Pattern extraction from text

Convolutional network (CNN) trying to mimic a regular expression. This is a proof of principle that a CNN is able to build text patterns by itself to extract information from a text to minimize a loss function.

Prerequisites

An installation of the conda package manager is required. A sufficient (according to needed packages in the LATEXWriter) installation of latex (lualatex) is needed. The fastest way to obtain conda is to install Miniconda.

Setup

A script for easy installation of the conda enviroments is included.

./setup.sh

tensorboard

Beside LaTeX, we use tensorboard for visualization during training.
You can activate it and make it visualize all stored log-files via, e.g., (adjust paths and envs):

bash -c "source /home/chambroc/miniconda3/bin/activate bob && tensorboard --logdir=./output/tensorboard"

References

• see slides in doc folder