Skip to content

Latest commit

 

History

History
107 lines (74 loc) · 3.75 KB

README.md

File metadata and controls

107 lines (74 loc) · 3.75 KB
Raw

ai-label-trainer-service

A docker container for training an AI model using a neural network for labeling text.

Executes a job, defined by the ai-job-service

Getting started

services:
  label-trainer:
    image: redpencil/ai-label-trainer
    environment:
      LOG_LEVEL: "debug"
      TASK: "predicates"
    links:
      - db:database
    volumes:
      - ./config/label-trainer/constants.py:/app/constants.py
      - ./share:/share
  #    deploy:
  #      resources:
  #        limits:
  #          cpus: 4

The TASK variable is the type of task that starts this training loop using the job-run-loop. This task should be used when creating a job with the ai-job-service.

You probably need to define the amount of virtual cpu cores that this service allowed to use to limit usage. Note that performance will also decrease with less resources.

If an Nvidia GPU is available, it is advised to change the Dockerfile so it allows the use of a GPU.

After training, the job is updated with the id of the file containing the trained model, stored in the triplestore.

Configuration

In config/label-trainer/constants.py:

import torch

SEP_TOKEN = '[SEP]'
CLS_TOKEN = '[CLS]'
MODEL_FILE_PATH = '/share/model/predicate-model.pth'
BATCH_SIZE = 4
NUM_EPOCHS = 3
GRADIENT_ACCUMULATION_STEPS = 8
MAX_CLASS_SIZE = 500
DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

The MAX_CLASS_SIZE defines the maximum amount of items per class that sould be used for training. Keep in mind that biasness due to class imbalance can occur if the number of items per class are very different. A larger possibly increases performance, but also learning time.

Make sure to mount this python file in the correct location of the container as shown in the docker-compose entry above!

Tensorboard

Tensorboard was used to log training loss and accuracy in a dashboard. All data will be written to share/tb/. In order to start up the dashboard, make sure tensorboard is installed locally with:

pip install tensorboard

starting tensorboard can now be done by running:

tensorboard --logdir=share/tb

Tensorboard is now available at http://localhost:6006 if that port is available, otherwise see terminal for correct port.

Reference

Environment variables

  • LOG_LEVEL takes the same options as defined in the Python logging module.

  • MU_SPARQL_ENDPOINT is used to configure the SPARQL query endpoint.

    • By default this is set to http://database:8890/sparql. In that case the triple store used in the backend should be linked to the microservice container as database.

  • MU_SPARQL_UPDATEPOINT is used to configure the SPARQL update endpoint.

    • By default this is set to http://database:8890/sparql. In that case the triple store used in the backend should be linked to the microservice container as database.

  • MU_APPLICATION_GRAPH specifies the graph in the triple store the microservice will work in.

    • By default this is set to http://mu.semte.ch/application. The graph name can be used in the service via settings.graph.

  • MU_SPARQL_TIMEOUT is used to configure the timeout (in seconds) for SPARQL queries.

Improvements

Every time the model is trained, it is stored on disk under the file name, defined in the config.py file. This means that if training is done two times, the original model can be overwritten. The link to this file in the triple store will not be changed. So, the older model will no longer be available while the id for this one will give the new model.