Add support for TrOCR Model

[ToluClassics]

Add Support for TrOCR(https://huggingface.co/microsoft/trocr-base-printed) Model

[LaurentMazare]

Looks like a great addition, thanks! Just one small nitpick.

[LaurentMazare]

Merged, thanks a lot! Would be great to advertise this new model a bit if you felt like posting on twitter / reddit / discord so that potential users are aware of its existance.

[katopz]

Thanks for nice addition! Anyway I have some issue for microsoft/trocr-large-printed when I try it with

    let vb = {
        let model = match args.model {
            Some(model) => std::path::PathBuf::from(model),
            None => match args.which {
                // Somehow this one is 404 
                Which::Base => Api::new()?
                    .repo(hf_hub::Repo::with_revision(
                        "microsoft/trocr-base-printed".to_string(),
                        hf_hub::RepoType::Model,
                        "main".to_string(),
                    ))
                    .get("model.safetensors")?,
                Which::Large => Api::new()?
                    .repo(hf_hub::Repo::with_revision(
                        "microsoft/trocr-large-printed".to_string(),
                        hf_hub::RepoType::Model,
                        "main".to_string(),
                    ))
                    .get("model.safetensors")?,
            },
        };
        println!("model: {:?}", model);
        unsafe { VarBuilder::from_mmaped_safetensors(&[model], DType::F32, &device)? }
    };

and

cargo run --example trocr --release --  --which large --cpu --image candle-examples/examples/trocr/assets/trocr.png

i got an error

Error: shape mismatch for encoder.embeddings.cls_token, expected: [1, 1, 768], got: [1, 1, 1024]

Not sure what I miss?

[ToluClassics]

Hi @katopz , I think we might be missing the encoder configurations for the large model.
Here's where the encoder config is for base

candle/candle-transformers/src/models/vit.rs

Line 38 in 6d83d42

pub fn microsoft_trocr_base_handwritten() -> Self {

[katopz]

Hi @katopz , I think we might be missing the encoder configurations for the large model. Here's where the encoder config is for base

candle/candle-transformers/src/models/vit.rs

Line 38 in 6d83d42

pub fn microsoft_trocr_base_handwritten() -> Self {

Oh, right! Anyway I hit another error after try to change hiden_size to 1024. 🤔

Error: shape mismatch for encoder.encoder.layer.0.attention.attention.query.weight, expected: [1020, 1024], got: [1024, 1024]