Python Panel¶

A Python panel is a fully custom panel that allows you to use your own methods to explore, visualize, and analyze your data. When writing a Python Panel, you can use all of the standard data science Python modules, including: comet_ml, matplotlib, numpy, pandas, plotly, scikit-learn, seaborn, and scipy. For a full list, see below.

Use the Python panel in Comet to create custom visualizations when they are not already offered in the built-in panels accessible from the Panels Gallery.

About the Python Panel Editor¶

The Panel Editor for the Python panel type has a unique layout, different from the default editor for all other panels.

The Python Panel Editor consists of two panes:

• The Editor pane to the left is used to provide and edit your custom code, set visualization options, add a description, and apply any relevant filters.
• The Preview pane displays a preview of the panel as currently defined in the editor, and running in the current project.You can also update the default panel name from here.

Get started with the Python Panel¶

Python Panels are powered by Comet Compute Beta which allow you to run you code on a personal secure server. Compute Compute uses CPython 3.9, so all of your regular Python libraries should work. In addition, you can use the comet_ml.ui (described below) or any streamlit code.

A simple Python Panel¶

Get a feel for Python Panels with this simple example.

Enter the following code in the Code tab:

from comet_ml import ui

ui.display("My first Python Panel!")

Click Run to display the output in the Panel Preview area on the right-hand side.

To display items on the Panel canvas, you need to import the comet_ml.ui library, and call ui.display(). See ui for addition details on comet_ml.ui.display().

See more examples on Python Panels.

Python SDK methods for Python Panels¶

API, ui, and APIExperiment classes in the Python SDK provide methods that you can call in your Python Panels code.

API¶

The Comet API class includes a number of methods that are only available within Python Panels:

• API.get_panel_options(): The Panel options can be edited when in the Options tab of the Panel Code editor, and set when you are adding a Panel instance to a View.

• API.get_panel_experiment_keys(): This will return a list containing all of the current Experiment keys. Specifically:

  • If on a Project View, the Experiment keys from the visible experiment table page.
  • If a filter is set, all of the matching Experiments.
  • If a filter is set for this Panel, the matching Experiments.

• API.get_panel_experiments(): Like the API.get_experiment_keys() method, but returns a list of APIExperiments rather than just their keys.

• API.get_panel_project_id(): Get Project ID. That information can be useful to retrieve other Project-level information.
• API.get_panel_project_name(): Get Project name. That can be useful in creating report-like Panels.
• API.get_panel_workspace(): Get Workspace name. That can be useful in creating report-like Panels.
• API.get_panel_metrics_names(): Get the names of all metrics logged for all experiments in this Project.

ui¶

The Comet ui methods can be used within Python Panels to control the rendering of Panels. ui contains three sets of methods:

• Display methods: To visualize different kinds of Python objects on the Panel canvas area
• Widgets methods: To add interactivity to Panels through elements that have a GUI representation
• Utility methods: To update the styling of Panels

Examples:

To use comet_ml.ui, you need only import it:

from comet_ml import ui

choice = ui.dropdown("Choose one:", ["A", "B", "C"])
ui.display("You picked", choice)

streamlit¶

Python Panels running on Comet Compute can also use streamlit functions. The only limitation is that the streamlit sidebar is hidden.

Examples:

To use streamlit, you need only import it:

import streamlit as st

choice = st.selectbox("Choose one:", ["A", "B", "C"])
st.write("You picked", choice)

See docs.streamlit.io for more information.

APIExperiment¶

When you call methods such as api.get_panel_experiments() or api.query() then you get back a list of APIExperiment objects.

The Comet APIExperiment methods can be used within Python Panels. All get_* methods in this class are available in Python Panels. Here are some of the main ones:

• APIExperiment.get_asset(): Get an asset.
• APIExperiment.get_asset_list(): Get a list of assets associated with the Experiment.
• APIExperiment.get_metadata(): Get the metadata associated with this Experiment.
• APIExperiment.get_metrics(): Get all of the logged metrics.
• APIExperiment.get_metrics_summary(): Return the Experiment metrics summary.
• APIExperiment.get_model_asset_list(): Get an Experiment model's asset list by model name.

Debugging¶

Note that using print() will not display items in the panel display. However, you can simply put the value that you wish to display on a line by itself. For example, change this:

# Don't do this:
print(value)

into the following:

# Do this
value

value can be any valid Python value (string, number, bool, etc). You can even get information on Python modules, functions, and objects with this. This can be useful for debugging. For more information, see Streamlit magic commands.

Use standard Python modules¶

Python Panels allows you to use many other Python support modules, including:

• altair
• biopython
• bokeh
• boto3
• comet_ml
• fastparquet
• geopandas
• jsonschema
• matplotlib
• msgpack
• nltk
• numpy
• opencv-python
• pandas
• Pillow (PIL - Python Image Library)
• plotly
• scikit-learn
• scipy
• simplejson
• streamlit

To use a Python package that is not listed here, contact us through our Slack channel.

Know the limitations¶

• Python Panels do not automatically update upon receiving new logged metrics. This is by design, as your Python Panel may be expensive to compute, and you may not wish to have it automatically refresh whenever new data is received.
• These are the supported widget types, but you can also use any streamlit widget.
• All of your Python Panels share the same Python environment. There are a few widely-used ideoms that you will want to avoid because of this. For example:

Instead of using:

for experiment_key in metrics:
    for metric in metrics[experiment_key]["metrics"]:
        # DO NOT DO THIS:
        plt.plot(metric['steps'], metric['values'])
ui.display(plt)

you should use:

figure, ax = plt.subplots()
for experiment_key in metrics:
    for metric in metrics[experiment_key]["metrics"]:
        ax.plot(metric['steps'], metric['values'])
ui.display(figure)

That is, insead of using matplotlib's pyplot global API, you should use local figure and ax.

Backward Incompatibilities¶

There are a few differences between the previous version of Python Panels and those supported by Comet Compute. In the following list, st stands for streamlit, ui stands for comet_ml.ui, and api stands for comet_ml.API().

• ui.sidebar() has been removed, and the sidebar created with st.sidebar() is not visible. Users can use st.columns() or st.tabs() instead
• api.project_id has been removed; use api.get_panel_project_id() instead
• All of the api.get_experiment_*() methods have been moved to APIExperiment().get_*() methods
• The module js has been removed
• ui.get_theme_names() has been removed
• The theme option in ui.display() is no longer supported
• Experiment.get_asset(..., "raw") should now be Experiment.get_asset(..., "binary")
• It used to be the case if a user defined a main() method that it would be called automatically and was used to separate long-running code from other code. Support for main() has been removed. Now, users can use streamlit cache and state functions.

If you find any additional limitations or have a request, let us know.

Learn more¶

• Python SDK: API
• Python SDK: ui
• Python SDK: APIExperiment
• Examples for Comet Python Panels