#bindings #tensorflow #machine-learning #ffi #cross-platform


A safe Rust wrapper of TensorFlow Lite C API supporting x86_64 and ARM (iOS, Android)

11 unstable releases (3 breaking)

0.6.0 Apr 13, 2023
0.5.2 Nov 22, 2022
0.5.1 Aug 18, 2022
0.4.1 Feb 1, 2022
0.3.2 Oct 4, 2021

#171 in Machine learning

Download history 12/week @ 2023-11-03 20/week @ 2023-11-10 105/week @ 2023-11-17 24/week @ 2023-11-24 37/week @ 2023-12-01 38/week @ 2023-12-08 33/week @ 2023-12-15 25/week @ 2023-12-22 11/week @ 2024-01-05 2/week @ 2024-01-12 10/week @ 2024-01-19 50/week @ 2024-01-26 24/week @ 2024-02-02 24/week @ 2024-02-09 118/week @ 2024-02-16

218 downloads per month

MIT license


Contains (Zip file, 1.5MB) build-res/docsrs_res.zip

Crates.io Docs.rs

This crate is a safe Rust wrapper of TensorFlow Lite C API. Its API is very similar to that of TensorFlow Lite Swift API.

Supported Targets

Targets below are tested. However, others may work, too.

  • iOS: aarch64-apple-ios and x86_64-apple-ios
  • MacOS: x86_64-apple-darwin
  • Linux: x86_64-unknown-linux-gnu
  • Android: aarch64-linux-android and armv7-linux-androideabi
  • Windows (see details)

See compilation section to see build instructions for your target. Please read Optimized Build section carefully.


  • xnnpack - Compiles XNNPACK and allows you to use XNNPACK delegate. See details of XNNPACK on here.
  • xnnpack_qs8 - Compiles XNNPACK with additional build flags to accelerate inference of operators with symmetric quantization. See details in this blog post. Implies xnnpack.
  • xnnpack_qu8 - Similar to xnnpack_qs8, but accelerates few operators with asymmetric quantization. Implies xnnpack.

Note: xnnpack is already enabled for iOS, but xnnpack_qs8 and xnnpack_qu8 should be enabled manually.


The example below shows running inference on a TensorFlow Lite model.

use tflitec::interpreter::{Interpreter, Options};
use tflitec::tensor;
use tflitec::model::Model;

// Create interpreter options
let mut options = Options::default();
options.thread_count = 1;

// Load example model which outputs y = 3 * x
let model = Model::new("tests/add.bin")?;
// Or initialize with model bytes if it is not available as a file
// let model_data = std::fs::read("tests/add.bin")?;
// let model = Model::from_bytes(&model_data)?;

// Create interpreter
let interpreter = Interpreter::new(&model, Some(options))?;
// Resize input
let input_shape = tensor::Shape::new(vec![10, 8, 8, 3]);
let input_element_count = input_shape.dimensions().iter().copied().reduce(std::ops::Mul::mul).unwrap();
interpreter.resize_input(0, input_shape)?;
// Allocate tensors if you just created Interpreter or resized its inputs

// Create dummy input
let data = (0..input_element_count).map(|x| x as f32).collect::<Vec<f32>>();

let input_tensor = interpreter.input(0)?;
assert_eq!(input_tensor.data_type(), tensor::DataType::Float32);

// Copy input to buffer of first tensor (with index 0)
// You have 2 options:
// Set data using Tensor handle if you have it already
// Or set data using Interpreter:
assert!(interpreter.copy(&data[..], 0).is_ok());

// Invoke interpreter

// Get output tensor
let output_tensor = interpreter.output(0)?;

assert_eq!(output_tensor.shape().dimensions(), &vec![10, 8, 8, 3]);
let output_vector = output_tensor.data::<f32>().to_vec();
let expected: Vec<f32> = data.iter().map(|e| e * 3.0).collect();
assert_eq!(expected, output_vector);
# // The line below is needed for doctest, please ignore it
# Ok::<(), Box<dyn std::error::Error>>(())

Prebuilt Library Support

As described in the compilation section, libtensorflowlite_c is built during compilation and this step may take a few minutes. To allow reusing prebuilt library, one can set TFLITEC_PREBUILT_PATH or TFLITEC_PREBUILT_PATH_<NORMALIZED_TARGET> environment variables (the latter has precedence). NORMALIZED_TARGET is the target triple which is converted to uppercase and underscores, as in the cargo configuration environment variables. Below you can find example values for different TARGETs:

  • TFLITEC_PREBUILT_PATH_AARCH64_APPLE_IOS=/path/to/TensorFlowLiteC.framework
  • TFLITEC_PREBUILT_PATH_ARMV7_LINUX_ANDROIDEABI=/path/to/libtensorflowlite_c.so
  • TFLITEC_PREBUILT_PATH_X86_64_APPLE_DARWIN=/path/to/libtensorflowlite_c.dylib
  • TFLITEC_PREBUILT_PATH_X86_64_PC_WINDOWS_MSVC=/path/to/tensorflowlite_c.dll. Note that, the prebuilt .dll file must have the corresponding .lib file under the same directory.

You can find these files under the OUT_DIR after you compile the library for the first time, then copy them to a persistent path and set environment variable.

XNNPACK support

You can activate xnnpack features with a prebuilt library, too. However, you must have built that library with XNNPACK, otherwise you will see a linking error.

Local Header Directory Support

Some tensorflow header files are downloaded from GitHub during compilation with or without prebuild binary. However, some users may have difficulty to access GitHub. Hence, one can pass header directory with TFLITEC_HEADER_DIR or TFLITEC_HEADER_DIR_<NORMALIZED_TARGET> environment variables (the latter has precedence). See the example command below:

TFLITEC_HEADER_DIR=/path/to/tensorflow_v2.9.1_headers cargo build --release
# Structure of /path/to/tensorflow_v2.9.1_headers is given below:
# tensorflow_v2.9.1_headers
# └── tensorflow
#     └── lite
#         ├── c
#         │ ├── c_api.h            # Required
#         │ ├── c_api_types.h      # Required
#         │ └── common.h           # Required if xnnpack enabled
#         └── delegates
#             └── xnnpack
#                 └── xnnpack_delegate.h  # Required if xnnpack enabled


This library builds libtensorflowlite_c dynamic library and must be linked to it. This is not an issue if you build and run a binary target with cargo run. However, if you run your binary directly, you must have libtensorflowlite_c dynamic library in your library search path. You can either copy built library under target/{release,debug}/build/tflitec-*/out to one of the system library search paths (such as /usr/lib or /usr/local/lib) or add that directory (out) to search path.

Similarly, if you distribute a prebuilt library depending on this, you must distribute libtensorflowlite_c, too. Or, document this warning in your library to instruct your users.


Current version of the crate builds tag v2.9.1 of the tensorflow project. Compiled dynamic library or Framework will be available under OUT_DIR (see cargo documentation) of tflitec. You won't need this most of the time, because the crate output is linked appropriately. In addition, it may be better to read prebuilt library support section to make your builds faster. For all environments and targets you will need to have:

Optimized Build

To build TensorFlow for your machine with native optimizations or pass other --copts to Bazel, set environment variable below:

TFLITEC_BAZEL_COPTS="OPT1 OPT2 ..." # space seperated values will be passed as `--copt=OPTN` to bazel
TFLITEC_BAZEL_COPTS="-march=native" # for native optimized build
# You can set target specific opts by appending normalized target to variable name

Some OSs or targets may require additional steps.


  • Android NDK
  • Following environment variables should be set appropriately to build TensorFlow for android:
  • Bindgen needs extra arguments, so set the environment variable below:
# Set appropriate host tag and target name.
# see https://developer.android.com/ndk/guides/other_build_systems
HOST_TAG=darwin-x86_64 # as example
TARGET_TRIPLE=arm-linux-androideabi # as example
-I${ANDROID_NDK_HOME}/sources/cxx-stl/llvm-libc++/include/ \
-I${ANDROID_NDK_HOME}/sysroot/usr/include/ \
  • (Recommended) cargo-ndk simplifies cargo build process. Recent version of the tool has --bindgen flag which sets BINDGEN_EXTRA_CLANG_ARGS variable appropriately. Hence, you can skip the step above.


Windows support is experimental. It is tested on Windows 10. You should follow instructions in the Setup for Windows section on TensorFlow Build Instructions for Windows. In other words, you should install following before build:

  • Python 3.8.x 64 bit (the instructions suggest 3.6.x but this package is tested with 3.8.x)
  • Bazel
  • MSYS2
  • Visual C++ Build Tools 2019

Do not forget to add relevant paths to %PATH% environment variable by following the TensorFlow Build Instructions for Windows carefully (the only exception is the Python version).

No runtime deps

~70K SLoC