Move to pytest about mlos HOT 6 CLOSED

we need a pytest.ini to store the pattern that matches our test, and then we can replace the command in CI / the test scripts, both of which should be quite easy, but we need to make sure we actually run all the tests.

Then we can start using some of the fancier features gradually if we like to, but really there's nothing else we need to do. Happy to make a PR in a bit.

from mlos.

For the record, one of the goals here is to be able to get some better stats on how long tests take.
See Also #67

from mlos.

parametrizable tests, better error messages, plugins for coverage and some other plugins.

pytest is a bit magic, it parses the AST to figure out a way to give you a good error, so if you write

b = 5
assert b == 6

It tells you "assert failed 5 != 6", while unittest or just plain python would say "AssertionError(False)".

I don't recall seeing a package not using pytest recently. There used to be nosetests, but that is no longer maintained and most packages moved from nosetests to pytest.
Examples of packages that use pytest are numpy, pandas, sklearn, requests, https://github.com/microsoft/hummingbird, https://github.com/microsoft/DeepSpeed, https://github.com/microsoft/coax, ...

from mlos.

The other reason we were discussing this was in the course of troubleshooting #67 and trying to figure out which tests were taking a long time.

I had written this nasty oneliner to try and identify the seconds elapsed that each test took:

last_ts=''; ./scripts/run-python-tests.sh | logger -s -t python-test 2>&1 | grep 'python-test: test_' | while read line; do echo "$line"; ts=$(echo "$line" | egrep -o '^<13>Sep 10 [0-9:]+ ' | sed 's/^<13>//' | xargs -I{} date -d"{}" +%s); if [ -n "$last_ts" ]; then let time_took=ts-last_ts; echo $time_took; fi; last_ts=$ts; done | egrep -B1 '^[1-9][0-9]+'
<13>Sep 10 21:01:39 python-test: test_random_search_optimizer (mlos.Optimizers.ExperimentDesigner.NumericOptimizers.unit_tests.TestRandomSearchOptimizer.TestRandomSearchOptimizer) ... ok
85
--
<13>Sep 10 21:01:50 python-test: test_random_function_configs (mlos.Optimizers.ExperimentDesigner.UtilityFunctions.unit_tests.TestConfidenceBoundUtilityFunction.TestConfidenceBoundUtilityFunction) ... ok
11
--
<13>Sep 10 21:02:04 python-test: test_default_homogeneous_random_forest_model (mlos.Optimizers.RegressionModels.unit_tests.TestHomogeneousRandomForestRegressionModel.TestHomogeneousRandomForestRegressionModel) ...
14
--
<13>Sep 10 21:02:52 python-test: test_hierarchical_quadratic_cold_start (mlos.Optimizers.unit_tests.TestBayesianOptimizer.TestBayesianOptimizer) ...  9881.271554996536
12
<13>Sep 10 21:04:00 python-test: test_hierarchical_quadratic_cold_start_random_configs (mlos.Optimizers.unit_tests.TestBayesianOptimizer.TestBayesianOptimizer) ...
68
<13>Sep 10 21:04:15 python-test: test_translating_dataframe_from_categorical_hierarchical_to_discrete_flat_hypergrid (mlos.Spaces.HypergridAdapters.unit_tests.TestCategoricalToDiscreteHypergridAdapter.TestCategoricalToDiscreteHypergridAdapter) ... ok
15

However, @amueller pointed out that pytest just gives us that with a CLI option.

from mlos.

I'm still educating myself about the relative merits of these two frameworks. What features of pytest particularly take your fancy? I know you mentioned parameterizable tests in the past...

from mlos.

That sounds really neat. What does it take to migrate?

from mlos.