Feature Request: huggingface model converter about huggingface.js HOT 9 CLOSED

Optimum is an open source library that lets one do many of those conversions locally (it's powered most/all of those Spaces):

• https://huggingface.co/docs/optimum/index
• https://github.com/huggingface/optimum

cc @mfuntowicz!

from huggingface.js.

@xenova you have just fully defined a JavaScript Dev's nightmare.

"Just load Gradio, to run the Python code, in an App, that can be put in an iFrame, in the web browser." LOL

On a more serious level it does look do-able, I will check around for some similar working examples. I need to be able to:

1. load a tensorflowJS model in .json format with it's .bin binary shard files
2. convert the model, let's say to Keras just to test. I am used to the command line but probably better to use a python script. Here is the link to the tensorflowjs-converter https://github.com/tensorflow/tfjs/tree/master/tfjs-converter
3. Download the converted model

If I can get that working I can probably do the entire process.

I am a bit confused by the whole PR thing the other converter examples did. Does the code send a PR which then automatically does the conversion? Seems strange. Can the pipeline just do what I listed above or does it need to do the whole PR thing?

from huggingface.js.

cc @xenova 👀

from huggingface.js.

we have a few converters on the Hub already, for instance:

• https://huggingface.co/spaces/safetensors/convert
• https://huggingface.co/spaces/onnx/export
• https://huggingface.co/spaces/huggingface-projects/transformers-to-coreml

so definitely open to building more!

from huggingface.js.

@hpssjellis Hi there! 👋 as mentioned above, we have a few converters already available as spaces on the Hub, but the only "JS-ready" format would be ONNX (thanks to onnxruntime-web).

From your first message, it seems like quite a specific use-case: starting with a TensorflowJS model and ending with a C-header file (with many steps in the middle).

That said, it might be worth looking into building a HF space that handles the full pipeline. Is that something you'd be interested in? As a follow up, since you seem to already have a dockerfile/setup to do it all, might I suggest you create a HF space yourself for it? Considering the typical size of these models (i.e., quite small), I am quite sure you won't need anything above the basic (free) version.

from huggingface.js.

Thanks @julien-c and @xenova those look great. I will definitely look into HF space. Yes my use case is very specific but also very helpful for moving ML models to micro-controllers. Here is an example if you are interested . Works on a cell phone without a connected device since the Cell Phone Motion button captures the x,y,z from the cell phone.

https://hpssjellis.github.io/tinyMLjs/public/acceleration/a00-best-acceleration.html

If HF Space can make a pipeline that I can drop any tensorflowJS file and output a c-header file that would be really helpful.

from huggingface.js.

If HF Space can make a pipeline that I can drop any tensorflowJS file and output a c-header file that would be really helpful.

That should be possible 👍 I recommend that you check out Gradio, which seamlessly integrates with Hugging Face spaces (docs). It handles most of the things you are looking to add (e.g., file uploads).

Also, since Gradio runs in python, you can directly use the tensorflowjs converter you linked above. Alternatively, you can just set up a list of os.system calls - whichever works.

from huggingface.js.

"Just load Gradio, to run the Python code, in an App, that can be put in an iFrame, in the web browser." LOL

😅 Haha, yeah I totally understand. The Gradio docs are quite extensive, and the above example spaces should help provide you with some idea on how to get started.

I am a bit confused by the whole PR thing the other converter examples did. Does the code send a PR which then automatically does the conversion? Seems strange. Can the pipeline just do what I listed above or does it need to do the whole PR thing?

The reason they submit a PR is just to contribute the converted model the original repo (for others to use too). In your case, probably the easiest would be to create and upload the converted files to a new model repo (under your account, either private or public). For that you can use the Hub python library to do that: https://huggingface.co/docs/huggingface_hub/guides/upload

from huggingface.js.

we have a few converters on the Hub already, for instance:

• https://huggingface.co/spaces/safetensors/convert
• https://huggingface.co/spaces/onnx/export
• https://huggingface.co/spaces/huggingface-projects/transformers-to-coreml

so definitely open to building more!

@julien-c @xenova these converters seem awesome but they are all cloud side (which makes sense for Hugggingface). My group often has to download a model and then work in the field or just without great internet. I am looking for ways to convert models such as downloaded Huggingface models but also models made by the students offsite. Does anyone have any suggestions:

I have found these sites

https://www.tensorflow.org/lite/models/convert

https://github.com/hhsecond/ml2rt

https://www.freecodecamp.org/news/transform-machine-learning-models-into-native-code-with-zero-dependencies/

https://madewithml.com/courses/foundations/utilities/

Much like my Javascript questions which ended with this very useful set of static webpages I am trying to let students manipulate locally the datasets and models that they have access to.

from huggingface.js.