Description

This package provides an implementation of conditional random field (CRF) in PyTorch. This implementation borrows mostly from AllenNLP CRF module with some modifications.

NOTE: This software is still in alpha version; every minor version change introduces backward incompatibility.

Requirements

• Python 3.6

• PyTorch 0.3.0

Installation

You can install with pip

pip install pytorch-crf

Or, you can install from Github directly

pip install git+https://github.com/kmkurn/pytorch-crf#egg=pytorch_crf

Examples

In the examples below, we will assume that these lines have been executed

>>> import torch
>>> from torchcrf import CRF
>>> seq_length, batch_size, num_tags = 3, 2, 5
>>> emissions = torch.autograd.Variable(torch.randn(seq_length, batch_size, num_tags), requires_grad=True)
>>> tags = torch.autograd.Variable(torch.LongTensor([[0, 1], [2, 4], [3, 1]]))  # (seq_length, batch_size)
>>> model = CRF(num_tags)
>>> # Initialize model parameters
... for p in model.parameters():
...    _ = torch.nn.init.uniform(p, -1, 1)
...
>>>

>>> model(emissions, tags)
Variable containing:
-10.0635
[torch.FloatTensor of size 1]

>>> mask = torch.autograd.Variable(torch.ByteTensor([[1, 1], [1, 1], [1, 0]]))  # (seq_length, batch_size)
>>> model(emissions, tags, mask=mask)
Variable containing:
-8.4981
[torch.FloatTensor of size 1]

>>> model.decode(emissions)
[[3, 1, 3], [0, 1, 0]]

>>> model.decode(emissions, mask=mask)
[[3, 1, 3], [0, 1, 0]]

License

MIT. See LICENSE.txt for details.

Contributing

Contributions are welcome! Please follow these instructions to setup dependencies and running the tests and linter. Make a pull request once your contribution is ready.

pip install -r requirements.txt
pip install -e .