NLP support with Huggingface tokenizers¶

This module contains the NLP support with Huggingface tokenizers implementation.

This is an implementation from Huggingface tokenizers RUST API.

Documentation¶

The latest javadocs can be found on here.

You can also build the latest javadocs locally using the following command:

./gradlew javadoc

The javadocs output is built in the build/doc/javadoc folder.

Installation¶

You can pull the module from the central Maven repository by including the following dependency in your pom.xml file:

<dependency>
    <groupId>ai.djl.huggingface</groupId>
    <artifactId>tokenizers</artifactId>
    <version>0.30.0</version>
</dependency>

Usage¶

Use DJL HuggingFace model converter¶

If you are trying to convert a complete HuggingFace (transformers) model, you can try to use our all-in-one conversion solution to convert to Java:

Currently, this converter supports the following tasks:

fill-mask
question-answering
sentence-similarity
text-classification
token-classification

Install `djl-converter`¶

You can install djl-converter from djl master branch or clone the repository and install from source:

# install release version of djl-converter
pip install https://publish.djl.ai/djl_converter/djl_converter-0.30.0-py3-none-any.whl

# install from djl master branch
pip install "git+https://github.com/deepjavalibrary/djl.git#subdirectory=extensions/tokenizers/src/main/python"

# install djl-convert from local djl repo
git clone https://github.com/deepjavalibrary/djl.git
cd djl/extensions/tokenizers/src/main/python
python3 -m pip install -e .

# install optimum if you want to convert to OnnxRuntime
pip install optimum

# convert a single model to TorchScript, Onnxruntime or Rust
djl-convert --help

# import models as DJL Model Zoo
djl-import --help

Convert Huggingface model to torchscript¶

djl-convert -m deepset/bert-base-cased-squad2

This will find converted model in model/bert-base-cased-squad2/ folder:

djl-convert -m deepset/bert-base-cased-squad2

Convert Huggingface model to OnnxRuntime¶

djl-convert -m deepset/bert-base-cased-squad2 -f OnnxRuntime

Convert Huggingface model to Rust¶

djl-convert -m deepset/bert-base-cased-squad2 -f Rust

Load converted model¶

Then, all you need to do, is to load this model in DJL:

Criteria<QAInput, String> criteria = Criteria.builder()
        .setTypes(QAInput.class, String.class)
        .optModelPath(Paths.get("model/bert-base-cased-squad2/"))
        .optTranslatorFactory(new DeferredTranslatorFactory())
        .optProgress(new ProgressBar()).build();

Import multiple Huggingface Hub models into DJL model zoo¶

djl-import -m deepset/bert-base-cased-squad2

This will generate a zip file into your local djl model zoo folder structure:

model/nlp/question_answer/ai/djl/huggingface/pytorch/deepset/bert-base-cased-squad2/0.0.1/bert-base-cased-squad2.zip

From HuggingFace AutoTokenizer¶

In most of the cases, you can easily use a pre-existing tokenizer in DJL:

Python

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/msmarco-distilbert-dot-v5")

Java

HuggingFaceTokenizer tokenizer = HuggingFaceTokenizer.newInstance("sentence-transformers/msmarco-distilbert-dot-v5");

This way requires network connection to huggingface repo. The way to determine if you can use this way is through looking into the "Files and versions" in HuggingFace model tab and see if there is a tokenizer.json.

If there is a tokenizer.json, you can get it directly through DJL. Otherwise, use the other way below to obtain a tokenizer.json.

From HuggingFace Pipeline¶

If you are trying to get tokenizer from a HuggingFace pipeline, you can use the followings to extract tokenizer.json file.