Using custom Python classes in your Streamlit app

If you are building a complex Streamlit app or working with existing code, you may have custom Python classes defined in your script. Common examples include the following:

  • Defining a @dataclass to store related data within your app.
  • Defining an Enum class to represent a fixed set of options or values.
  • Defining custom interfaces to external services or databases not covered by st.connection.

Because Streamlit reruns your script after every user interaction, custom classes may be redefined multiple times within the same Streamlit session. This may result in unwanted effects, especially with class and instance comparisons. Read on to understand this common pitfall and how to avoid it.

We begin by covering some general-purpose patterns you can use for different types of custom classes, and follow with a few more technical details explaining why this matters. Finally, we go into more detail about Using Enum classes specifically, and describe a configuration option which can make them more convenient.

This is the recommended, general solution. If possible, move class definitions into their own module file and import them into your app script. As long as you are not editing the file where your class is defined, Streamlit will not re-import it with each rerun. Therefore, if a class is defined in an external file and imported into your script, the class will not be redefined during the session.

Example: Move your class definition

Try running the following Streamlit app where MyClass is defined within the page's script. isinstance() will return True on the first script run then return False on each rerun thereafter.

# app.py import streamlit as st # MyClass gets redefined every time app.py reruns class MyClass: def __init__(self, var1, var2): self.var1 = var1 self.var2 = var2 if "my_instance" not in st.session_state: st.session_state.my_instance = MyClass("foo", "bar") # Displays True on the first run then False on every rerun st.write(isinstance(st.session_state.my_instance, MyClass)) st.button("Rerun")

If you move the class definition out of app.py into another file, you can make isinstance() consistently return True. Consider the following file structure:

myproject/ ├── my_class.py └── app.py
# my_class.py class MyClass: def __init__(self, var1, var2): self.var1 = var1 self.var2 = var2
# app.py import streamlit as st from my_class import MyClass # MyClass doesn't get redefined with each rerun if "my_instance" not in st.session_state: st.session_state.my_instance = MyClass("foo", "bar") # Displays True on every rerun st.write(isinstance(st.session_state.my_instance, MyClass)) st.button("Rerun")

Streamlit only reloads code in imported modules when it detects the code has changed. Thus, if you are actively editing the file where your class is defined, you may need to stop and restart your Streamlit server to avoid an undesirable class redefinition mid-session.

For classes that store data (like dataclasses), you may be more interested in comparing the internally stored values rather than the class itself. If you define a custom __eq__ method, you can force comparisons to be made on the internally stored values.

Example: Define __eq__

Try running the following Streamlit app and observe how the comparison is True on the first run then False on every rerun thereafter.

import streamlit as st from dataclasses import dataclass @dataclass class MyDataclass: var1: int var2: float if "my_dataclass" not in st.session_state: st.session_state.my_dataclass = MyDataclass(1, 5.5) # Displays True on the first run the False on every rerun st.session_state.my_dataclass == MyDataclass(1, 5.5) st.button("Rerun")

Since MyDataclass gets redefined with each rerun, the instance stored in Session State will not be equal to any instance defined in a later script run. You can fix this by forcing a comparison of internal values as follows:

import streamlit as st from dataclasses import dataclass @dataclass class MyDataclass: var1: int var2: float def __eq__(self, other): # An instance of MyDataclass is equal to another object if the object # contains the same fields with the same values return (self.var1, self.var2) == (other.var1, other.var2) if "my_dataclass" not in st.session_state: st.session_state.my_dataclass = MyDataclass(1, 5.5) # Displays True on every rerun st.session_state.my_dataclass == MyDataclass(1, 5.5) st.button("Rerun")

The default Python __eq__ implementation for a regular class or @dataclass depends on the in-memory ID of the class or class instance. To avoid problems in Streamlit, your custom __eq__ method should not depend the type() of self and other.

Another option for classes that store data is to define serialization and deserialization methods like to_str and from_str for your class. You can use these to store class instance data in st.session_state rather than storing the class instance itself. Similar to pattern 2, this is a way to force comparison of the internal data and bypass the changing in-memory IDs.

Example: Save your class instance as a string

Using the same example from pattern 2, this can be done as follows:

import streamlit as st from dataclasses import dataclass @dataclass class MyDataclass: var1: int var2: float def to_str(self): return f"{self.var1},{self.var2}" @classmethod def from_str(cls, serial_str): values = serial_str.split(",") var1 = int(values[0]) var2 = float(values[1]) return cls(var1, var2) if "my_dataclass" not in st.session_state: st.session_state.my_dataclass = MyDataclass(1, 5.5).to_str() # Displays True on every rerun MyDataclass.from_str(st.session_state.my_dataclass) == MyDataclass(1, 5.5) st.button("Rerun")

For classes that are used as resources (database connections, state managers, APIs), consider using the cached singleton pattern. Use @st.cache_resource to decorate a @staticmethod of your class to generate a single, cached instance of the class. For example:

import streamlit as st class MyResource: def __init__(self, api_url: str): self._url = api_url @st.cache_resource(ttl=300) @staticmethod def get_resource_manager(api_url: str): return MyResource(api_url) # This is cached until Session State is cleared or 5 minutes has elapsed. resource_manager = MyResource.get_resource_manager("http://example.com/api/")

When you use one of Streamlit's caching decorators on a function, Streamlit doesn't use the function object to look up cached values. Instead, Streamlit's caching decorators index return values using the function's qualified name and module. So, even though Streamlit redefines MyResource with each script run, st.cache_resource is unaffected by this. get_resource_manager() will return its cached value with each rerun, until the value expires.

So what's really happening here? We'll consider a simple example to illustrate why this is a pitfall. Feel free to skip this section if you don't want to deal more details. You can jump ahead to learn about Using Enum classes.

Set aside Streamlit for a moment and think about this simple Python script:

from dataclasses import dataclass @dataclass class Student: student_id: int name: str Marshall_A = Student(1, "Marshall") Marshall_B = Student(1, "Marshall") # This is True (because a dataclass will compare two of its instances by value) Marshall_A == Marshall_B # Redefine the class @dataclass class Student: student_id: int name: str Marshall_C = Student(1, "Marshall") # This is False Marshall_A == Marshall_C

In this example, the dataclass Student is defined twice. All three Marshalls have the same internal values. If you compare Marshall_A and Marshall_B they will be equal because they were both created from the first definition of Student. However, if you compare Marshall_A and Marshall_C they will not be equal because Marshall_C was created from the second definition of Student. Even though both Student dataclasses are defined exactly the same, they have different in-memory IDs and are therefore different.

In Streamlit, you probably don't have the same class written twice in your page script. However, the rerun logic of Streamlit creates the same effect. Let's use the above example for an analogy. If you define a class in one script run and save an instance in Session State, then a later rerun will redefine the class and you may end up comparing a Mashall_C in your rerun to a Marshall_A in Session State. Since widgets rely on Session State under the hood, this is where things can get confusing.

Several Streamlit UI elements, such as st.selectbox or st.radio, accept multiple-choice options via an options argument. The user of your application can typically select one or more of these options. The selected value is returned by the widget function. For example:

number = st.selectbox("Pick a number, any number", options=[1, 2, 3]) # number == whatever value the user has selected from the UI.

When you call a function like st.selectbox and pass an Iterable to options, the Iterable and current selection are saved into a hidden portion of Session State called the Widget Metadata.

When the user of your application interacts with the st.selectbox widget, the broswer sends the index of their selection to your Streamlit server. This index is used to determine which values from the original options list, saved in the Widget Metadata from the previous page execution, are returned to your application.

The key detail is that the value returned by st.selectbox (or similar widget function) is from an Iterable saved in Session State during a previous execution of the page, NOT the values passed to options on the current execution. There are a number of architectural reasons why Streamlit is designed this way, which we won't go into here. However, this is how we end up comparing instances of different classes when we think we are comparing instances of the same class.

The above explanation might be a bit confusing, so here's a pathological example to illustrate the idea.

import streamlit as st from dataclasses import dataclass @dataclass class Student: student_id: int name: str Marshall_A = Student(1, "Marshall") if "B" not in st.session_state: st.session_state.B = Student(1, "Marshall") Marshall_B = st.session_state.B options = [Marshall_A,Marshall_B] selected = st.selectbox("Pick", options) # This comparison does not return expected results: selected == Marshall_A # This comparison evaluates as expected: selected == Marshall_B

As a final note, we used @dataclass in the example for this section to illustrate a point, but in fact it is possible to encounter these same problems with classes, in general. Any class which checks class identity inside of a comparison operator—such as __eq__ or __gt__—can exhibit these issues.

The Enum class from the Python standard library is a powerful way to define custom symbolic names that can be used as options for st.multiselect or st.selectbox in place of str values.

For example, you might add the following to your streamlit page:

from enum import Enum import streamlit as st # class syntax class Color(Enum): RED = 1 GREEN = 2 BLUE = 3 selected_colors = set(st.multiselect("Pick colors", options=Color)) if selected_colors == {Color.RED, Color.GREEN}: st.write("Hooray, you found the color YELLOW!")

If you're using the latest version of Streamlit, this Streamlit page will work as it appears it should. When a user picks both Color.RED and Color.GREEN, they are shown the special message.

However, if you've read the rest of this page you might notice something tricky going on. Specifically, the Enum class Color gets redefined every time this script is run. In Python, if you define two Enum classes with the same class name, members, and values, the classes and their members are still considered unique from each other. This should cause the above if condition to always evaluate to False. In any script rerun, the Color values returned by st.multiselect would be of a different class than the Color defined in that script run.

If you run the snippet above with Streamlit version 1.28.0 or less, you will not be able see the special message. Thankfully, as of version 1.29.0, Streamlit introduced a configuration option to greatly simplify the problem. That's where the enabled-by-default enumCoercion configuration option comes in.

When enumCoercion is enabled, Streamlit tries to recognize when you are using an element like st.multiselect or st.selectbox with a set of Enum members as options.

If Streamlit detects this, it will convert the widget's returned values to members of the Enum class defined in the latest script run. This is something we call automatic Enum coercion.

This behavior is configurable via the enumCoercion setting in your Streamlit config.toml file. It is enabled by default, and may be disabled or set to a stricter set of matching criteria.

If you find that you still encounter issues with enumCoercion enabled, consider using the custom class patterns described above, such as moving your Enum class definition to a separate module file.

forum

Still have questions?

Our forums are full of helpful information and Streamlit experts.