CONTEXT v4: Neural network code for text categorization in C++ on GPU

• Shallow CNN (convolutional neural networks) [JZ15a]
• Shallow CNN enhanced with unsupervised embeddings (embeddings trained in an unsupervised manner) [JZ15b]
• Deep pyramid CNN (DPCNN) enhanced with unsupervised embeddings [JZ17]
• LSTM enhanced with unsupervised embeddings [JZ16a]

Looking for a tool?

• Try DPCNN [JZ17] if your labeled training data is large (e.g., 1M documents).
• Try shallow CNN [JZ15a] if your labeled training data is small (e.g., 50K documents).
• Also try enhancement of your networks with unsupervised embeddings [JZ15b, JZ16b, JZ17]
  • if you have good unlabeled data (large and in-domain) or
  • if your training data is large enough to be used as unlabeled data

System requirements    This code runs only on a CUDA-capable GPU such as Tesla K20. That is, your system must have a GPU and an appropriate version of CUDA installed. The provided makefile and example shell scripts are for Unix-like systems. Testing was done on Linux. In principle, the C++ code should compile and run also in other systems (e.g., Windows), but no guarantee. See README for more details.

Download & Documentation

• conText-v4.00a.tar.gz  (CONTEXT v4.00 code)
• README
• User guide (pdf)
• Optional resources
  • elec2.tar.gz (required for training with Elec training data of various sizes as in Fig.6 of [JZ15a])
  • Info on RCV1 used in [JZ15a,15b,16a]
  • Unlabeled data, etc.

Getting started

1. Download the code and extract the files, and read README.
2. Go to the top directory and build executables by make, after customizing makefile as needed.
3. To confirm installation, go to examples/ and enter ./sample.sh to train and test a network on small data.
   (See README for installation trouble shooting.)
4. Read Section 1 (Overview) of User Guide to get an idea.
5. Try some shell scripts at examples/. There is a table of the scripts in Section 1.6 of User Guide.

Data Source

License

References