references/best-practices.md: Code quality standards and project structure

code-validator.py

Check code quality:

• Linting (flake8, black)
• Type checking (mypy)
• Import validation

Tips

• Start with a working baseline, then modify
• Keep experiment configs versioned
• Log everything (parameters, seeds, results)
• Use Weights & Biases or MLflow for tracking