pytorch-rex 0.0.9

A toolkit for Relation Extraction and more...

A toolkit for Relation & Event eXtraction (REx) and more...

This project has not been finished yet, so be careful when using it, or wait until the first release comes out.

This project is suffering from the second-system effect. I would like to cut some features to make this going smoothly.

Accelerate seems to be a very sweet wrapper for multi-GPU, TPU training, we highly recommend you to use such frameworks, instead of adding hard codes on your own.

⚙️Installation

Make sure you have installed all the dependencies below.

• Python>=3.6
  • torch>=1.2.0 : project is developed with torch==1.7.1, should be compatable with >=1.2.0 versions
  • numpy>=1.19.0
  • scikit-learn>=0.21.3
  • omegaconf>=2.0.6
  • loguru>=0.5.3
  • tqdm>=4.61.1
  • transformers>=4.8.2

$ git clone https://github.com/Spico197/REx.git
$ cd REx
$ pip install -e .

# or you can download and install from pypi, not recommend for now
$ pip install pytorch-rex -i https://pypi.org/simple

🚀QuickStart

Checkout the examples folder.

Name	Model	Dataset	Task
SentRE-MCML	PCNN	IPRE	Sentence-level Multi-class multi-label relation classification
BagRE	PCNN+ONE	NYT10	Bag-level relation classification (Multi-Instance Learning, MIL)
JointERE	CasRel	WebNLG	Jointly entity relation extraction

✈️Abilities

Dataset

• IPRE preprocess
• NYT10

Tasks

• Chinese sentence-level relation extraction
• English bag-level relation extraction

Modules & Models

• Piecewise CNN
• PCNN + ONE
• PCNN + ATT

✉️Update

• v0.0.9: use argparse instead of click, move loguru logger into rex.utils.logging, add hierarchical config setting
• v0.0.8: fix config loading, change default of makedirs and dump_configfile to be True
• v0.0.7: fix recursive import bug
• v0.0.6: integrate omega conf loading into the inner task, add load_*_data option to data managers
• v0.0.5: update ffn
• v0.0.4: return detailed classification information in mc_prf1, support nested dict tensor movement
• v0.0.3: fix packaging bug in setup.py
• v0.0.2: add black formatter and pytest testing
• v0.0.1: change LabelEncoder.to_binary_labels into convert_to_multi_hot or convert_to_one_hot

🔑LICENCE

MIT