This starter include a Pipfile to install tensorflow + bert-serving-server.

• tensorflow 1.12.0 (using gpu support version as default), read: https://www.tensorflow.org/install
• bert-serving-server http, from https://github.com/hanxiao/bert-as-service
• bert model BERT-Base, Multilingual Cased, from: https://github.com/google-research/bert

• Git
• Python3
• pipenv

• Nvidia GPU
• CUDA 9.0
• CUDNN 7.4.2

Clone this starter:

$ git clone https://github.com/jk195417/bert-as-service-starter.git
$ cd bert-as-service-starter

Install dependencies:

$ pipenv update

Download and unzip BERT-Base, Multilingual Cased into /models dir

If you don't want GPU support:

$ pipenv uninstall tensorflow-gpu
$ pipenv install tensorflow

1 worker need 1G RAM, example: GTX 1060 6G can open 4 workers max.

# Start server at localhost:8125
$ pipenv run bert-serving-start -model_dir=./models/multi_cased_L-12_H-768_A-12 -num_worker=4 -http_port=8125 -http_max_connect=20

Press ctrl+c twice to shutdown server, then you can clean up those tmp dirs.