edgefirst-tflite

Ergonomic Rust API for TensorFlow Lite inference with DMA-BUF zero-copy and NPU-accelerated preprocessing.

Usage

[dependencies]
edgefirst-tflite = "0.1"

use edgefirst_tflite::{Library, Model, Interpreter};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let lib = Library::new()?;
    let model = Model::from_file(&lib, "model.tflite")?;

    let mut interpreter = Interpreter::builder(&lib)?
        .num_threads(4)
        .build(&model)?;

    interpreter.invoke()?;

    for (i, tensor) in interpreter.outputs()?.iter().enumerate() {
        println!("output[{i}]: {tensor}");
    }
    Ok(())
}

API Tour

The main entry points are:

• Library -- Load the TFLite shared library (auto-discovery or explicit path)
• Model -- Load a model from a file or byte buffer
• Interpreter -- Run inference via a builder pattern
• Tensor / TensorMut -- Type-safe tensor access with shape and quantization info
• Delegate -- Hardware acceleration via external delegates
• DmaBuf -- Zero-copy DMA-BUF operations (feature: dmabuf)
• CameraAdaptor -- NPU preprocessing configuration (feature: camera_adaptor)
• Metadata -- Model metadata extraction (feature: metadata)

Feature Flags

License

Apache-2.0