How to decrease inference time of LiLT?

[piegu]

Hi,

I'm using Hugging Face libraries in order to run LiLT.
How can I decrease inference time? Which code to use?

I've already try BetterTransformer (Optimum) and ONNX but none of them accepts LiLT model.

• BetterTransformer: NotImplementedError: The model type lilt is not yet supported to be used with BetterTransformer.
• ONNX: KeyError: "lilt is not supported yet.

Thank you.

Note: I asked this question here, too: jpWang/LiLT#42

[piegu]

Issue opened in the Optimum library: huggingface/optimum#1024

[bkocis]

Have you considered making a smaller model? What is your model size?

[NielsRogge]

One thing you can try (especially if you're using a multilingual model like https://huggingface.co/nielsr/lilt-xlm-roberta-base), then you can remove token embeddings of tokens of languages that you don't need.

See this blog post for more info: https://medium.com/@coding-otter/reduce-your-transformers-model-size-by-removing-unwanted-tokens-and-word-embeddings-eec08166d2f9