tira 0.0.67

Simple access to the TIRA API.

The TIRA Client

This is a python client for TIRA.io.

Setup REST Client to Access Non-Public Endpoints

To access non-public endpoints, you will need an authentication via an API key to ensure that you have the correct access credentials. Please generate your API key online at tira.io/admin/api/keys and create a credentials file at ~/.tira/.tira-settings.json with the following content:

{"api_key": "<YOUR-API-KEY>"}

Download The results of some Submission

You can download runs of published and unblinded submissions via:

from tira.rest_api_client import Client

tira = Client()
output = tira.get_run_output('<task>/<team>/<approach>', '<dataset>')

As an example, you can download all baseline BM25 runs submitted to TIREx via:

from tira.rest_api_client import Client

tira = Client()
datasets = ['antique-test-20230107-training', 'argsme-touche-2020-task-1-20230209-training', 'argsme-touche-2021-task-1-20230209-training', 'clueweb09-en-trec-web-2009-20230107-training', 'clueweb09-en-trec-web-2010-20230107-training', 'clueweb09-en-trec-web-2011-20230107-training', 'clueweb09-en-trec-web-2012-20230107-training', 'clueweb12-touche-2020-task-2-20230209-training', 'clueweb12-touche-2021-task-2-20230209-training', 'clueweb12-trec-web-2013-20230107-training', 'clueweb12-trec-web-2014-20230107-training', 'cord19-fulltext-trec-covid-20230107-training', 'cranfield-20230107-training', 'disks45-nocr-trec-robust-2004-20230209-training', 'disks45-nocr-trec7-20230209-training', 'disks45-nocr-trec8-20230209-training', 'gov-trec-web-2002-20230209-training', 'gov-trec-web-2003-20230209-training', 'gov-trec-web-2004-20230209-training', 'gov2-trec-tb-2004-20230209-training', 'gov2-trec-tb-2005-20230209-training', 'gov2-trec-tb-2006-20230209-training', 'medline-2004-trec-genomics-2004-20230107-training', 'medline-2004-trec-genomics-2005-20230107-training', 'medline-2017-trec-pm-2017-20230211-training', 'medline-2017-trec-pm-2018-20230211-training', 'msmarco-passage-trec-dl-2019-judged-20230107-training', 'msmarco-passage-trec-dl-2020-judged-20230107-training', 'nfcorpus-test-20230107-training', 'vaswani-20230107-training', 'wapo-v2-trec-core-2018-20230107-training']

for dataset in datasets:
    output = tira.get_run_output('ir-benchmarks/tira-ir-starter/BM25 Re-Rank (tira-ir-starter-pyterrier)', dataset)

Export datasets

You can export datasets if you are the owner or if the dataset is public. Export a dataset via the cli:

tira-run --export-dataset '<task>/<tira-dataset>' --output-directory tira-dataset

Export a dataset via the python API:

from tira.rest_api_client import Client

tira = Client()
tira.download_dataset('<task>', '<tira-dataset>')

Running Jupyter Notebooks with TIRA

docker build -t tira/submission-base-image:1.0.0 -f Dockerfile .

Testing the model locally can be done using the following command:

tira-run \
  --input-directory ${PWD}/input \
  --output-directory ${PWD}/output \
  --image tira/submission-base-image:1.0.0 \
  --command 'tira-run-notebook --input $inputDataset --output $outputDir /workspace/template-notebook.ipynb'

---

Afterwards you can push the image to TIRA

docker push tira/submission-base-image:1.0.0

and set the command:

tira-run-notebook --input $inputDataset --output $outputDir /workspace/template-notebook.ipynb

---

Finally, if the actual processing in notebook is toggled via is_running_as_inference_server() (as seen in the template notebook) and your notebook defines a function named predict in the format

def predict(input_list: List) -> List:

you can start an inference server for your model with:

PORT=8001

docker run --rm -it --init \
  -v "$PWD/logs:/workspace/logs" \
  -p $PORT:$PORT \
  tira/submission-base-image:1.0.0 \
  tira-run-inference-server --notebook /workspace/template-notebook.ipynb --port $PORT

Exemplary request for a server running on localhost:8001 are

# POST (JSON list as payload)
curl -X POST -H "application/json" \
  -d "[\"element 1\", \"element 2\", \"element 3\"]" \
  localhost:8001

and

# GET (JSON object string(s) passed to the 'payload' parameter)
curl "localhost:8001?payload=\"element+1\"&payload=\"element+2\"&payload=\"element+3\""