SigNetworks: Sequential Path Signature Networks

SigNetworks (sig-networks) is a PyTorch package for training and evaluating neural networks for longitudinal NLP classification tasks. sig-networks is a library that applies models first developed in Sequential Path Signature Networks for Personalised Longitudinal Language Modeling by Tseriotou et al. (2023) which presented a novel extension of neural sequential models using the notion of path signatures from rough path theory.

For more details of this library, see our EACL demo paper Sig-Networks Toolkit: Signature Networks for Longitudinal Language Modelling.

For some quick examples without having to install anything on your personal machine, we have a few examples on Deepnote - you will need to duplicate the project to run it yourself though:

• SWNU Network example on Anno-MI dataset

Installation

SigNetworks is available on PyPI and can be installed with pip:

pip install sig_networks

Note that currently sig_networks only supports Python 3.8 since it relies on the signatory library. However, it is possible to install signatory with more recent Python and PyTorch versions if you install it from source. See the installation guide in the signatory documentation for more details.

Signatory/Torch

SigNetworks depends on the patrick-kidger/signatory library for differentiable computation of path signatures/log-signatures in PyTorch. Please see the signatory documentation for installation instructions of the signatory library.

A common signatory installation issue is that the installation requires that you already have PyTorch installed. In this case, you can try the following:

# install PyTorch
pip install torch==1.9.0
# install signatory
pip install signatory==1.2.6.1.9.0
# install sig_networks
pip install sig_networks

If you encounter any issues with the installation of signatory, please see the FAQs in the signatory documentation.

Repo structure

The key parts of the libary are found in src/:

• src/sig_networks/ contains the source code for the models and includes PyTorch modules for the various components of the models (see below for more usage details)
• src/scripts/ contains some helper scripts for training and evaluating the models

Usage

The key components in the signature-window models presented in (see Sequential Path Signature Networks for Personalised Longitudinal Language Modeling for full details) are written as PyTorch modules which can be used in a modular fashion. The key components are:

• The Signature Window Network Units (SWNUs): sig_networks.SWNU
  • There also exists the SWLSTM module which does not include the 1D-convolution layer at the start of the SWNU: sig_networks.SWLSTM
• The Signature Window (Multihead-)Attention Units (SWMHAUs): sig_networks.SWMHAU
  • As with the SWNU, there also exists the SWMHA module which does not include the 1D-convolution layer at the start of the SWMHAU: sig_networks.SWMHA
• The SWNU-Network model: sig_networks.SWNUNetwork
• The SWMHAU-Network model: sig_networks.SWMHAUNetwork
• The SeqSigNet model: sig_networks.SeqSigNet
• The SeqSigNet-Attention-Encoder model: sig_networks.SeqSigNetAttentionEncoder
• The SeqSigNet-Attention-BiLSTM model: sig_networks.SeqSigNetAttentionBiLSTM

Also see our demo paper Sig-Networks Toolkit: Signature Networks for Longitudinal Language Modelling for details of each of these model components.

nlpsig: Preparing the data

The functionality to prepare the data for the SW-models and constructing paths and inputs are found in the nlpsig library which can be easily installed using pip and comes as a dependency in sig-networks.

Paths can be constructed using the nlpsig.PrepareData class. Furthermore, there is functionality within the nlpsig.TextEncoder and nlpsig.SentenceEncoder classes to obtain embeddings using transformers to be used as the channels in the paths. Since we want to take path signatures within the SW-models, we need to ensure that the number of channels in the path are low enough that we can take the path signatures efficiently. To enable this, there are also a number of dimensionality reduction methods in the nlpsig library - see nlpsig.DimReduce.

For full details, see the nlpsig GitHub repo and there are examples of using the library in the examples/ directory.

Note that for obtaining inputs to the SWNU-/SWMHA-Networks and the SeqSigNet family models, there are helper functions in the scripts (see e.g. obtain_SWNUNetwork_input in src/scripts/swnu_network_functions.py and obtain_SeqSigNet_input in src/scripts/seqsignet_functions.py). There is also examples run-throughs in the examples/ directory:

• Training a SWNU-Network model for Anno-MI client-talk-type prediction
  • Note that this notebook is also on Deepnote: Anno-MI SWNU example
  • Note that you will need to duplicate the project so that you can run the notebook yourself
• Training a SeqSigNet model for Anno-MI client-talk-type prediction

Using the SWNU and SWMHAU modules

The Signature Window units (SWNU and SWMHAU) accept a batch of streams and returns a batch of feature representations.

For example:

from sig_networks.swnu import SWNU
import torch

# initialise a SWNU object
swnu = SWNU(
    input_channels=10,
    hidden_dim=5,
    log_signature=False,
    sig_depth=3,
    pooling="signature",
    BiLSTM=True,
)

# create a three-dimensional tensor of batched streams
# shape [batch, length, channels] where channels = 10
streams = torch.randn(2, 20, 10)

# pass the streams through the SWNU
features = swnu(streams)

# features is a two-dimensional tensor of shape [batch, signature_channels]
features.shape

The SWMHAU is similar to the SWNU, but rather than having an LSTM to process the signature streams, we have a multihead-attention layer. For example:

from sig_networks.swmhau import SWMHAU
import torch

# initialise a SWMHAU object
swmhau = SWMHAU(
    input_channels=10,
    output_channels=5,
    log_signature=False,
    sig_depth=3,
    num_heads=5,
    num_layers=1,
    dropout_rate=0.1,
    pooling="signature",
)

# create a three-dimensional tensor of batched streams
# shape [batch, length, channels] where channels = 10
streams = torch.randn(2, 20, 10)

# pass the streams through the SWMHAU
features = swmhau(streams)

# features is a two-dimensional tensor of shape [batch, signature_channels]
features.shape

Note in the above, we used the pooling="signature" option. This means that at the end of the SWNU/SWMHAU, we will take a final signature of the streams to get a fixed-length feature representation for each item in the batch. There are other options such as taking the final LSTM hidden state for SWNU (set pooling="lstm"), or using a CLS pooling for SWMHAU (set pooling="cls"). There is another option where pooling=None which means that the SWNU/SWMHAU where no pooling is applied at the end of the SWNU/SWMHAU and the output is a three-dimensional tensor of shape [batch, length, hidden_dim].

Using the network models

The library also has the SWNU-Network and SeqSigNet models as introduced in Sequential Path Signature Networks for Personalised Longitudinal Language Modeling.

Since then, there have been developments of other models which utilise the SWNUs and SWMHAUs discussed above. Each of these models are avaliable as PyTorch modules which can be initialised and trained in the usual way.

For SWNU-Network and SWMHAU-Network models, they expect two inputs:

1. path: a batch of streams of shape [batch, length, channels] - these get processed by the SWNU/SWMHAU
2. features: a batch of features of shape [batch, features] - these get concatenated with the output of the SWNU/SWMHAU to be fed into a FFN layer

For SeqSigNet models (e.g. SeqSigNet, SeqSigNet-Attention-Encoder, SeqSigNet-Attention-BiLSTM), they also expect two inputs but the path is slightly different:

1. path: a batch of streams of shape [batch, units, length, channels] - each of the units for each batch will get processed by the SWNU/SWMHAU. Afterwards, there is a global network to process the outputs of the SWNU/SWMHAU in order to pool the outputs into a single fixed-length feature represenation for the history. The global network can either be a BiLSTM (in the case of SeqSigNet and SeqSigNet-Attention-BiLSTM) or a Transformer Encoder (in the case of SeqSigNet-Attention-Encoder).
2. features: a batch of features of shape [batch, features] - these get concatenated with the output of the global network (either BiLSTM or a Transformer Encoder) that processes the outputs of SWNU and SWMHAU to be fed into a FFN layer

Example experiments

In the examples/ directory, there are some example experiments using the library to comapre the SW-models with other baseline models such as a simple FFN, a BiLSTM model on the sentence-transformer representations and using a pre-trained Transformer model for classification.

Pre-commit and linters

To take advantage of pre-commit, which will automatically format your code and run some basic checks before you commit:

pip install pre-commit  # or brew install pre-commit on macOS
pre-commit install  # will install a pre-commit hook into the git repo

After doing this, each time you commit, some linters will be applied to format the codebase. You can also/alternatively run pre-commit run --all-files to run the checks.