Checking Library Versions

Before you write any machine learning code, you need the right tools installed. The first real application in this material is built around a hobby botanist who has found some iris flowers and wants to tell their species apart. For every iris she has measured, she has recorded four quantities in centimeters: the length and width of the petals, and the length and width of the sepals. Those four numbers describe one flower.

The application comes with a version requirement attached. The requirement is short, but it teaches a lesson that applies to every library you will use: the version of a library you have installed determines which functions and import paths your code can rely on. This article uses that one requirement to show how to check any version requirement you meet in the same way.

Start with the library itself. scikit-learn is a Python library for machine learning. Different releases of it do not always offer the same modules or the same import paths. A program that runs cleanly on one release can fail on an older one, because a function or module it depends on was not yet available.

Read the chain from left to right. The requirement says what release the material expects. The installed release is what your Python environment actually provides. An import statement is resolved against the installed release, not against the requirement written in the text. So the code that runs depends on the installed release, and a gap between the two shows up as a failed import rather than as a subtle wrong answer.

The number 0.18 is a floor, not a target. Anything at 0.18 or above is what the material expects. You are not being asked to install one exact release. The material also recommends a release at least as recent as the authors used, so a newer release is a safe choice, while an older one below 0.18 is not.

The reason the floor sits at 0.18 is the model_selection module. That module was added in version 0.18. Any code that imports from model_selection depends on the release that contains it. This is what makes the version number concrete: it is not an arbitrary cutoff, but the first release in which a specific module exists.

Consider what happens on an earlier release. If you install a version below 0.18, imports from model_selection will not resolve as written. You will then need to adjust those imports so they match what your installed version provides. The underlying idea is the same one shown in the pipeline above: the import is only as good as the release it is resolved against.

The botanist's situation is a supervised learning problem. In supervised learning, we have measurements for which we already know the correct output. The known correct outputs are what set this setting apart: the model is given examples together with their answers and learns the relationship between them.

The idea of a label is simple, but it is easy to confuse with the measurements. The measurements describe the flower. The label names what the flower is. The botanist wants a model that learns from measurements whose species is already known, so that it can predict the species of a new iris from its measurements alone.

Read the diagram as a mapping. One flower's four measurements are the features. Each flower carries exactly one label, which is its species. The label must be one of the classes, which are the full set of allowed answers. The measurements tell the model what to look at, and the label tells it what the right answer was for that flower.

Within supervised learning, this application is a classification problem. In a classification problem, we want to predict one of several options. Here the options are the species of iris, and the possible outputs are called classes.

Every iris in the dataset belongs to one of three classes: setosa, versicolor, or virginica. Because there are exactly three possible outputs, the problem is described as a three-class classification problem. The number of classes is a property of the problem, and it tells you how many distinct answers the model is allowed to give.

The two terms in this article are easy to mix up because they describe the same problem from different angles. Supervised learning describes how the model is trained: it learns from data where the correct outputs are known. Classification describes what kind of output is wanted: one option out of several named categories. The iris application is both at once.

The method you learn here works for any version requirement. Start by finding out which release is installed, then compare it with the floor the material states. The decisions below follow that order, and each branch tells you what to do next.