Couple of questions about multilingual support and ONNX deployment

[carlos-chourio]

Hello!

Sorry in advance if my questions are too basic.
I’m new to AI and still missing some base knowledge.

I have two quick questions:

1. Does the model support multilingual data?

My guess is that if I replace the English-only BERT with a multilingual version, it should work, but I’m not entirely sure if there are limitations or additional steps required.

2. Can the model be deployed using another inference provider?

I’m building an application in C#, where running ONNX models is straightforward.
Is it possible to train the adaptive classifier in Python and then export the final model to ONNX for inference in a .NET application?
Or does the architecture rely on something that would not translate well to an ONNX runtime?

Thank you in advance!

And seriously. This is such an amazing work.
This project looks incredible. Thank you for your contributions to the community

[codelion]

Hi,

1. Yes you can use a base model with multilingual version and it would work. You can see an example here - https://github.com/codelion/adaptive-classifier/blob/main/examples/advanced_usage.py#L250 that uses distilbert/distilbert-base-multilingual-cased

2. You can use ONNX, in fact when you push to HF, by default we will convert all the models to ONNX and if when you use them on CPU they will default to using ONNX for inference. See https://huggingface.co/adaptive-classifier/browsesafe/tree/main/onnx or any other model on the adaptive-classifier org on HF.

[carlos-chourio]

Hello. First of all thank you for your help.

I successfully trained a model using distilbert-base-multilingual-cased and I am getting great results (~80% accuracy) in Python with my testing dataset.

Now, I am trying to use this trained model in a C# application using ONNX, but I cannot get the predictions to match the Python results.

What I have done so far:

I exported the model to ONNX and grabbed the examples.json file generated by the save method.

• In C#, I am using the same distilbert tokenizer. I verified that the Token IDs in C# match the Python ones exactly.
• I run the ONNX model and get the output tensor.
• I try to find the class by calculating the average of the output vectors and comparing them to the vectors inside examples.json using Cosine Similarity, I have also tried using centroids.

The Problem I am facing:

The predictions in C# are completely different from the Python predict() method, and the confidence scores are also wrong

My Question:

I suspect I am calculating the embeddings wrong in C#.

• Could you help me a bit to understand the logic behind the predict() function?
• Do I just average the ONNX output?, Do I compare against the examples.json directly, or is there another step (like extra weights or layers) I need to apply first?

Any help on how to manually calculate the prediction from the ONNX output would be really helpful!

Thank you very much in advance,

[carlos-chourio]

Hello again,

I have some good news.
I successfully managed to replicate the inference logic in C#

After studying the source code (specifically the predict method in classifier.py, memory.py and model.py for the AdaptiveHead). I realized that my previous approach was missing two key components:

1. Switching from Cosine Similarity to Squared Euclidean Distance for the prototype memory comparisons.
2. Exporting the Adaptive Head weights and blending the neural network's predictions with the memory predictions (using the 70/30 weighted split) (although for my data I noticed that 80/20 worked like 1% better).

Once I implemented those changes, my C# results matched almost entirely the Python results (There is a slight 0.4% difference in results but I guess it has something to do with using float32 for inference).

Since this might be useful for others looking to run your models in a .NET/ONNX environment, I was thinking of creating a separate repository demonstrating this C# implementation. I wanted to ask first if you are okay with me publishing that?

Thank you again for this excellent library and your contributions to the community.

[codelion]

Glad it worked out, and yes feel free to create the repo with the C# implementation.