Introduction

UNF(Universal NLP Framework) is built on pytorch and torchtext. Its design philosophy is：

• modularity: specifically, on the one hand, it is convenient to quickly run some nlp-related tasks; on the other hand, it is convenient for secondary development and research to implement some new models or technologies.
• efficiency: supports distributed training and half-precision training, which is convenient for quickly training the model, although the current support is relatively crude
• comprehensive: support pytorch trace into static graph, support c ++ server, provide web-server for debugging tools

Support Tasks

Now, support text classification and sequence labeling related tasks.

Related Papers

• Convolutional Neural Networks for Sentence Classification 2014
• Bag of Tricks for Efficient Text Classification 2016
• Deep Pyramid Convolutional Neural Networks for Text Categorization 2017, ACL
• Hierarchical Attention Networks for Document Classification 2017, ACL
• A STRUCTURED SELF-ATTENTIVE SENTENCE EMBEDDING 2017,ICLR
• Joint Embedding of Words and Labels for Text Classification2018,ACL
• Neural Architectures for Named Entity 2016,ACL
• Semi-supervised Multitask Learning for Sequence Labeling 2017, ACL

Framwork

Module relation

Requirement

python3

pip3 install -r requirement.txt

Training

#quick start
python3 train_flow.py

Only 5 line code need

#data loader
data_loader = DataLoader(data_loader_conf)
train_iter, dev_iter, test_iter = data_loader.generate_dataset()

#model loader
model, model_conf = ModelLoader.from_params(model_conf, data_loader.fields)

#learner loader
learner = LearnerLoader.from_params(model, train_iter, dev_iter, learner_conf, test_iter=test_iter, fields=data_loader.fields, model_conf=model_conf)

#learning
learner.learn()

train_flow.py is the demo code for training，Run directly！！

conf for multu-gpu and mixed precision

"use_fp16": False,
"multi_gpu": False

tensorboard demo

Python inference

#quick start
python3 score_flow.py

#core code
from models.predictor import Predictor

predictor = Predictor(model_path, device, model_type)
logits = predictor.predict(input)

(0.18, -0.67)

C++ inference

step1: Trace dynamic graph to static graph

#quick start
python3 trace.py

#core code
net = globals()[model_cls](**config.__dict__)
net.load_state_dict_trace(torch.load("%s/best.th" % model_path))
net.eval()

mock_input = net.mock_input_data()
tr = torch.jit.trace(net, mock_input)
tr.save("trace/%s" % save_path)

step2: c++ serving

• install cmake
• download libtorch and unzip to trace folder

cd trace
cmake -DCMAKE_PREFIX_PATH=libtorch .

make

./predict trace.pt predict_vocab.txt
output: 2.2128 -2.3287

RESTFUL-API web demo

cd web_server
python run.py