Update readme.md

Author: tirthajyoti

Pytest/readme.md

@@ -46,11 +46,14 @@ test_linear_model.py::test_raised_exception PASSED
 
 ### Notes on the test module
 
-- Note, how the `test_linear_model.py` contains 9 functions with names starting with `test...`. Those contain the actual test code. It also has a couple of data constructor functions whose names do not start with `test...` and they are ignored by Pytest.
+- Note, how the `test_linear_model.py` contains 9 functions with names starting with `test...`. Those contain the actual test code. 
+- It also has a couple of data constructor functions (`random_data_constructor` and `fixed_data_constructor`) whose names do not start with `test...` and they are ignored by Pytest. The `random_data_constructor` even takes a `noise_mag` argument which is used to control the magnitude of noise to test the expected behavior of a linear regression algorithm. Refer to the `test_noise_impact` function for this.
 
 - Note that we need to import a variety of libraries to test all kind of things e.g. we imported libraries like `joblib`, `os`, `sklearn`, `numpy`, and of course, the `train_linear_model` function from the `linear_model` module.
 - Note the **clear and distinctive names** for the testing functions e.g. `test_model_return_object()` which only checks the returned object from the `train_linear_model` function, or the `test_model_save_load()` which checks whether the saved model can be loaded properly (but does not try to make predictions or anything). **Always write short and crisp test functions with a singular focus**. 
-- For checking the predictions i.e. whether the trained model really works or not, we have the `test_loaded_model_works()` function which uses a fixed data generator with no noise (as compared to other cases, where we can use a random data generator with random noise). It passes on the fixed `X` and `y` data, loads the trained model, checks if the R^2 ([regression coefficient](https://www.geeksforgeeks.org/python-coefficient-of-determination-r2-score/)) scores are perfectly equal to 1.0 (true for a fixed dataset with no noise) and then compare the model predictions with the original ground truth `y` vector. Note, how it uses a **special Numpy testing function `np.testing.assert_allclose` instead of the regular `assert` statement**. This is to avoid any potential numerical precision issues associated with the model data i.e. Numpy arrays and the prediction algorithm involving linear algebra operations.
-- Take a look at the `random_data_constructor` and `fixed_data_constructor` functions too to see how they are designed and used in the test code. The `random_data_constructor` even takes a `noise_mag` argument which is used to control the magnitude of noise to test the expected behavior of a linear regression algorithm. Refer to the `test_noise_impact` function for this.
-- Finally, take a look at the `test_wrong_input_raises_assertion` function which tests **if the original modeling function raises the correct Exception and messages** when fed with various types of wrong input arguments. Always remember that **Pytest (or any testing module for that matter) is not to be used for error-handling** (e.g. a wrong input type). The developer (ML sofware engineer in this case) must write the necessary error-checking and handling code. In this example, we write a bunch of `assert` statements in the original `train_linear_model` function and wrap them around a whole `try...except` block. The test code only checks if we are returning the `AssertionType` error for those wrong input cases and if the correct error message is being printed for the end user.
-- The `test_raised_exception` function tests the **rise of other exception types (i.e. different from the `AssertionError` that could be raised by the `assert` statements in the function module)** during runtime using a special Pytest feature - [the `pytest.raises` context manager](https://docs.pytest.org/en/reorganize-docs/new-docs/user/pytest_raises.html).
+- For checking the predictions i.e. whether the trained model really works or not, we have the `test_loaded_model_works()` function which uses a fixed data generator with no noise (as compared to other cases, where we can use a random data generator with random noise). It passes on the fixed `X` and `y` data, loads the trained model, checks if the R^2 ([regression coefficient](https://www.geeksforgeeks.org/python-coefficient-of-determination-r2-score/)) scores are perfectly equal to 1.0 (true for a fixed dataset with no noise) and then compare the model predictions with the original ground truth `y` vector. 
+- Note, how the aforementioned function uses a **special Numpy testing function `np.testing.assert_allclose` instead of the regular `assert` statement**. This is to avoid any potential numerical precision issues associated with the model data i.e. Numpy arrays and the prediction algorithm involving linear algebra operations.
+- The `test_model_works_data_range_sign_change` function tests the expected behavior of a linear regression estimator - that the regression scores will still be 1.0 no matter the range of the data (scaling the data by 10e-9 or 10e+9). It also changes the expected behavior if the data flips sign somehow.
+- The `test_additive_invariance` function tests the additive invariance property of the linear regression estimator. **Similarly, you should think about the special properties of your ML estimator and write customized tests for them**.
+- Take a look at the `test_wrong_input_raises_assertion` function which tests **if the original modeling function raises the correct Exception and messages** when fed with various types of wrong input arguments. Always remember that **Pytest (or any testing module for that matter) is not to be used for error-handling** (e.g. a wrong input type). The developer (ML sofware engineer in this case) must write the necessary error-checking and handling code. In this example, we write a bunch of `assert` statements in the original `train_linear_model` function and wrap them around a whole `try...except` block. The test code only checks if we are returning the `AssertionType` error for those wrong input cases and if the correct error message is being printed for the end user.
+- Finally, the `test_raised_exception` function tests the **rise of other exception types (i.e. different from the `AssertionError` that could be raised by the `assert` statements in the function module)** during runtime using a special Pytest feature - [the `pytest.raises` context manager](https://docs.pytest.org/en/reorganize-docs/new-docs/user/pytest_raises.html).