Joshua Lochner

[New task] Visual Question Answering

2

### Feature request Currently, the [visual-question-answering pipeline/task](https://huggingface.co/tasks/visual-question-answering) in transformers is not supported for onnx export: https://github.com/huggingface/optimum/blob/618a483d463b9c2e0b0ba3b859ff74c6279b5161/optimum/exporters/tasks.py#L147-L170 Here are some models which can be used for testing: https://huggingface.co/models?other=visual-question-answering ### Motivation This...

Create marqo-fashionCLIP demo

1

Model: https://huggingface.co/Marqo/marqo-fashionCLIP

Disable pretty-print when saving tokenizer.json files

1

### Feature request As the vocabulary of newer models, like Llama 3 or Gemma, increases in size, so does the size of the tokenizer, which includes the vocabulary as JSON...

[bug] Incorrect results for floating point numbers. `p.number_to_words(0.000001) != 'one hundred and six'`

2

When converting numbers like 0.000001 to strings, python prints it as 1e-06, leading to problems with the result. For example: ```py >>> import inflect >>> p = inflect.engine() >>> p.number_to_words(0.000001)...

WebGPU fixes

1

Makes multi-session management much easier, adding a global web inference lock to ensure only one session can run at any given time, preventing "Session has already started" errors.

[V3 proposal] Improved defaults for quantization and device selection

5

### Feature request Currently, Transformers.js V3 defaults to use CPU (WASM) instead of GPU (WebGPU) due to lack of support and instability across browsers (specifically Firefox and Safari, and Chrome...

Great library! I inspected the models with [Netron](https://netron.app/), and noticed there was some room for optimization. Using a tool called [onnxslim](https://github.com/inisis/OnnxSlim), you can remove a bunch of redundant nodes (nothing...

[WIP] Electron demo for Transformers.js v3

4

Running with `npm run start` works well, but I'm still running into issues when building the application with `npm run make`. First: > Uncaught Exception: Error: dlopen(.../electron.app/Contents/Resources/app.asar.unpacked/node_modules/onnxruntime-node/bin/napi-v3/darwin/arm64/onnxruntime_binding.node, 0x0001): Library not...

[ONNX] Add dynamic shapes support (& in-browser inference w/ Transformers.js)

13

# Description This PR adds support for exporting RF-DETR to ONNX with dynamic input and output shapes. This means you can now supply images of variable batch_size, height, and width,...