Can not get JINA CLIP v1 running

[SeanPedersen]

I am trying to get JINA CLIP v1 ONNX model to run but I am stuck with following code. Would appreciate any help.

use ort::session::{builder::GraphOptimizationLevel, Session};
use tokenizers::Tokenizer;

fn main() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
    // Load the ONNX model
    let model = Session::builder()?
        .with_optimization_level(GraphOptimizationLevel::Level3)?
        .with_intra_threads(4)?
        .commit_from_file("text_model.onnx")?;

    // Load the tokenizer
    let tokenizer = Tokenizer::from_pretrained("jinaai/jina-clip-v1", None)?;

    // Tokenize input
    let input = tokenizer.encode(
        "hello world hello world hello world hello world hello world hello world hello world hello",
        true,
    )?;
    println!("Tokenized input: {:?}", input);

    // Prepare input for ONNX inference
    let input_ids = vec![input.get_ids().to_vec()];

    // Run inference
    let outputs = model.run(ort::inputs!["text" => input_ids]?)?;
    let predictions = outputs["output0"].try_extract_tensor::<f32>()?;
    println!("Inference output: {:?}", outputs);

    Ok(())
}

Cargo.toml

[package]
name = "rust-onnx"
version = "0.1.0"
edition = "2021"

[dependencies]
ort = "=2.0.0-rc.9"
tokenizers = { version = "0.21.0", features = ["http"] }

[decahedron1]

What's the error you're encountering?

[SeanPedersen]

I get following error:

   Compiling rust-onnx v0.1.0 (/Users/Studio-MacMini/Projects/rust-onnx)
error[E0277]: the trait bound `Value: From<Vec<Vec<u32>>>` is not satisfied
  --> src/main.rs:25:29
   |
25 |     let outputs = model.run(ort::inputs!["text" => input_ids]?)?;
   |                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ the trait `From<Vec<Vec<u32>>>` is not implemented for `Value`, which is required by `Value: TryFrom<Vec<Vec<u32>>>`
   |
   = help: the following other types implement trait `From<T>`:
             `Value` implements `From<Value<DynTensorValueType>>`
             `Value` implements `From<Value<TensorValueType<T>>>`
   = note: required for `Vec<Vec<u32>>` to implement `Into<Value>`
   = note: required for `Value` to implement `TryFrom<Vec<Vec<u32>>>`
   = note: this error originates in the macro `ort::inputs` (in Nightly builds, run with -Z macro-backtrace for more info)

For more information about this error, try `rustc --explain E0277`.
error: could not compile `rust-onnx` (bin "rust-onnx") due to 1 previous error

[decahedron1]

The input IDs need to be converted to something that can be cast into a tensor first.

This can be done by creating an Arc<Box<[i64]>> from the encoded tokens:

ort/examples/gpt2/examples/gpt2-no-ndarray.rs

Line 48 in 123c449

let mut tokens = Arc::new(tokens.get_ids().iter().map(|i| *i as i64).collect::<Vec<_>>().into_boxed_slice());

and passing a (Vec<i64>, Arc<Box<[i64]>>) as an input:

ort/examples/gpt2/examples/gpt2-no-ndarray.rs

Line 56 in 123c449

let input = (vec![1, 1, tokens.len() as i64], Arc::clone(&tokens));

The exact shape of the input tensor depends on what the model expects.

[SeanPedersen]

Thx for the help, got it working:

use ort::session::{builder::GraphOptimizationLevel, Session};
use std::sync::Arc;
use tokenizers::Tokenizer;

fn main() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
    // Load the ONNX model
    let model = Session::builder()?
        .with_optimization_level(GraphOptimizationLevel::Level3)?
        .with_intra_threads(4)?
        .commit_from_file("text_model.onnx")?;

    // Load the tokenizer
    let tokenizer = Tokenizer::from_pretrained("jinaai/jina-clip-v1", None)?;

    // Tokenize input
    let input = tokenizer.encode(
        "hello world hello world hello world hello world hello world hello world hello world hello",
        true,
    )?;
    println!("Tokenized input: {:?}", input);

    // Prepare input for ONNX inference - convert to proper tensor format
    let tokens = Arc::new(
        input
            .get_ids()
            .iter()
            .map(|i| *i as i64)
            .collect::<Vec<_>>()
            .into_boxed_slice(),
    );

    // Create input with shape information - adjust shape based on model requirements
    // Common shapes are [1, sequence_length] or [batch_size, sequence_length]
    let input_shape = vec![1, tokens.len() as i64]; // Adjust shape as needed
    let model_input = (input_shape, Arc::clone(&tokens));

    // Run inference with properly formatted input
    let outputs = model.run(ort::inputs!["input_ids" => model_input]?)?;
    // Extract the embedding tensor from outputs
    let text_embeds = outputs.get("text_embeds").unwrap();

    // Convert to a Vec<f32> for easier display
    let embeds: Vec<f32> = text_embeds
        .try_extract_tensor::<f32>()
        .unwrap()
        .as_slice()
        .unwrap()
        .to_vec();

    // Print the first few values of the embedding vector
    println!("Embedding vector shape: {:?}", text_embeds.shape());
    // Print the full embedding if needed
    println!("Full embedding vector: {:?}", embeds);

    Ok(())
}