InCoder: A Generative Model for Code Infilling and Synthesis

InCoder is a generative model designed for code infilling and synthesis. Developed by Facebook AI Research and released on GitHub, it offers two model sizes: a 1.3B parameter model and a larger 6.7B parameter model. This powerful tool leverages the HuggingFace Transformers library for ease of use and integration.

Key Features

• Code Infilling: InCoder excels at completing partially written code, significantly speeding up development and reducing errors.
• Code Synthesis: It can generate code from natural language descriptions, bridging the gap between human intent and code implementation.
• Two Model Sizes: Users can choose between a smaller, faster 1.3B parameter model and a more powerful 6.7B parameter model, depending on their needs and computational resources.
• HuggingFace Integration: Seamless integration with the popular HuggingFace Transformers library simplifies usage and deployment.
• Custom Tokenizer: InCoder uses a custom tokenizer optimized for code, ensuring accurate tokenization and improved model performance.

Usage and Examples

The GitHub repository provides example scripts (example_usage.py and example_batched_usage.py) demonstrating how to use InCoder for both single and batched code generation. These examples showcase the model's infilling capabilities and provide clear instructions for implementation.

The model requires the following libraries:

• pytorch
• tokenizers (>= 0.12.1)
• transformers

Installation can be done using pip:

pip install torch
pip install 'tokenizers>=0.12'
pip install transformers

Model Access

The pre-trained models are available on the HuggingFace Model Hub:

• 6.7B parameter model: facebook/incoder-6B
• 1.3B parameter model: facebook/incoder-1B

Comparison with Other Code Generation Models

Compared to other code generation models like Codex or TabNine, InCoder distinguishes itself through its focus on infilling and its availability through the widely used HuggingFace ecosystem. While benchmarks comparing it directly to all competitors are not readily available in a single source, its performance is notable for its ability to handle incomplete code snippets effectively.

Conclusion

InCoder represents a significant advancement in code generation, offering a powerful and accessible tool for developers. Its ease of use, coupled with its strong performance in code infilling and synthesis, makes it a valuable asset for improving coding efficiency and productivity.