FeatherCNN is a high-performance lightweight CNN inference library developed by Tencent AI Platform Department. Originating from the game AI project for King of Glory (Chinese: 王者荣耀), FeatherCNN aims to build a neural model for MOBA game AI and run it efficiently on mobile devices. Currently, it targets ARM CPUs, with plans to extend support to other architectures in the future.

Main Features of FeatherCNN

• High Performance: FeatherCNN delivers state-of-the-art inference computing performance across various devices, including mobile phones (iOS/Android), embedded devices (Linux), and ARM-based servers (Linux).
• Easy Deployment: The library is designed to eliminate third-party dependencies, facilitating straightforward deployment on mobile platforms.
• Featherweight: The compiled FeatherCNN library is compact, typically only a few hundred KBs in size.

Technical Architecture and Implementation

FeatherCNN is built with a focus on performance and efficiency. It utilizes TensorGEMM for fast inference computation on ARM architectures. The library’s architecture is designed to optimize resource usage while maintaining high throughput, making it ideal for mobile and embedded applications.

Setup and Installation Process

To get started with FeatherCNN, follow these steps:

1. Clone the repository using the command:

git clone -b master --single-branch https://github.com/tencent/FeatherCNN.git

2. Follow the detailed instructions for building from source available in the Build From Source guide.
3. Refer to the specific guides for iOS and Android platforms for tailored setup instructions.

Usage Examples and API Overview

FeatherCNN supports model format conversion and provides runtime interfaces for inference. Here’s a brief overview of how to use the library:

Model Format Conversion

FeatherCNN accepts Caffemodels and merges the structure file (.prototxt) and the weight file (.caffemodel) into a single binary model (.feathermodel). The conversion tool requires protobuf, but it is not needed for the library itself.

To initialize the network for inference, use the following code:

feather::Net forward_net(num_threads);
forward_net.InitFromPath(FILE_PATH_TO_FEATHERMODEL);

To perform forward computation, use:

forward_net.Forward(PTR_TO_YOUR_INPUT_DATA);

Extracting output data can be done with:

forward_net.ExtractBlob(PTR_TO_YOUR_OUTPUT_BUFFER, BLOB_NAME);

Additionally, you can retrieve the blob’s data size using:

size_t data_size = 0;
forward_net.GetBlobDataSize(&data_size, BLOB_NAME);

Performance Benchmarks

FeatherCNN has been tested on various devices, showcasing its performance capabilities. For detailed benchmarks, visit the Benchmarks page.

Community and Contribution

FeatherCNN welcomes contributions from the community. If you encounter any issues or have suggestions for enhancements, please open an issue in the repository. Join the community discussions on Telegram or QQ: 728147343.

Conclusion

FeatherCNN stands out as a robust solution for developers looking to implement lightweight CNN inference in mobile and embedded applications. With its high performance, easy deployment, and compact size, it is an excellent choice for enhancing AI capabilities in various domains.

Resources

For more information, visit the FeatherCNN GitHub Repository.

FAQ Section