pytorch-mcrf 0.0.3

Multiple CRF implementation for PyTorch

Masked CRF

NOT Official Pytorch implemented Masked CRF.

Installation

Dependencies

• Python >= 3.6
  • torch == 1.5.1 (better > 1.0)
  • tqdm == 4.53.0
  • pyyaml == 5.3.1

Download and Install

1. Editable installation:

pip install -e .

2. Install from PyPI:

pip install pytorch-mcrf -i https://pypi.org/simple

Quick Start

Settings

All the settings are in config.yaml, you can change model settings from this file.

Run

python run.py

Results

Micro-F1 results

WeiboNER

Batch Size	Optimizer	Learning Rate	Max Seqence Length
8	SGD	0.015	128

Method	Dev	Test	#Illegal Tags
PlainCRF	57.564	51.733	4
MaskedCRF ( decoding only )	55.662	51.351	0
MaskedCRF	55.380	50.287	0

MSRA

Batch Size	Optimizer	Learning Rate	Max Seqence Length
8	SGD	0.015	250

Model is selected directly from test set since there is no official dev set.

Method	Test	#Illegal Tags
PlainCRF	86.274	1
MaskedCRF ( decoding only )	86.285	0
MaskedCRF	84.714	0

Acknowledgements

• Official: MaskedCRF
• pytorch-crf
• allennlp
• ChineseWeiboNER

UPDATES

• v0.0.2: fix setuptools packages finding issue