Scikit-learn style model finetuning for NLP

Finetune is a library that allows users to leverage state-of-the-art pretrained NLP models for a wide variety of downstream tasks.

Finetune currently supports TensorFlow implementations of the following models:

1. BERT, from "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding"
2. RoBERTa, from "RoBERTa: A Robustly Optimized BERT Pretraining Approach"
3. GPT, from "Improving Language Understanding by Generative Pre-Training"
4. GPT2, from "Language Models are Unsupervised Multitask Learners"
5. TextCNN, from "Convolutional Neural Networks for Sentence Classification"
6. Temporal Convolution Network, from "An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling"
7. DistilBERT from "Smaller, faster, cheaper, lighter: Introducing DistilBERT, a distilled version of BERT"

Section	Description
API Tour	Base models, configurables, and more
Installation	How to install using pip or directly from source
Finetune with Docker	Finetune and inference within a Docker Container
Documentation	Full API documentation

Finetune API Tour

Finetuning the base language model is as easy as calling Classifier.fit:

model = Classifier()               # Load base model
model.fit(trainX, trainY)          # Finetune base model on custom data
model.save(path)                   # Serialize the model to disk
...
model = Classifier.load(path)      # Reload models from disk at any time
predictions = model.predict(testX) # [{'class_1': 0.23, 'class_2': 0.54, ..}, ..]

Choose your desired base model from finetune.base_models:

from finetune.base_models import BERT, RoBERTa, GPT, GPT2, TextCNN, TCN
model = Classifier(base_model=BERT)

Optimize your model with a variety of configurables. A detailed list of all config items can be found in the finetune docs.

model = Classifier(low_memory_mode=True, lr_schedule="warmup_linear", max_length=512, l2_reg=0.01, oversample=True, ...)

The library supports finetuning for a number of tasks. A detailed description of all target models can be found in the finetune API reference.

from finetune import *
models = (Classifier, MultiLabelClassifier, MultiFieldClassifier, MultipleChoice, # Classify one or more inputs into one or more classes
          Regressor, OrdinalRegressor, MultifieldRegressor,                       # Regress on one or more inputs
          SequenceLabeler, Association,                                           # Extract tokens from a given class, or infer relationships between them
          Comparison, ComparisonRegressor, ComparisonOrdinalRegressor,            # Compare two documents for a given task
          LanguageModel, MultiTask,                                               # Further pretrain your base models
          DeploymentModel                                                         # Wrapper to optimize your serialized models for a production environment
          )

For example usage of each of these target types, see the finetune/datasets directory. For purposes of simplicity and runtime these examples use smaller versions of the published datasets.

If you have large amounts of unlabeled training data and only a small amount of labeled training data, you can finetune in two steps for best performance.

model = Classifier()               # Load base model
model.fit(unlabeledX)              # Finetune base model on unlabeled training data
model.fit(trainX, trainY)          # Continue finetuning with a smaller amount of labeled data
predictions = model.predict(testX) # [{'class_1': 0.23, 'class_2': 0.54, ..}, ..]
model.save(path)                   # Serialize the model to disk

Installation

Finetune can be installed directly from PyPI by using pip

pip3 install finetune

or installed directly from source:

git clone -b master https://github.com/IndicoDataSolutions/finetune && cd finetune
python3 setup.py develop              # symlinks the git directory to your python path
pip3 install tensorflow-gpu --upgrade # or tensorflow-cpu
python3 -m spacy download en          # download spacy tokenizer

In order to run finetune on your host, you'll need a working copy of tensorflow-gpu >= 1.14.0 and up to date nvidia-driver versions.

You can optionally run the provided test suite to ensure installation completed successfully.

pip3 install pytest
pytest

Docker

If you'd prefer you can also run finetune in a docker container. The bash scripts provided assume you have a functional install of docker and nvidia-docker.

git clone https://github.com/IndicoDataSolutions/finetune && cd finetune

# For usage with NVIDIA GPUs
./docker/build_gpu_docker.sh      # builds a docker image
./docker/start_gpu_docker.sh      # starts a docker container in the background, forwards $PWD to /finetune

docker exec -it finetune bash # starts a bash session in the docker container

For CPU-only usage:

./docker/build_cpu_docker.sh
./docker/start_cpu_docker.sh

Documentation

Full documentation and an API Reference for finetune is available at finetune.indico.io.