Skip to content

Marian machine translation training pipeline for thousands of models

Notifications You must be signed in to change notification settings

hplt-project/OpusPocus

Repository files navigation

OpusPocus

Modular NLP pipeline manager.

OpusPocus is aimed at simplifying the description and execution of popular and custom NLP pipelines, including dataset preprocessing, model training, fine-tuning and evaluation. The pipeline manager supports execution using simple CLI (Bash) or common HPC schedulers (Slurm).

It uses OpusCleaner for data preparation and OpusTrainer for training scheduling (development in progress).

Structure

  • go.py - pipeline manager entry script
  • opuspocus/ - OpusPocus modules
  • opuspocus_cli/ - OpusPocus CLI subcommands
  • config/ - default configuration files (pipeline config, marian training config, ...)
  • examples/ - pipeline manager usage examples
  • scripts/ - helper scripts, at this moment not directly implemented in OpusPocus
  • tests/ - unit tests

Installation

  1. Install MarianNMT
$ ./scripts/install_marian_gpu.sh PATH_TO_CUDA CUDNN_VERSION [NUM_THREADS]

Alternatively, you can usel scripts/install_marian_cpu.sh for CPU version. Note that the scripts may require modification based on your system.

  1. (Optional) Setup the Python virtual environment (using virtualenv):
$ /usr/bin/virtualenv -p /usr/bin/python3.10 python-venv
  1. Install the Python dependencies.
(source python-venv/bin/activate  # if using virtual environment)
$ pip install --upgrade pip setuptools
$ pip install -r requirements.txt
  1. Setup the Python virtual environment for Opuscleaner. (OpusCleaner is currently not supported by Python>=3.10.)
$ /usr/bin/virtualenv -p /usr/bin/python3.9 opuscleaner-venv
  1. Activate the OpusCleaner virtualenv and install OpusCleaner's dependencies
$ source opuscleaner-venv/bin/activate
$ pip install --upgrade pip setuptools
$ pip install -r requirements-opuscleaner.txt

Usage (Simple Pipeline)

Either run the main script go.py or the subcommand scripts from opuspocus_cli/ directory. Run the scripts directly from the root directory for this repository. (You may need to add the path to the local OpusPocus repository directory to your PYTHONPATH.)

You can execute ./go.py --help for general description or ./go.py <subcommand> --help to list the available subcommand options.

Pipeline execution

Run ./go.py run (or opuspocus_cli/run) while providing a pipeline configuration file to execute a new pipeline:

$ ./go.py --pipeline-dir <pipeline_destination> --pipeline-donfig <config_file> --runner <runner>

Alternatively, run ./go.py run while providing an existing pipeline directory to rerun a failed pipeline execution:

$ ./go.py run --pipeline-dir <pipeline_dir> --runner <runner>

You can use --reinit to reinitialize the exitisting pipeline before running. You can use --resubmit-done to also execute pipeline steps in the DONE state.

Lastly, you can also stop and resubmit a running pipeline using --stop-previous-run

$ ./go.py run --pipeline-dir <pipeline_dir> --stop-previous-run

This is simialr to:

$ ./go.py stop --pipeline-dir <pipeline_dir>
$ ./go.py run --pipeline-dir <pipeline_dir>

Other subcommands

  • stop - stops the execution of a running pipeline
  • status- prints the status of a pipeline its steps
  • traceback - prints the dependency structure of a pipeline

Examples

I. Data preprocessing example

  1. Download the data and setup the dataset directory structure.
$ scripts/prepare_data.en-eu.sh
  1. Initialize and execute the (data preprocessing) pipeline.
$ mkdir -p experiments/en-eu/preprocess.simple
$ ./go.py run \
    --pipeline-config config/pipeline.preprocess.yml \
    --pipeline-dir experiments/en-eu/preprocess.simple \
	--runner bash
  • --pipeline-config (required) provides the details about the pipeline steps and their dependencies
  • --pipeline-dir (optional) overrides the pipeline.pipeline_dir value from the pipeline-config
  • --runner (required) runner to be used for pipeline execution. Use --runner slurm for more effective HPC execution (if Slurm is available)
  1. Check the pipeline status.
$ ./go.py traceback --pipeline-dir experiments/en-eu/preprocess.simple

OR

$ ./go.py status --pipeline-dir experiments/en-eu/preprocess.simple

II. Model training example (preprocessing follow-up)

  1. Check the preprocessing pipeline status (The data preprocessing pipeline must be finished, i.e. all steps must be in the DONE step)
$ ./go.py status --pipeline-dir experiments/en-eu/preprocess.simple
  1. Initialize and execute the training pipeline.
$ mkdir -p experiments/en-eu/train.simple
$ ./go.py run \
    --pipeline-config config/pipeline.train.simple.yml \
    --pipeline-dir experiments/en-eu/train.simple \
	--runner bash

(Advanced) Config modification examples

TBD

About

Marian machine translation training pipeline for thousands of models

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published