EuroPy is a declarative python testing framework designed for the machine learning (ML) pipeline. Inspired by other testing frameworks (such as JUnit and NUnit) EuroPy uses test decorators to register and execute tests.
EuroPy is designed to work within an Iron Python interactive environment (i.e. Jupyter Notebook) as well as within the typical python kernel.
As the use of machine learning becomes increasingly common, there is a greater need to address issues of reproducibility and the prevalence of ML systems' unforeseen adverse effects. Doing so requires a more structured approach to evaluating ML systems. EuroPy is a ML testing library that incorporates recommendations from existing literature to aid in the organization and reporting of tests in the ML pipeline. EuroPy is a lightweight testing framework designed to help data scientists and machine learning engineers write and generate results for custom tests to address factors such as Accuracy, Bias, Fairness, Data Bias, Minimum Functionality, etc. that are oftentimes not considered in the development of ML models and pipelines. By encouraging developers to consider potential sources of bias through ML testing, the EuroPy testing framework aims to minimize negative societal impact, address ethical debt, and foster responsible ML prior to deployment.
pip install europypython setup.py installCreating a test function is accomplished by simply decorating any function with decorators from europy.decorators. This will register your test function with the EuroPy lifecycle. In order to execute all registered tests you simply need to execute the europy.lifecycle.reporting.execute_tests function. Any parameters passed to the execute_tests function will also be passed to all registered functions.
from europy.decorator import bias, fairness
from europy.lifecycle.reporting import execute_tests, generate_report, report_model_details, report_model_params
@bias("My example bias test", "Lorem ipsum dolor sit amet, consectetur adipiscing elit.")
def foo(input):
assert input % 2 == 0
return input
@fairness("My example fairness test that fails")
@bias()
def foo_failure(input):
assert input % 2 == 1
return input
@bias("example_figure")
def save_image_example(plots: dict):
import matplotlib.pyplot as plt
import numpy as np
plt.style.use('fivethirtyeight')
x = np.linspace(0, 10)
# Fixing random state for reproducibility
np.random.seed(19680801)
fig, ax = plt.subplots()
ax.plot(x, np.sin(x) + x + np.random.randn(50))
ax.plot(x, np.sin(x) + 0.5 * x + np.random.randn(50))
ax.plot(x, np.sin(x) + 2 * x + np.random.randn(50))
ax.plot(x, np.sin(x) - 0.5 * x + np.random.randn(50))
ax.plot(x, np.sin(x) - 2 * x + np.random.randn(50))
ax.plot(x, np.sin(x) + np.random.randn(50))
ax.set_title("'fivethirtyeight' style sheet")
plots['fivethirtyeight_fig'] = plt
report_model_params('tests/param_example.yml')
report_model_details('tests/model_details_example.yml')
execute_tests(clear=True, input=1)
generate_report(export_type='markdown', clear_report=True)Pulled from this notebook
Interested in seeing a more concrete example? EuroPy has a sibling repository EuroPy Examples that contains a more detailed demonstration of how to use the EuroPy testing framework for a more concrete use case.
When the generate_report function is called, all recorded tests are flushed to a report file. The report can be exported in markdown or JSON format. An example of a testing report for the above example can be found HERE
EuroPy can also help with generating high level details of your model for use in the report. This is included with the @model_details() decorator which reads data from .yaml or .json file. An example can be found at tests/model_details_example.[yml|json].
from europy.decorator import model_details
from europy.lifecycle.model_details import ModelDetails
@model_details('model_details.yml') # will load details from file and pass into train
def train(details: ModelDetails=None):
# train
details.description = "something computed"
@model_details() # will load existing details from the current report
def test(detail: ModelDetails=None):
passTo include the model details in your test report simply add the model details decoratoer to a test inside of your test suite OR use the report_model_details function. Examples of both are below
from europy.decorator import model_details
from europy.lifecycle.model_details import ModelDetails
from europy.lifecycle.reporting import execute_tests, generate_report
@model_details('model-details.yml')
def loaded_model_details(details: ModelDetails = None):
# You can also manipulate your details object HERE if you wish
pass
execute_tests()
generate_report()from europy.lifecycle.reporting import execute_tests, generate_report, report_model_details
#... Write Tests
execute_tests()
report_model_details('model-details.yml')
generate_report()Keeping track of hyper-parameter tuning can be accomplished in configuration .yml or .json files and loaded into your code with the decorator using_params() or globally with load_global_params(). Doing so will automatically add the parameter details to your report. Parameters are loaded from
global:
num_epochs: 4
batch_size: 128
learning_rate: 0.0001
train:
test_split: 0.2
test:
title: "Testing Run 01"
batch_size: 256from europy.decorator import using_params
@using_params('params.yml')
def train(num_epochs: int=None, batch_size: int=None, learning_rate: float=None, test_split: float=None):
# will load global params first, then params matching func name
pass
@using_params('params.yml')
def test(title: str="", batch_size: int=None):
# will load batch_size from test, overriding the global definition
pass 
To include the model parameters in your test report simply add the using_params decorator to a test inside of your test suite OR use the report_model_params function. Examples of both are below
from europy.decorator import using_params
from europy.lifecycle.reporting import execute_tests, generate_report
@using_params('params.yml')
def test(title: str="", batch_size: int=None):
# will load batch_size from test, overriding the global definition
pass
execute_tests()
generate_report()from europy.lifecycle.reporting import execute_tests, generate_report, report_model_params
#... Write Tests to execute
execute_tests()
report_model_params('model-parameters.yml')
generate_report()EuroPy supports a number of test classes out of the box. These test classes have been created by surveying industry practitioners and creating a list of well known testing labels. When defining a testing function you may pass two optional parameters to any decorator: name and description. These parameters are used as metadata and will be included in the resulting test report. The supported labels are listed below.
@bias()@data_bias()@fairness()@transparency()@accountability()@accuracy()@unit()@integration()@minimum_functionality()
While we believe these decorators should be comprehensive we have provided two ways in which you can add your own custom label to the test reports. For the examples below we will be adding a custom "end-to-end" testing label to our testing function.
- Using the generic
@test()decorator you can provide a custom string label
from europy.decorator import test
from europy.lifecycle.reporting import execute_tests
EXAMPLE_LABEL_NAME = "end-to-end"
@test(EXAMPLE_LABEL_NAME, "My custom label test")
def foo(input):
assert input % 2 == 0
return input
execute_tests(100)- Defining your own test decorator using the decorator_factory
from europy.decorator import decorator_factory
from europy.lifecycle.reporting import execute_tests
def end_to_end(name: str = ""):
# here you can add multiple labels if you wish
labels = ["end-to-end"]
return decorator_factory(labels, name)
@end_to_end("Test with custom decorator")
def foo(input):
assert input % 2 == 0
return input
execute_tests(100)