TensorFlow And TensorFlow-Lite Plug-in For WasmEdge
Developers can use WASI-NN to inference the models. However, for the TensorFlow and TensorFlow-Lite users, the WASI-NN APIs could be more friendly to retrieve the input and output tensors. Therefore WasmEdge provides the TensorFlow-related plug-in and rust SDK for inferencing models in WASM.
This is not a WASI-NN compatible plug-in. If you are finding the plug-ins working with the WASI-NN crate, please follow the tensorflow-lite backend instead.
Prerequisite
Please ensure that you Rust and WasmEdge installed.
Developers will add the wasmedge_tensorflow_interface crate as a dependency to their Rust -> Wasm applications. For example, add the following line to the application's Cargo.toml file.
[dependencies]
wasmedge_tensorflow_interface = "0.3.0"
Developers will bring the functions of wasmedge_tensorflow_interface into scope within their Rust -> Wasm application's code. For example, adding the following code to the top of their main.rs.
use wasmedge_tensorflow_interface;
Image Loading And Conversion
In this crate, we provide several functions to decode and convert images into tensors using the WasmEdge-Image host functions.
To use these functions in WASM and execute in WasmEdge, users should install WasmEdge with WasmEdge-Image plug-in.
For decoding the JPEG images, there are:
// Function to decode JPEG from buffer and resize to RGB8 format.
pub fn load_jpg_image_to_rgb8(img_buf: &[u8], w: u32, h: u32) -> Vec<u8>
// Function to decode JPEG from buffer and resize to BGR8 format.
pub fn load_jpg_image_to_bgr8(img_buf: &[u8], w: u32, h: u32) -> Vec<u8>
// Function to decode JPEG from buffer and resize to RGB32F format.
pub fn load_jpg_image_to_rgb32f(img_buf: &[u8], w: u32, h: u32) -> Vec<f32>
// Function to decode JPEG from buffer and resize to BGR32F format.
pub fn load_jpg_image_to_rgb32f(img_buf: &[u8], w: u32, h: u32) -> Vec<f32>
For decoding the PNG images, there are:
// Function to decode PNG from buffer and resize to RGB8 format.
pub fn load_png_image_to_rgb8(img_buf: &[u8], w: u32, h: u32) -> Vec<u8>
// Function to decode PNG from buffer and resize to BGR8 format.
pub fn load_png_image_to_bgr8(img_buf: &[u8], w: u32, h: u32) -> Vec<u8>
// Function to decode PNG from buffer and resize to RGB32F format.
pub fn load_png_image_to_rgb32f(img_buf: &[u8], w: u32, h: u32) -> Vec<f32>
// Function to decode PNG from buffer and resize to BGR32F format.
pub fn load_png_image_to_rgb32f(img_buf: &[u8], w: u32, h: u32) -> Vec<f32>
Developers can load, decode, and resize image as following:
let mut file_img = File::open("sample.jpg").unwrap();
let mut img_buf = Vec::new();
file_img.read_to_end(&mut img_buf).unwrap();
let flat_img = wasmedge_tensorflow_interface::load_jpg_image_to_rgb32f(&img_buf, 224, 224);
// The flat_img is a vec<f32> which contains normalized image in rgb32f format and resized to 224x224.
Inferring TensorFlow And TensorFlow-Lite Models
For using the TFSession struct to inference the TensorFlow models and executing in WasmEdge, users should install the WasmEdge-TensorFlow plug-in with dependencies.
For using the TFLiteSession struct and to inference the TensorFlow-Lite models executing in WasmEdge, users should install the WasmEdge-TensorFlowLite plug-in with dependencies.
Create Session
First, developers should create a session to load the TensorFlow or TensorFlow-Lite model.
// The mod_buf is a vec<u8> which contains the model data.
let mut session = wasmedge_tensorflow_interface::TFSession::new(&mod_buf);
The above function creates the session for TensorFlow frozen models. Developers can use the new_from_saved_model function to create from saved-models:
// The mod_path is a &str which is the path to saved-model directory.
// The second argument is the list of tags.
let mut session = wasmedge_tensorflow_interface::TFSession::new_from_saved_model(model_path, &["serve"]);
Or use the TFLiteSession to create a session for inferring the tflite models.
// The mod_buf is a vec<u8> which contains the model data.
let mut session = wasmedge_tensorflow_interface::TFLiteSession::new(&mod_buf);
Prepare Input Tensors
// The flat_img is a vec<f32> which contains normalized image in rgb32f format.
session.add_input("input", &flat_img, &[1, 224, 224, 3])
.add_output("MobilenetV2/Predictions/Softmax");
Run TensorFlow Models
session.run();
Convert Output Tensors
let res_vec: Vec<f32> = session.get_output("MobilenetV2/Predictions/Softmax");
Build And Execution
After completing your code, you can follow the command to compile into WASM.
cargo build --target=wasm32-wasi
The output WASM file will be at target/wasm32-wasi/debug/ or target/wasm32-wasi/release.
Please refer to WasmEdge CLI for WASM execution.
Examples
Work in Progress