error[E0277]: the trait bound ort::Value: From<ArrayBase<OwnedRepr<f32>, Dim<[usize; 4]>>> is not satisfied

[Esword618]

I encountered an error when using ort 2.0.0-rc.4:

error[E0277]: the trait bound `ort::Value: From<ArrayBase<ViewRepr<&f32>, Dim<[usize; 4]>>>` is not satisfied
   --> src/main.rs:120:55
    |
120 |     let outputs: SessionOutputs = encoder_session.run(inputs!["pixel_values" => inputs.view()]?)?;
    |                                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ the trait `From<ArrayBase<ViewRepr<&f32>, Dim<[usize; 4]>>>` is not implemented for `ort::Value`, which is required by `ort::Value: TryFrom<ArrayBase<ViewRepr<&f32>, Dim<[usize; 4]>>>`
    |
    = help: the following other types implement trait `From<T>`:
              <ort::Value as From<ort::Value<DynTensorValueType>>>
              <ort::Value as From<ort::Value<TensorValueType<T>>>>
    = note: required for `ArrayBase<ViewRepr<&f32>, Dim<[usize; 4]>>` to implement `Into<ort::Value>`
    = note: required for `ort::Value` to implement `TryFrom<ArrayBase<ViewRepr<&f32>, Dim<[usize; 4]>>>`
    = note: this error originates in the macro `inputs` (in Nightly builds, run with -Z macro-backtrace for more info)

The following is my code:

use std::path::Path;
use tokenizers::tokenizer::{self, Result, Tokenizer, EncodeInput, TokenizerBuilder, AddedToken};
use tokenizers::models::bpe::{BPE, BpeBuilder, Merges, Vocab};
use serde::Deserialize;
use image::{self, GenericImageView, imageops::FilterType};
use ndarray::{Array, Array3, };
use std::fs::File;
use std::io::BufReader;
use ort::{inputs, Tensor, Value, GraphOptimizationLevel, Session, SessionOutputs, CPUExecutionProvider};

// 定义配置结构
#[derive(Deserialize)]
struct Size {
    height: u32,
    width: u32,
}

#[derive(Deserialize)]
struct PreprocessorConfig {
    do_normalize: bool,
    do_rescale: bool,
    do_resize: bool,
    image_processor_type: String,
    resample: u8, // 假设这里是一个 enum 生成的值，例如0到4
    rescale_factor: f32,
    size: Size,
}

fn main() -> Result<()> {
    // ==================图片特征提取===================
    // 读取配置文件
    let file = File::open("./onnx/preprocessor_config.json")?;
    let reader = BufReader::new(file);
    let config: PreprocessorConfig = serde_json::from_reader(reader)?;

    // 加载图像
    let img = image::open("img.png")?;

    let filter = match config.resample {
        0 => FilterType::Nearest,
        1 => FilterType::Triangle,
        2 => FilterType::CatmullRom,
        3 => FilterType::Gaussian,
        4 => FilterType::Lanczos3,
        _ => return Err("Invalid resample value".into()),
    };

    let img_resized = if config.do_resize {
        img.resize_exact(config.size.width, config.size.height, filter)
    } else {
        img
    };

    let mut img_ndarray = Array3::<f32>::zeros((config.size.height as usize, config.size.width as usize, 3));
    for (x, y, pixel) in img_resized.to_rgb8().enumerate_pixels() {
        let channels = pixel.0;
        img_ndarray[(y as usize, x as usize, 0)] = channels[0] as f32;
        img_ndarray[(y as usize, x as usize, 1)] = channels[1] as f32;
        img_ndarray[(y as usize, x as usize, 2)] = channels[2] as f32;
    }

    if config.do_rescale {
        img_ndarray = &img_ndarray * config.rescale_factor;
    }

    if config.do_normalize {
        // 这里假设标准化的方法是减去均值并除以标准差，用户可以根据实际需要调整
        let mean = img_ndarray.mean().unwrap_or(0.0);
        let std = img_ndarray.std(0.0);
        img_ndarray = (&img_ndarray - mean) / std;
    }
    img_ndarray = img_ndarray.permuted_axes([2, 0, 1]);
    let inputs  = img_ndarray.insert_axis(ndarray::Axis(0));

    println!("{:?}", inputs.shape());

    // ==================tokenizer加载===================
    let bpe_builder = BPE::from_file("./onnx/vocab.json", "./onnx/merges.txt");
    let bpe = bpe_builder
        .dropout(0.1)
        // .unk_token("[UNK]".into())
        .build()?;

    let mut tokenizer = Tokenizer::new(bpe);
    tokenizer.add_special_tokens(&*vec![
        AddedToken::from(String::from("<s>"), true).normalized(true),
        AddedToken::from(String::from("</s>"), true).normalized(true),
        AddedToken::from(String::from("<mask>"), true).normalized(true),
        AddedToken::from(String::from("<pad>"), true).normalized(true),
        AddedToken::from(String::from("<unk>"), true).normalized(true),
    ]);

    // let encoding = tokenizer.encode("My name is Esword!", false)?;
    // println!("{:?}", encoding.get_tokens());

    // ==================模型加载===================

    // Initialize tracing to receive debug messages from `ort`
    tracing_subscriber::fmt::init();

    // Create the ONNX Runtime environment, enabling CUDA execution providers for all sessions created in this process.
    ort::init()
        .with_name("Pix2Text")
        .with_execution_providers([CPUExecutionProvider::default().build()])
        .commit()?;

    let encoder_session = Session::builder()?
        .with_optimization_level(GraphOptimizationLevel::Level1)?
        .with_intra_threads(4)?
        .commit_from_file("onnx/encoder_model.onnx")?;

    let decoder_session = Session::builder()?
        .with_optimization_level(GraphOptimizationLevel::Level1)?
        .with_intra_threads(4)?
        .commit_from_file("onnx/decoder_model.onnx")?;


    // let cc = Value::from_array(&inputs)?;

    let outputs: SessionOutputs = encoder_session.run(inputs!["pixel_values" => inputs.view()]?)?;

    // let original_img = image::open(Path::new("img.png")).unwrap();
    // let (img_width, img_height) = (original_img.width(), original_img.height());
    // let img = original_img.resize_exact(448, 448, FilterType::Gaussian);
    // let mut input3 = Array::zeros((1, 3, 448, 448));
    // for pixel in img.pixels() {
    //     let x = pixel.0 as _;
    //     let y = pixel.1 as _;
    //     let [r, g, b, _] = pixel.2.0;
    //     input3[[0, 0, y, x]] = (r as f32) / 255.;
    //     input3[[0, 1, y, x]] = (g as f32) / 255.;
    //     input3[[0, 2, y, x]] = (b as f32) / 255.;
    // }
    // println!("{:?}", input3.shape());
    // let outputs: SessionOutputs = encoder_session.run(inputs!["pixel_values" => input3.view()]?)?;
    // let outputs: SessionOutputs = encoder_session.run(inputs![input3]?)?;
    
    Ok(())
}

[decahedron1]

Make sure you're using version 0.15 of the ndarray crate, and that you've enabled the ndarray feature for ort if you disabled its default features.

[Esword618]

OK, I understand. Thank you!