BERT for Named Entity Disambiguation — bert_ned

TL;DR

To build and run the container:

make wharfer-build && make wharfer-run

To run the full pipeline (data generation, training, evaluation) inside the container:

make full

Makefile for Docker/Wharfer

There are shorthands for the preceding Docker and Wharfer commands in a Makefile. Navigate to the directory with this repository and type make help for a list of instructions.

Docker

You can build and run containers with specific volumes or names if you disregard the 'make' instructions.

Run the image on a machine with a Cuda-enabled GPU to train the model or use the model for inference.

Build the Docker image with GPU (7.27GB):

docker build -t bert_ned .

For completeness, there is a CPU version of the Dockerfile (Dockerfile.CPU). The CPU container works for the data preparation scripts, but the model runs poorly on a CPU.

Build without GPU (3.19GB):

docker build -f Dockerfile.CPU -t bert_ned_cpu .

Docker run

To run the Docker image on an AD machine using files from /nfs/:

docker run -v /nfs/students/amund-faller-raheim/master_project_bert_ned/data:/bert_ned/data \
           -v /nfs/students/amund-faller-raheim/master_project_bert_ned/ex_data:/bert_ned/ex_data \
           -v /nfs/students/amund-faller-raheim/master_project_bert_ned/models:/bert_ned/models \
           -it --name bert_ned bert_ned

Please note: accessing files over NFS makes some of the operations quite slow. You can also copy the directories in /nfs/students/amund-faller-raheim/master_project_bert_ned to a local directory and mount those to the docker container.

E.g. cp /nfs/students/amund-faller-raheim/master_project_bert_ned /local/data/$(whoami)/ and

docker run -v /local/data/$(whoami)/data:/bert_ned/data \
           -v /local/data/$(whoami)/ex_data:/bert_ned/ex_data \
           -v /local/data/$(whoami)/models:/bert_ned/models \
           -it --name bert_ned bert_ned

Run scripts in container

Instructions should appear once the container is running. Type 'make help' to see a list of actions. (This is from a second makefile Makefile_scripts.)

You can use 'make' to run scripts in the container. For example:

make full

Which runs the script bert_ned_full_pipeline.py. This script can do the full process, from data generation to training and evaluation.

When running the scripts, make sure that the settings in config.ini are correct for your environment (or the container, by default). Leave the settings as provided to reproduce results.