Crate menoh

Menoh-rs

A simple wrapper of Menoh

Examples

extern crate menoh;
extern crate image;

use image::GenericImage;

const INPUT_VARIABLE_NAME: &str = "139900320569040";
const OUTPUT_VARIABLE_NAME: &str = "139898462888656";

const INPUT_BATCH_SIZE: i32 = 1;
const INPUT_CHANNEL_NUM: i32 = 1;
const INPUT_WIDTH: i32 = 28;
const INPUT_HEIGHT: i32 = 28;

const MODEL_PATH: &str = "tests/resource/mnist.onnx";
const IMAGE_PATH: &str = "tests/resource/1.png";

fn main() {
    // load category file
    let model_data = menoh::ModelData::new(MODEL_PATH).unwrap();
    let mut vpt_builder = menoh::VariableProfileTableBuilder::new().unwrap();

    let input_dims = vec![
        INPUT_BATCH_SIZE,
        INPUT_CHANNEL_NUM,
        INPUT_HEIGHT,
        INPUT_WIDTH,
    ];
    // register profile
    vpt_builder
        .add_input_profile(INPUT_VARIABLE_NAME, menoh::Dtype::Float, &input_dims)
        .unwrap();
    vpt_builder
        .add_output_profile(OUTPUT_VARIABLE_NAME, menoh::Dtype::Float)
        .unwrap();

    let vpt = vpt_builder
        .build_variable_profile_table(&model_data)
        .unwrap();

    // Create Buffer
    let mut im = to_input_vec(image::open(IMAGE_PATH).unwrap());
    let buffer = menoh::Buffer::new(&mut im);

    // Attach buffer to input variable.
    // This is not necessary operation.
    // Internal buffer is automatically generated by model.
    let mut model_builder = menoh::ModelBuilder::new(&vpt).unwrap();
    model_builder
        .attach_external_buffer(INPUT_VARIABLE_NAME, &buffer, &vpt)
        .unwrap();

    let mut model = model_builder
        .build_model(&model_data, menoh::Backend::MKL_DNN, "")
        .unwrap();

    model.run().unwrap();

    println!(
        "{:?}",
        model
            .get_internal_buffer::<f32>(OUTPUT_VARIABLE_NAME)
            .unwrap()
            .as_slice()
    );
}

// ###### Just create input data below here ######

fn resize_im(mut im: image::DynamicImage, width: u32, height: u32) -> image::DynamicImage {
    let im_w = im.width();
    let im_h = im.height();
    let shortest_edge = std::cmp::min(im_h, im_w);
    let im = im.crop(
        (im_w - shortest_edge) / 2,
        (im_h - shortest_edge) / 2,
        shortest_edge,
        shortest_edge,
    );
    im.resize_exact(width, height, image::FilterType::Nearest)
}

fn reorder_to_chw(im: &image::DynamicImage) -> Vec<f32> {
    let (im_h, im_w) = (im.height(), im.width());
    let mut input_im: Vec<f32> = vec![Default::default(); (im_h * im_w) as usize];
    for h in 0..im_h {
        for w in 0..im_w {
            input_im[(w * im_h + h) as usize] = im.get_pixel(h as u32, w as u32)[0] as f32;
        }
    }
    input_im
}

fn to_input_vec(im: image::DynamicImage) -> Vec<f32> {
    let im = resize_im(im, INPUT_WIDTH as _, INPUT_HEIGHT as _);
    let im = im.grayscale();
    reorder_to_chw(&im)
}

Modules

ffi	This module provides "raw" declarations and linkage for Menoh C API.

Structs

Buffer	Buffer, a safe wrapper of menoh buffer control scheme.
Model	Main struct running inference.
ModelBuilder	Builder of model.
ModelData	Represent model data defined by ONNX file.
VariableProfile
VariableProfileTable	Variable Profile table.
VariableProfileTableBuilder	Builder of Variable Profile Table.

Enums

Backend
Dtype	Dtype that accepted by menoh model.
Error

Traits

DtypeCompatible

Indicate compatible type with menoh dtype