Testing Machine Learning Systems: Code, Data and Models
๐ฌ Receive new lessons straight to your inbox (once a month) and join 30K+ developers in learning how to responsibly deliver value with ML.
Intuition
In this lesson, we'll learn how to test code, data and machine learning models to construct a machine learning system that we can reliably iterate on. Tests are a way for us to ensure that something works as intended. We're incentivized to implement tests and discover sources of error as early in the development cycle as possible so that we can decrease downstream costs and wasted time. Once we've designed our tests, we can automatically execute them every time we change or add to our codebase.
Tip
We highly recommend that you explore this lesson after completing the previous lessons since the topics (and code) are iteratively developed. We did, however, create the testing-ml repository for a quick overview with an interactive notebook.
Types of tests
There are four majors types of tests which are utilized at different points in the development cycle:
Unit tests
: tests on individual components that each have a single responsibility (ex. function that filters a list).Integration tests
: tests on the combined functionality of individual components (ex. data processing).System tests
: tests on the design of a system for expected outputs given inputs (ex. training, inference, etc.).Acceptance tests
: tests to verify that requirements have been met, usually referred to as User Acceptance Testing (UAT).Regression tests
: tests based on errors we've seen before to ensure new changes don't reintroduce them.
While ML systems are probabilistic in nature, they are composed of many deterministic components that can be tested in a similar manner as traditional software systems. The distinction between testing ML systems begins when we move from testing code to testing the data and models.

There are many other types of functional and non-functional tests as well, such as smoke tests (quick health checks), performance tests (load, stress), security tests, etc. but we can generalize all of these under the system tests above.
How should we test?
The framework to use when composing tests is the Arrange Act Assert methodology.
Arrange
: set up the different inputs to test on.Act
: apply the inputs on the component we want to test.Assert
: confirm that we received the expected output.
Cleaning
is an unofficial fourth step to this methodology because it's important to not leave remnants of a previous test which may affect subsequent tests. We can use packages such as pytest-randomly to test against state dependency by executing tests randomly.
In Python, there are many tools, such as unittest, pytest, etc. that allow us to easily implement our tests while adhering to the Arrange Act Assert framework. These tools come with powerful built-in functionality such as parametrization, filters, and more, to test many conditions at scale.
What should we test?
When arranging our inputs and asserting our expected outputs, what are some aspects of our inputs and outputs that we should be testing for?
- inputs: data types, format, length, edge cases (min/max, small/large, etc.)
- outputs: data types, formats, exceptions, intermediary and final outputs
๐ We'll cover specific details pertaining to what to test for regarding our data and models below.
Best practices
Regardless of the framework we use, it's important to strongly tie testing into the development process.
atomic
: when creating functions and classes, we need to ensure that they have a single responsibility so that we can easily test them. If not, we'll need to split them into more granular components.compose
: when we create new components, we want to compose tests to validate their functionality. It's a great way to ensure reliability and catch errors early on.reuse
: we should maintain central repositories where core functionality is tested at the source and reused across many projects. This significantly reduces testing efforts for each new project's code base.regression
: we want to account for new errors we come across with a regression test so we can ensure we don't reintroduce the same errors in the future.coverage
: we want to ensure 100% coverage for our codebase. This doesn't mean writing a test for every single line of code but rather accounting for every single line.automate
: in the event we forget to run our tests before committing to a repository, we want to auto run tests when we make changes to our codebase. We'll learn how to do this locally using pre-commit hooks and remotely via GitHub actions in subsequent lessons.
Test-driven development
Test-driven development (TDD) is the process of writing a test before writing the functionality to ensure that tests are always written. This is in contrast to writing functionality first and then composing tests afterwards. Here are our thoughts on this:
- good to write tests as we progress, but it does signify 100% correctness.
- initial time should be spent on design before ever getting into the code or tests.
Perfect coverage doesn't mean that our application is error free if those tests aren't meaningful and don't encompass the field of possible inputs, intermediates and outputs. Therefore, we should work towards better design and agility when facing errors, quickly resolving them and writing test cases around them to avoid next time.
Application
In our application, we'll be testing the code, data and models. We'll start by creating a separate tests
directory with code
subdirectory for testing our tagifai
scripts. We'll create subdirectories for testing data and models soon below.
mkdir tests
cd tests
mkdir app config model tagifai
touch <SCRIPTS>
cd ../
tests/
โโโ code/
โ โโโ test_data.py
โ โโโ test_evaluate.py
โ โโโ test_main.py
โ โโโ test_predict.py
โ โโโ test_utils.py
Feel free to write the tests and organize them in these scripts after learning about all the concepts in this lesson. We suggest using our tests
directory on GitHub as a reference.
Notice that our
tagifai/train.py
script does not have it's respectivetests/code/test_train.py
. Some scripts have large functions (ex.train.train()
,train.optimize()
,predict.predict()
, etc.) with dependencies (ex. artifacts) and it makes sense to test them viatests/code/test_main.py
.
๐ป Code
We'll start by testing our code and we'll use pytest as our testing framework for it's powerful builtin features such as parametrization, fixtures, markers and more.
pip install pytest==7.1.2
Since pytest is not integral to the core machine learning operations (ie. only a developer would need to run tests), let's create a separate list in our setup.py
and add it to our extras_require
:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 |
|
We created an explicit test
option because a user will want to only download the testing packages. We'll see this in action when we use CI/CD workflows to run tests via GitHub Actions.
Configuration
Pytest expects tests to be organized under a tests
directory by default. However, we can also add to our existing pyproject.toml
file to configure any other test directories as well. Once in the directory, pytest looks for python scripts starting with tests_*.py
but we can configure it to read any other file patterns as well.
1 2 3 4 |
|
Assertions
Let's see what a sample test and it's results look like. Assume we have a simple function that determines whether a fruit is crisp or not:
1 2 3 4 5 6 7 8 9 10 11 |
|
To test this function, we can use assert statements to map inputs with expected outputs. The statement following the word assert
must return True.
1 2 3 4 5 6 7 8 |
|
We can also have assertions about exceptions like we do in lines 6-8 where all the operations under the with statement are expected to raise the specified exception.
Example of using assert
in our project
1 2 3 4 5 6 7 8 9 10 11 12 |
|
Execution
We can execute our tests above using several different levels of granularity:
python3 -m pytest # all tests
python3 -m pytest tests/food # tests under a directory
python3 -m pytest tests/food/test_fruits.py # tests for a single file
python3 -m pytest tests/food/test_fruits.py::test_is_crisp # tests for a single function
Running our specific test above would produce the following output:
python3 -m pytest tests/food/test_fruits.py::test_is_crisp
tests/food/test_fruits.py::test_is_crisp . [100%]
Had any of our assertions in this test failed, we would see the failed assertions, along with the expected and actual output from our function.
tests/food/test_fruits.py F [100%] def test_is_crisp(): > assert is_crisp(fruit="orange") E AssertionError: assert False E + where False = is_crisp(fruit='orange')
Tip
It's important to test for the variety of inputs and expected outputs that we outlined above and to never assume that a test is trivial. In our example above, it's important that we test for both "apple" and "Apple" in the event that our function didn't account for casing!
Classes
We can also test classes and their respective functions by creating test classes. Within our test class, we can optionally define functions which will automatically be executed when we setup or teardown a class instance or use a class method.
setup_class
: set up the state for any class instance.teardown_class
: teardown the state created in setup_class.setup_method
: called before every method to setup any state.teardown_method
: called after every method to teardown any state.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 |
|
We can execute all the tests for our class by specifying the class name:
python3 -m pytest tests/food/test_fruits.py::TestFruit
tests/food/test_fruits.py::TestFruit . [100%]
Example of testing a class
in our project
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 |
|
Parametrize
So far, in our tests, we've had to create individual assert statements to validate different combinations of inputs and expected outputs. However, there's a bit of redundancy here because the inputs always feed into our functions as arguments and the outputs are compared with our expected outputs. To remove this redundancy, pytest has the @pytest.mark.parametrize
decorator which allows us to represent our inputs and outputs as parameters.
1 2 3 4 5 6 7 8 9 10 |
|
python3 -m pytest tests/food/test_is_crisp_parametrize.py ... [100%]
[Line 2]
: define the names of the parameters under the decorator, ex. "fruit, crisp" (note that this is one string).[Lines 3-7]
: provide a list of combinations of values for the parameters from Step 1.[Line 9]
: pass in parameter names to the test function.[Line 10]
: include necessary assert statements which will be executed for each of the combinations in the list from Step 2.
Similarly, we could pass in an exception as the expected result as well:
1 2 3 4 5 6 7 8 9 |
|
Example of parametrize
in our project
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 |
|
Fixtures
Parametrization allows us to reduce redundancy inside test functions but what about reducing redundancy across different test functions? For example, suppose that different functions all have a dataframe as an input. Here, we can use pytest's builtin fixture, which is a function that is executed before the test function.
1 2 3 4 5 6 7 |
|
Note that the name of fixture and the input to the test function are identical (
my_fruit
).
We can apply fixtures to classes as well where the fixture function will be invoked when any method in the class is called.
1 2 3 |
|
Example of fixtures
in our project
In our project, we use fixtures to efficiently pass a set of inputs (ex. Pandas DataFrame) to different testing functions that require them (cleaning, splitting, etc.).
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 |
|
Note that we don't use the df
fixture directly (we pass in df.copy()
) inside our parametrized test function. If we did, then we'd be changing df
's values after each parametrization.
Tip
When creating fixtures around datasets, it's best practice to create a simplified version that still adheres to the same schema. For example, in our dataframe fixture above, we're creating a smaller dataframe that still has the same column names as our actual dataframe. While we could have loaded our actual dataset, it can cause issues as our dataset changes over time (new labels, removed labels, very large dataset, etc.)
Fixtures can have different scopes depending on how we want to use them. For example our df
fixture has the module scope because we don't want to keep recreating it after every test but, instead, we want to create it just once for all the tests in our module (tests/test_data.py
).
function
: fixture is destroyed after every test.[default]
class
: fixture is destroyed after the last test in the class.module
: fixture is destroyed after the last test in the module (script).package
: fixture is destroyed after the last test in the package.session
: fixture is destroyed after the last test of the session.
Functions are lowest level scope while sessions are the highest level. The highest level scoped fixtures are executed first.
Typically, when we have many fixtures in a particular test file, we can organize them all in a
fixtures.py
script and invoke them as needed.
Markers
We've been able to execute our tests at various levels of granularity (all tests, script, function, etc.) but we can create custom granularity by using markers. We've already used one type of marker (parametrize) but there are several other builtin markers as well. For example, the skipif
marker allows us to skip execution of a test if a condition is met. For example, supposed we only wanted to test training our model if a GPU is available:
1 2 3 4 5 6 |
|
We can also create our own custom markers with the exception of a few reserved marker names.
1 2 3 |
|
We can execute them by using the -m
flag which requires a (case-sensitive) marker expression like below:
pytest -m "fruits" # runs all tests marked with `fruits`
pytest -m "not fruits" # runs all tests besides those marked with `fruits`
Tip
The proper way to use markers is to explicitly list the ones we've created in our pyproject.toml file. Here we can specify that all markers must be defined in this file with the --strict-markers
flag and then declare our markers (with some info about them) in our markers
list:
1 2 3 |
|
1 2 3 4 5 6 7 8 |
|
pytest --markers
and we'll receive an error when we're trying to use a new marker that's not defined here.
Coverage
As we're developing tests for our application's components, it's important to know how well we're covering our code base and to know if we've missed anything. We can use the Coverage library to track and visualize how much of our codebase our tests account for. With pytest, it's even easier to use this package thanks to the pytest-cov plugin.
pip install pytest-cov==2.10.1
And we'll add this to our setup.py
script:
1 2 3 4 5 |
|
python3 -m pytest --cov tagifai --cov-report html

Here we're asking for coverage for all the code in our tagifai and app directories and to generate the report in HTML format. When we run this, we'll see the tests from our tests directory executing while the coverage plugin is keeping tracking of which lines in our application are being executed. Once our tests are complete, we can view the generated report (default is htmlcov/index.html
) and click on individual files to see which parts were not covered by any tests. This is especially useful when we forget to test for certain conditions, exceptions, etc.

Warning
Though we have 100% coverage, this does not mean that our application is perfect. Coverage only indicates that a piece of code executed in a test, not necessarily that every part of it was tested, let alone thoroughly tested. Therefore, coverage should never be used as a representation of correctness. However, it is very useful to maintain coverage at 100% so we can know when new functionality has yet to be tested. In our CI/CD lesson, we'll see how to use GitHub actions to make 100% coverage a requirement when pushing to specific branches.
Exclusions
Sometimes it doesn't make sense to write tests to cover every single line in our application yet we still want to account for these lines so we can maintain 100% coverage. We have two levels of purview when applying exclusions:
-
Excusing lines by adding this comment
# pragma: no cover, <MESSAGE>
1 2 3 4
if trial: # pragma: no cover, optuna pruning trial.report(val_loss, epoch) if trial.should_prune(): raise optuna.TrialPruned()
-
Excluding files by specifying them in our
pyproject.toml
configuration:
1 2 3 |
|
The main point is that we were able to add justification to these exclusions through comments so our team can follow our reasoning.
Now that we have a foundation for testing traditional software, let's dive into testing our data and models in the context of machine learning systems.
๐ข Data
So far, we've used unit and integration tests to test the functions that interact with our data but we haven't tested the validity of the data itself. We're going to use the great expectations library to test what our data is expected to look like. It's a library that allows us to create expectations as to what our data should look like in a standardized way. It also provides modules to seamlessly connect with backend data sources such as local file systems, S3, databases, etc. Let's explore the library by implementing the expectations we'll need for our application.
๐ Follow along interactive notebook in the testing-ml repository as we implement the concepts below.
pip install great-expectations==0.15.15
And we'll add this to our setup.py
script:
1 2 3 4 5 6 |
|
First we'll load the data we'd like to apply our expectations on. We can load our data from a variety of sources (filesystem, database, cloud etc.) which we can then wrap around a Dataset module (Pandas / Spark DataFrame, SQLAlchemy).
1 2 3 4 |
|
1 2 3 4 5 6 |
|
Expectations
When it comes to creating expectations as to what our data should look like, we want to think about our entire dataset and all the features (columns) within it.
# Presence of specific features
df.expect_table_columns_to_match_ordered_list(
column_list=["id", "created_on", "title", "description", "tag"]
)
# Unique combinations of features (detect data leaks!)
df.expect_compound_columns_to_be_unique(column_list=["title", "description"])
# Missing values
df.expect_column_values_to_not_be_null(column="tag")
# Unique values
df.expect_column_values_to_be_unique(column="id")
# Type adherence
df.expect_column_values_to_be_of_type(column="title", type_="str")
# List (categorical) / range (continuous) of allowed values
tags = ["computer-vision", "graph-learning", "reinforcement-learning",
"natural-language-processing", "mlops", "time-series"]
df.expect_column_values_to_be_in_set(column="tag", value_set=tags)
Each of these expectations will create an output with details about success or failure, expected and observed values, expectations raised, etc. For example, the expectation df.expect_column_values_to_be_of_type(column="title", type_="str")
would produce the following if successful:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 |
|
and if we have a failed expectation (ex. df.expect_column_values_to_be_of_type(column="title", type_="int")
), we'd receive this output(notice the counts and examples for what caused the failure):
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 |
|
There are just a few of the different expectations that we can create. Be sure to explore all the expectations, including custom expectations. Here are some other popular expectations that don't pertain to our specific dataset but are widely applicable:
- feature value relationships with other feature values โ
expect_column_pair_values_a_to_be_greater_than_b
- row count (exact or range) of samples โ
expect_table_row_count_to_be_between
- value statistics (mean, std, median, max, min, sum, etc.) โ
expect_column_mean_to_be_between
Organization
When it comes to organizing expectations, it's recommended to start with table-level ones and then move on to individual feature columns.
Table expectations
1 2 3 4 5 6 |
|
Column expectations
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 |
|
We can group all the expectations together to create an Expectation Suite object which we can use to validate any Dataset module.
1 2 3 |
|
1 2 3 4 5 6 7 8 9 10 11 |
|
Projects
So far we've worked with the Great Expectations library at the adhoc script / notebook level but we can further organize our expectations by creating a Project.
cd tests
great_expectations init
tests/great_expectations
directory with the following structure:
tests/great_expectations/
โโโ checkpoints/
โโโ expectations/
โโโ plugins/
โโโ uncommitted/
โโโ .gitignore
โโโ great_expectations.yml
Data source
The first step is to establish our datasource
which tells Great Expectations where our data lives:
great_expectations datasource new
What data would you like Great Expectations to connect to?
1. Files on a filesystem (for processing with Pandas or Spark) ๐
2. Relational database (SQL)
What are you processing your files with?
1. Pandas ๐
2. PySpark
Enter the path of the root directory where the data files are stored: ../data
Run the cells in the generated notebook and change the datasource_name
to local_data
. After we run the cells, we can close the notebook (and end the process on the terminal with Ctrl + c) and we can see the Datasource being added to great_expectations.yml
.
Suites
Create expectations manually, interactively or automatically and save them as suites (a set of expectations for a particular data asset).
great_expectations suite new
How would you like to create your Expectation Suite?
1. Manually, without interacting with a sample batch of data (default)
2. Interactively, with a sample batch of data ๐
3. Automatically, using a profiler
Which data asset (accessible by data connector "default_inferred_data_connector_name") would you like to use?
1. labeled_projects.csv
2. projects.csv ๐
3. tags.csv
Name the new Expectation Suite [projects.csv.warning]: projects
tags.csv
and labeled_projects.csv
.

Expectations for projects.csv
Table expectations
1 2 3 4 |
|
Column expectations:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 |
|
Expectations for tags.csv
Table expectations
1 2 |
|
Column expectations:
1 2 3 4 5 6 |
|
Expectations for labeled_projects.csv
Table expectations
1 2 3 4 |
|
Column expectations:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 |
|
All of these expectations have been saved under great_expectations/expectations
:
great_expectations/
โโโ expectations/
โ โโโ labeled_projects.csv
โ โโโ projects.csv
โ โโโ tags.csv
And we can also list the suites with:
great_expectations suite list
Using v3 (Batch Request) API 3 Expectation Suites found: - labeled_projects - projects - tags
To edit a suite, we can execute the follow CLI command:
great_expectations suite edit <SUITE_NAME>
Checkpoints
Create Checkpoints where a Suite of Expectations are applied to a specific data asset. This is a great way of programmatically applying checkpoints on our existing and new data sources.
cd tests
great_expectations checkpoint new CHECKPOINT_NAME
great_expectations checkpoint new projects
great_expectations checkpoint new tags
great_expectations checkpoint new labeled_projects
data_asset_name
(which data asset to run the checkpoint suite on) and expectation_suite_name
(name of the suite to use). For example, the projects
checkpoint would use the projects.csv
data asset and the projects
suite.
Checkpoints can share the same suite, as long the schema and validations are applicable.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 |
|
Validate autofills
Be sure to ensure that the datasource_name
, data_asset_name
and expectation_suite_name
are all what we want them to be (Great Expectations autofills those with assumptions which may not always be accurate).
Repeat these same steps for the tags
and labeled_projects
checkpoints and then we're ready to execute them:
great_expectations checkpoint run projects
great_expectations checkpoint run tags
great_expectations checkpoint run labeled_projects

At the end of this lesson, we'll create a target in our Makefile
that run all these tests (code, data and models) and we'll automate their execution in our pre-commit lesson.
Note
We've applied expectations on our source dataset but there are many other key areas to test the data as well. For example, the intermediate outputs from processes such as cleaning, augmentation, splitting, preprocessing, tokenization, etc.
Documentation
When we create expectations using the CLI application, Great Expectations automatically generates documentation for our tests. It also stores information about validation runs and their results. We can launch the generate data documentation with the following command: great_expectations docs build

By default, Great Expectations stores our expectations, results and metrics locally but for production, we'll want to set up remote metadata stores.
Production
The advantage of using a library such as great expectations, as opposed to isolated assert statements is that we can:
- reduce redundant efforts for creating tests across data modalities
- automatically create testing checkpoints to execute as our dataset grows
- automatically generate documentation on expectations and report on runs
- easily connect with backend data sources such as local file systems, S3, databases, etc.
Many of these expectations will be executed when the data is extracted, loaded and transformed during our DataOps workflows. Typically, the data will be extracted from a source (database, API, etc.) and loaded into a data system (ex. data warehouse) before being transformed there (ex. using dbt) for downstream applications. Throughout these tasks, Great Expectations checkpoint validations can be run to ensure the validity of the data and the changes applied to it. We'll see a simplified version of when data validation should occur in our data workflows in the orchestration lesson.

Learn more about different data systems in our data stack lesson if you're not familiar with them.
๐ค Models
The final aspect of testing ML systems involves how to test machine learning models during training, evaluation, inference and deployment.
Training
We want to write tests iteratively while we're developing our training pipelines so we can catch errors quickly. This is especially important because, unlike traditional software, ML systems can run to completion without throwing any exceptions / errors but can produce incorrect systems. We also want to catch errors quickly to save on time and compute.
- Check shapes and values of model output
1
assert model(inputs).shape == torch.Size([len(inputs), num_classes])
- Check for decreasing loss after one batch of training
1
assert epoch_loss < prev_epoch_loss
- Overfit on a batch
1 2
accuracy = train(model, inputs=batches[0]) assert accuracy == pytest.approx(0.95, abs=0.05) # 0.95 ยฑ 0.05
- Train to completion (tests early stopping, saving, etc.)
1 2 3
train(model) assert learning_rate >= min_learning_rate assert artifacts
- On different devices
1 2
assert train(model, device=torch.device("cpu")) assert train(model, device=torch.device("cuda"))
Note
You can mark the compute intensive tests with a pytest marker and only execute them when there is a change being made to system affecting the model.
1 2 3 |
|
Behavioral testing
Behavioral testing is the process of testing input data and expected outputs while treating the model as a black box (model agnostic evaluation). They don't necessarily have to be adversarial in nature but more along the types of perturbations we may expect to see in the real world once our model is deployed. A landmark paper on this topic is Beyond Accuracy: Behavioral Testing of NLP Models with CheckList which breaks down behavioral testing into three types of tests:
invariance
: Changes should not affect outputs.1 2 3 4
# INVariance via verb injection (changes should not affect outputs) tokens = ["revolutionized", "disrupted"] texts = [f"Transformers applied to NLP have {token} the ML field." for token in tokens] predict.predict(texts=texts, artifacts=artifacts)
['natural-language-processing', 'natural-language-processing']
directional
: Change should affect outputs.1 2 3 4
# DIRectional expectations (changes with known outputs) tokens = ["text classification", "image classification"] texts = [f"ML applied to {token}." for token in tokens] predict.predict(texts=texts, artifacts=artifacts)
['natural-language-processing', 'computer-vision']
minimum functionality
: Simple combination of inputs and expected outputs.1 2 3 4
# Minimum Functionality Tests (simple input/output pairs) tokens = ["natural language processing", "mlops"] texts = [f"{token} is the next big wave in machine learning." for token in tokens] predict.predict(texts=texts, artifacts=artifacts)
['natural-language-processing', 'mlops']
Adversarial testing
Each of these types of tests can also include adversarial tests such as testing with common biased tokens or noisy tokens, etc.
1 2 3 4 5 |
|
['natural-language-processing', 'other']
And we can convert these tests into systematic parameterized tests:
mkdir tests/model
touch tests/model/test_behavioral.py
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 |
|
View tests/model/test_behavioral.py
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 |
|
Inference
When our model is deployed, most users will be using it for inference (directly / indirectly), so it's very important that we test all aspects of it.
Loading artifacts
This is the first time we're not loading our components from in-memory so we want to ensure that the required artifacts (model weights, encoders, config, etc.) are all able to be loaded.
1 2 3 |
|
Prediction
Once we have our artifacts loaded, we're readying to test our prediction pipelines. We should test samples with just one input, as well as a batch of inputs (ex. padding can have unintended consequences sometimes).
1 2 3 4 5 6 7 8 9 10 11 12 |
|
Makefile
Let's create a target in our Makefile
that will allow us to execute all of our tests with one call:
# Test
.PHONY: test
test:
pytest -m "not training"
cd tests && great_expectations checkpoint run projects
cd tests && great_expectations checkpoint run tags
cd tests && great_expectations checkpoint run labeled_projects
make test
Testing vs. monitoring
We'll conclude by talking about the similarities and distinctions between testing and monitoring. They're both integral parts of the ML development pipeline and depend on each other for iteration. Testing is assuring that our system (code, data and models) passes the expectations that we've established offline. Whereas, monitoring involves that these expectations continue to pass online on live production data while also ensuring that their data distributions are comparable to the reference window (typically subset of training data) through \(t_n\). When these conditions no longer hold true, we need to inspect more closely (retraining may not always fix our root problem).
With monitoring, there are quite a few distinct concerns that we didn't have to consider during testing since it involves (live) data we have yet to see.
- features and prediction distributions (drift), typing, schema mismatches, etc.
- determining model performance (rolling and window metrics on overall and slices of data) using indirect signals (since labels may not be readily available).
- in situations with large data, we need to know which data points to label and upsample for training.
- identifying anomalies and outliers.
We'll cover all of these concepts in much more depth (and code) in our monitoring lesson.
Resources
- Great Expectations
- The ML Test Score: A Rubric for ML Production Readiness and Technical Debt Reduction
- Beyond Accuracy: Behavioral Testing of NLP Models with CheckList
- Robustness Gym: Unifying the NLP Evaluation Landscape
To cite this content, please use:
1 2 3 4 5 6 |
|