Cross Language Information Retrieval pipeline
Patapsco - the SCALE 2021 Pipeline
Patapsco requires Python 3.6+ and Java 11+.
Installing Patapsco with Anaconda will add Java into the virtual environment. If not using Anaconda, you will need to check your Java version or enable the java module on the grid.
To check your Java version:
On the grid, enable Java with:
module add java
Create a Python virtual environment using venv or conda.
Installing with conda is recommended and will install the gpu-enabled version of pytorch. As of June 2021, CUDA 11.1.1 will be installed into the environment by default. You do not need to load any CUDA modules on the grid to use the GPUs.
Create and activate the conda environment:
conda env create --file environment.yml conda activate patapsco
pip install --editable .
With Python's venv module
Create and activate the virtual environment:
python3 -m venv venv source venv/bin/activate
You may need to upgrade your pip:
pip install -U pip pip install -U wheel
Install Patapsco and its dependencies:
pip install --editable .
Note: python virtual environments do not work properly on the HLTCOE grid!
If you do not have a C++ compiler or cannot install pytrec_eval, then comment out the lines in environment.yml and setup.py that specify pytrec_eval. Example:
- pip: - pyserini # - pytrec_eval
You will be able to run Patapsco, but not score your runs.
Patapsco was designed to create CLIR runs and not for training CLIR components (like reranking models). It is expected that the artifacts generated by Patapsco could be used for training, but that the training happens outside of Patapsco.
Patapsco consists of two pipelines:
- Stage 1: creates an index from the documents
- Stage 2: retrieves results for queries from the indexes and reranks the results
A pipeline consists of a sequence of tasks.
- Stage 1 tasks:
- text processing of documents (character normalization, tokenization, etc.)
- Stage 2 tasks:
- extract query from topic
- text processing of query (same as document processing)
- retrieval of results
- reranking of results
When a run is complete, its output is written to a run directory. Tasks also store artifacts in the run directory that can be used for other runs. For example, an index created in one run can be used in another.
Patapsco can run partial pipelines. For example, a user can run just stage 1 to generate an index. Or a user may run only stage 2 and have it start with processed queries and a prebuilt index.
Patapsco uses YAML or JSON files for configuration.
The stage 1 and stage 2 pipelines are built from the configuration.
The output including any artifacts (like processed queries or an index) are stored in a run directory.
For more information on configuration, see
After installing Patapsco, a sample run is started with:
By default, the output for the run is written to a
runs directory in the working directory.
If a run is complete, Patapsco will not overwrite it.
To turn on more detailed logging and full exception stack traces, use the debug flag:
patapsco --debug samples/configs/eng_basic.yml
Any variable in the configuration can be overriden on the command line:
patapsco --set run.name=my_test_run samples/configs/eng_basic.yml
A run's output file plus the configuration used to generate the run can be submitted at the website: https://scale21.org
Use issues on Gitlab to report bugs or request new features. For a bug report include
- a description of what was expected and what actually happened
- any stack trace or error message
- the configuration file if the bug only happens with that configuration
Developers should install Patapsco in editable mode along with development dependencies:
pip install -e .[dev]
To run the unit tests, run:
Some tests load models and are normally skipped. To run those:
The code should conform to the PEP8 style except for leniency on line length.
To update the code, you can use autopep8. To run it on a file:
autopep8 -i [path to file]
To test PEP8 compliance, run:
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.