This package leverages the power of the 🤗Tokenizers library (built with Rust) to process the input text. It then uses TensorFlow.js to run the DistilBERT-cased model fine-tuned for Question Answering (87.1 F1 score on SQuAD v1.1 dev set, compared to 88.7 for BERT-base-cased). DistilBERT is used by default, but you can use other models available in the 🤗Transformers library in one additional line of code!
It can run models in SavedModel and TFJS formats locally, as well as remote models thanks to TensorFlow Serving.
npm install question-answering@latestThe following example will automatically download the default DistilBERT model in SavedModel format if not already present, along with the required vocabulary / tokenizer files. It will then run the model and return the answer to the question.
import { QAClient } from "question-answering"; // When using Typescript or Babel
// const { QAClient } = require("question-answering"); // When using vanilla JS
const text = `
Super Bowl 50 was an American football game to determine the champion of the National Football League (NFL) for the 2015 season.
The American Football Conference (AFC) champion Denver Broncos defeated the National Football Conference (NFC) champion Carolina Panthers 24–10 to earn their third Super Bowl title. The game was played on February 7, 2016, at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California.
As this was the 50th Super Bowl, the league emphasized the "golden anniversary" with various gold-themed initiatives, as well as temporarily suspending the tradition of naming each Super Bowl game with Roman numerals (under which the game would have been known as "Super Bowl L"), so that the logo could prominently feature the Arabic numerals 50.
`;
const question = "Who won the Super Bowl?";
const qaClient = await QAClient.fromOptions();
const answer = await qaClient.predict(question, text);
console.log(answer); // { text: 'Denver Broncos', score: 0.3 }You can also download the model and vocabulary / tokenizer files separately by using the CLI.
The above example internally makes use of the default DistilBERT-cased model in the SavedModel format. The library is also compatible with any other DistilBERT-based model, as well as any BERT-based and RoBERTa-based models, both in SavedModel and TFJS formats. The following models are available in SavedModel format from the Hugging Face model hub thanks to the amazing NLP community 🤗:
a-ware/mobilebert-squadv2a-ware/roberta-large-squadv2bert-large-cased-whole-word-masking-finetuned-squadbert-large-uncased-whole-word-masking-finetuned-squaddeepset/bert-base-cased-squad2deepset/bert-large-uncased-whole-word-masking-squad2deepset/roberta-base-squad2distilbert-base-cased-distilled-squad(default) (also available in TFJS format)distilbert-base-uncased-distilled-squadhenryk/bert-base-multilingual-cased-finetuned-dutch-squad2ktrapeznikov/biobert_v1.1_pubmed_squad_v2ktrapeznikov/scibert_scivocab_uncased_squad_v2mrm8488/bert-base-spanish-wwm-cased-finetuned-spa-squad2-esmrm8488/distill-bert-base-spanish-wwm-cased-finetuned-spa-squad2-esmrm8488/spanbert-finetuned-squadv2NeuML/bert-small-cord19qatwmkn9/bert-base-uncased-squad2
To specify a model to use with the library, you need to instantiate a model class that you'll then pass to the QAClient:
import { initModel, QAClient } from "question-answering"; // When using Typescript or Babel
// const { initModel, QAClient } = require("question-answering"); // When using vanilla JS
const text = ...
const question = ...
const model = await initModel({ name: "deepset/roberta-base-squad2" });
const qaClient = await QAClient.fromOptions({ model });
const answer = await qaClient.predict(question, text);
console.log(answer); // { text: 'Denver Broncos', score: 0.46 }Note that using a model hosted on Hugging Face is not a requirement: you can use any compatible model (including any from the HF hub not already available in SavedModel or TFJS format that you converted yourself) by passing the correct local path for the model and vocabulary files in the options.
To use a TFJS model, you simply need to pass tfjs to the runtime param of initModel (defaults to saved_model):
const model = await initModel({ name: "distilbert-base-cased-distilled-squad", runtime: RuntimeType.TFJS });As with any SavedModel hosted in the HF model hub, the required files for the TFJS models will be automatically downloaded the first time. You can also download them manually using the CLI.
Using remote models with TensorFlow Serving
If your model is in the SavedModel format, you may prefer to host it on a dedicated server. Here is a simple example using Docker locally:
# Inside our project root, download DistilBERT-cased to its default `.models` location
npx question-answering download
# Download the TensorFlow Serving Docker image
docker pull tensorflow/serving
# Start TensorFlow Serving container and open the REST API port.
# Notice that in the `target` path we add a `/1`:
# this is required by TFX which is expecting the models to be "versioned"
docker run -t --rm -p 8501:8501 \
--mount type=bind,source="$(pwd)/.models/distilbert-cased/",target="/models/cased/1" \
-e MODEL_NAME=cased \
tensorflow/serving &In the code, you just have to pass remote as runtime and the server endpoint as path:
const model = await initModel({
name: "distilbert-base-cased-distilled-squad",
path: "http://localhost:8501/v1/models/cased",
runtime: RuntimeType.Remote
});
const qaClient = await QAClient.fromOptions({ model });You can choose to download the model and associated vocab file(s) manually using the CLI. For example to download the deepset/roberta-base-squad2 model:
npx question-answering download deepset/roberta-base-squad2By default, the files are downloaded inside a
.modelsdirectory at the root of your project; you can provide a custom directory by using the--diroption of the CLI. You can also use--format tfjsto download a model in TFJS format (if available). To check all the options of the CLI:npx question-answering download --help.
The QAClient.fromOptions params object has a tokenizer field which can either be a set of options relative to the tokenizer files, or an instance of a class extending the abstract Tokenizer class. To extend this class, you can create your own or, if you simply need to adjust some options, you can import and use the provided initTokenizer method, which will instantiate such a class for you.
Thanks to the native execution of SavedModel format in TFJS, the performance of such models is similar to the one using TensorFlow in Python.
Specifically, here are the results of a benchmark using question-answering with the default DistilBERT-cased model:
- Running entirely locally (both SavedModel and TFJS formats)
- Using a (pseudo) remote model server (i.e. local Docker with TensorFlow Serving running the SavedModel format)
- Using the Question Answering pipeline in the 🤗Transformers library.
Shorts texts are texts between 500 and 1000 characters, long texts are between 4000 and 5000 characters. You can check the
question-answering benchmark script here (the transformers one is equivalent). Benchmark run on a standard 2019 MacBook Pro running on macOS 10.15.2.
There is a known incompatibility in the TFJS library with some types. If you encounter errors when building your project, make sure to pass the --skipLibCheck flag when using the Typescript CLI, or to add skipLibCheck: true to your tsconfig.json file under compilerOptions. See here for more information.
Due to a bug in TFJS, the use of @tensorflow/tfjs-node to load or execute SavedModel models independently from the library is not recommended for now, since it could overwrite the TF backend used internally by the library. In the case where you would have to do so, make sure to require both question-answering and @tensorflow/tfjs-node in your code before making any use of either of them.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent