A high level framework and library for running, training, and deploying state-of-the-art Natural Language Processing (NLP) models for end to end tasks.
AdaptNLP allows users ranging from beginner python coders to experienced machine learning engineers to leverage state-of-the-art NLP models and training techniques in one easy-to-use python package.
Built atop Zalando Research's Flair and Hugging Face's Transformers library, AdaptNLP provides Machine Learning Researchers and Scientists a modular and adaptive approach to a variety of NLP tasks with an Easy API for training, inference, and deploying NLP-based microservices.
Key Features
- Full Guides and API Documentation
- Tutorial Jupyter/Google Colab Notebooks
- Unified API for NLP Tasks with SOTA Pretrained Models (Adaptable with Flair and Transformer's Models)
- Token Tagging
- Sequence Classification
- Embeddings
- Question Answering
- Summarization
- Translation
- More in development
- Training and Fine-tuning Interface
- Jeremy's ULM-FIT approach for transfer learning in NLP
- Fine-tuning Transformer's language models and task-specific predictive heads like Flair's
SequenceClassifier
- Rapid NLP Model Deployment with Sebastián's FastAPI Framework
- Containerized FastAPI app
- Immediately deploy any custom trained Flair or AdaptNLP model
- Dockerizing AdaptNLP with GPUs
- Easily build and run AdaptNLP containers leveraging NVIDIA GPUs with Docker
To avoid dependency clustering and issues, it would be wise to install AdaptNLP in a virtual environment. To create a new python 3.6+ virtual environment, run this command and then activate it however your operating system specifies:
python -m venv venv-adaptnlp
Install using pip in your virtual environment:
pip install adaptnlp
If you want to work on AdaptNLP, pip install adaptnlp[dev]
will install its development tools.
PyTorch needs to manually installed on Windows environments. If it's not already installed, proceed to http://pytorch.org/get-started/locally to select your preferences and then run the given install command. Note that the current version of PyTorch we use relies on cuda 10.1.
Install using pip:
pip install adaptnlp
If you want to work on AdaptNLP, pip install adaptnlp[dev]
will install its development tools.
Once you have installed AdaptNLP, here are a few examples of what you can run with AdaptNLP modules:
from adaptnlp import EasyTokenTagger
## Example Text
example_text = "Novetta's headquarters is located in Mclean, Virginia."
## Load the token tagger module and tag text with the NER model
tagger = EasyTokenTagger()
sentences = tagger.tag_text(text=example_text, model_name_or_path="ner")
## Output tagged token span results in Flair's Sentence object model
for sentence in sentences:
for entity in sentence.get_spans("ner"):
print(entity)
from adaptnlp import EasySequenceClassifier
## Example Text
example_text = "Novetta is a great company that was chosen as one of top 50 great places to work!"
## Load the sequence classifier module and classify sequence of text with the english sentiment model
classifier = EasySequenceClassifier()
sentences = classifier.tag_text(text=example_text, mini_batch_size=1, model_name_or_path="en-sentiment")
## Output labeled text results in Flair's Sentence object model
for sentence in sentences:
print(sentence.labels)
from adaptnlp import EasyQuestionAnswering
## Example Query and Context
query = "What is the meaning of life?"
context = "Machine Learning is the meaning of life."
top_n = 5
## Load the QA module and run inference on results
qa = EasyQuestionAnswering()
best_answer, best_n_answers = qa.predict_qa(query=query, context=context, n_best_size=top_n, mini_batch_size=1, model_name_or_path="distilbert-base-uncased-distilled-squad")
## Output top answer as well as top 5 answers
print(best_answer)
print(best_n_answers)
from adaptnlp import EasySummarizer
# Text from encyclopedia Britannica on Einstein
text = """Einstein would write that two “wonders” deeply affected his early years. The first was his encounter with a compass at age five.
He was mystified that invisible forces could deflect the needle. This would lead to a lifelong fascination with invisible forces.
The second wonder came at age 12 when he discovered a book of geometry, which he devoured, calling it his 'sacred little geometry
book'. Einstein became deeply religious at age 12, even composing several songs in praise of God and chanting religious songs on
the way to school. This began to change, however, after he read science books that contradicted his religious beliefs. This challenge
to established authority left a deep and lasting impression. At the Luitpold Gymnasium, Einstein often felt out of place and victimized
by a Prussian-style educational system that seemed to stifle originality and creativity. One teacher even told him that he would
never amount to anything."""
summarizer = EasySummarizer()
# Summarize
summaries = summarizer.summarize(text = text, model_name_or_path="t5-small", mini_batch_size=1, num_beams = 4, min_length=0, max_length=100, early_stopping=True)
print("Summaries:\n")
for s in summaries:
print(s, "\n")
from adaptnlp import EasyTranslator
text = ["Machine learning will take over the world very soon.",
"Machines can speak in many languages.",]
translator = EasyTranslator()
# Translate
translations = translator.translate(text = text, t5_prefix="translate English to German", model_name_or_path="t5-small", mini_batch_size=1, min_length=0, max_length=100, early_stopping=True)
print("Translations:\n")
for t in translations:
print(t, "\n")
from adaptnlp import EasyDocumentEmbeddings, SequenceClassifierTrainer
# Specify corpus data directory and model output directory
corpus = "Path/to/data/directory"
OUTPUT_DIR = "Path/to/output/directory"
# Instantiate AdaptNLP easy document embeddings module, which can take in a variable number of embeddings to make `Stacked Embeddings`.
# You may also use custom Transformers LM models by specifying the path the the language model
doc_embeddings = EasyDocumentEmbeddings(model_name_or_path="bert-base-cased", methods = ["rnn"])
# Instantiate Sequence Classifier Trainer by loading in the data, data column map, and embeddings as an encoder
sc_trainer = SequenceClassifierTrainer(corpus=corpus, encoder=doc_embeddings, column_name_map={0: "text", 1:"label"})
# Find Learning Rate
learning_rate = sc_trainer.find_learning_rate(output_dir=OUTPUT_DIR)
# Train Using Flair's Sequence Classification Head
sc_trainer.train(output_dir=OUTPUT_DIR, learning_rate=learning_rate, max_epochs=150)
# Predict text labels with the trained model using `EasySequenceClassifier`
from adaptnlp import EasySequenceClassifier
example_text = '''Where was the Queen's wedding held? '''
classifier = EasySequenceClassifier()
sentences = classifier.tag_text(example_text, model_name_or_path=OUTPUT_DIR / "final-model.pt")
print("Label output:\n")
for sentence in sentences:
print(sentence.labels)
from adaptnlp import LMFineTuner
# Specify Text Data File Paths
train_data_file = "Path/to/train.csv"
eval_data_file = "Path/to/test.csv"
# Instantiate Finetuner with Desired Language Model
finetuner = LMFineTuner(train_data_file=train_data_file, eval_data_file=eval_data_file, model_type="bert", model_name_or_path="bert-base-cased")
finetuner.freeze()
# Find Optimal Learning Rate
learning_rate = finetuner.find_learning_rate(base_path="Path/to/base/directory")
finetuner.freeze()
# Train and Save Fine Tuned Language Models
finetuner.train_one_cycle(output_dir="Path/to/output/directory", learning_rate=learning_rate)
Look in the Tutorials directory for a quick introduction to the library and its very simple and straight forward use cases:
NLP Tasks
- Token Classification: NER, POS, Chunk, and Frame Tagging
- Sequence Classification: Sentiment
- Embeddings: Transformer Embeddings e.g. BERT, XLM, GPT2, XLNet, roBERTa, ALBERT
- Question Answering: Span-based Question Answering Model
- Summarization: Abstractive and Extractive
- Translation: Seq2Seq
Custom Fine-Tuning and Training with Transformer Models
Checkout the documentation for more information.
We use FastAPI for standing up endpoints for serving state-of-the-art NLP models with AdaptNLP.
The REST directory contains more detail on deploying a REST API locally or with docker in a very easy and fast way.
AdaptNLP official docker images are up on Docker Hub.
Images have AdaptNLP installed from source in developer mode with tutorial notebooks available.
Images can build with GPU support if NVIDA-Docker is correctly installed.
Simply run an image with AdaptNLP installed from source in developer mode by running:
docker run -it --rm achangnovetta/adaptnlp:latest
Run an image with AdaptNLP running on GPUs if you have nvidia drivers and nvidia-docker 19.03+ installed:
docker run -it --rm --gpus all achangnovetta/adaptnlp:latest
Build docker image and run container with the following commands in the directory of the Dockerfile to create a container with adaptnlp installed and ready to go
Note: A container with GPUs enabled requires Docker version 19.03+ and nvida-docker installed
docker build -t achangnovetta/adaptnlp:latest .
docker run -it --rm achangnovetta/adaptnlp:latest
If you want to use CUDA compatible GPUs
docker run -it --rm --gpus all achangnovetta/adaptnlp:latest
Please contact the author Andrew Chang at [email protected] with questions or comments regarding AdaptNLP.
Follow us on Twitter at @achang1618 and @AdaptNLP for updates and NLP dialogue.
This project is licensed under the terms of the Apache 2.0 license.