Python Panels - BETA

Welcome to the Comet Python Panel Beta!

Comet Python Code Panels allow you to write custom visualizations the same way that you write all of your other scripts and using the Python modules that you know and love.

When writing a Python Panel, you can use all of the standard data science Python modules, including: astropy, autograd, biopython, bokeh, comet_ml, matplotlib, nltk, numpy, pandas, pillow, scikit-image, scikit-learn, and scipy.

If you would like to jump to the examples, they are here:

Steps to creating a Custom Python Panel

  1. Go to a Project or Experiment View.

  2. Click on "Add Panel" (either in the panel area on the Project View, or on the Panel Tab of the Experiment View).

  3. Click on + Create New

  4. That should display the following screen:

The left-hand portion of the Panel editor contains tabs for:

Code tab

Here you enter on the Python code for this panel. Every saved change creates a new version that you can always revert to at any time.

Keyboard short cuts:

PC short cut Mac short cut Functionality
Ctrl-S Cmd-S Save
Ctrl-[ Cmd-[ Indent
Ctrl-] Cmd-] Unindent
Ctrl-F Cmd-F Start searching
Ctrl-G Cmd-G Find next
Shift-Ctrl-G Shift-Cmd-G Find previous
Shift-Ctrl-F Cmd-Option-F Replace
Shift-Ctrl-R Shift-Cmd-Option-F Replace all

Options tab

These are options that you can set for each instance of a Panel. They are useful for having items that can be varied without having to edit the code. For example, you could have an option for a title of a chart.

Description tab

A description of the Panel and thumbnail of the Panel. Click the image to create or edit the thumbnail. We suggest putting author, and details about how the Panel works in the description. This is searchable text on the Panel Gallery.

Query tab

You can provide a default query for your Panel. For example, perhaps your Panel only works for experiments with a particular hyperparameter value. You can define the query here. Of course, many Panels will work without setting a specific filter.

You can override this when you add an instance of the Panel to a Project View.

A Simple Python Panel

Enter the following code in the Code tab:

```python from comet_ml import ui

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

Click on the Run button 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().

Info

Python Panels do not automatically update with new experiments nor 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.

The following sections provides details on accessing Comet-logged data via the API, details on the User Interface ui and end-to-end sample panels.

API

To use comet_ml.API, first import it and make an instance of it:

```python from comet_ml import API

api = API() ```

Then you will have access to the following methods. You can find additional documentation on the full Comet Python Panels API SDK.

API Panel Methods

These methods provide functionality related to the current panel.

API.get_panel_options()

Get the JSON Options data as a Python dictionary.

Example:

```python from comet_ml import API

api = API() options = 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()

Get the experiment keys selected for this Panel.

Example:

```python from comet_ml import API

api = API() experiment_keys = 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 Query is set for this Panel, the matching experiments

API.get_panel_experiments()

Get the APIExperiments selected for this Panel.

Example:

```python from comet_ml import API

api = API() experiment = api.get_panel_experiments() ```

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

You can find additional documentation at Python Panels APIExperiment.

API.get_panel_project_id()

Get the project_id for this Panel.

Example:

```python from comet_ml import API

api = API() project_id = api.get_panel_project_id() ```

The project_id can be useful to retrieve other Project-level information.

API.get_panel_project_name()

Get the project_name for this Panel.

Example:

```python from comet_ml import API

api = API() project_name = api.get_panel_project_name() ```

The Project name could be useful in creating report-like Panels.

API.get_panel_workspace()

Get the workspace (name) for this Panel.

Example:

```python from comet_ml import API

api = API() workspace = api.get_panel_workspace() ```

The Workspace name could 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.

Example:

```python from comet_ml import API, ui

api = API() metrics_names = api.get_panel_metrics_names() metric_name = ui.dropdown("Select a metric name:", metrics_names) ```

As shown, the metrics_names could be useful in creating a dropdown selection for plotting.

Info

Note that this list includes all metrics, and system metrics. If you wish, you could use only those names starting (or not starting) with "sys".

ui

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

```python from comet_ml import ui

ui.display("Hello") ```

ui contains the following methods:

ui Display methods

  • ui.display()
  • ui.display_figure()
  • ui.display_image()
  • ui.display_text()
  • ui.display_markdown()

These methods are described below.

ui.display()

The ui.display() method is used to visualize different kinds of Python objects on the Panel canvas area.

Signature and Use:

python ui.display(*items, format=None, **kwargs)

You can use ui.display() on one (or more) of any of the following types of Python items:

  1. Pillow Python Images Library (PIL) images
  2. HTML strings (including SVG images)
  3. Matplotlib figures
  4. Plotly plots
  5. Pandas' Dataframes
  6. Any object that has a _repr_*_ method

In addition, there are specialized display methods for text, markdown, and images represented as raw strings (as logged as image assets, for example). We'll describe each of those methods below.

If you wish to send multiple items to the display area, pass them to ui.display() or call ui.display() repeatedly:

```python from comet_ml import ui

ui.display("hello") ui.display("world")

or

ui.display("hello", "world") ```

The format argument can be either "text", "markdown", "html" or None. The default is "html"/None.

kwargs are optional, and varying depending on the type of item being displayed. For example, if the item is a pandas Dataframe, then you may also pass in these keyword arguments:

  • theme: the name of a color theme (see below)
  • font_size: name of the size, eg 'normal'
  • font_family: name of the font family
  • text_align: which side to align to
  • width: the width of table, eg, 'auto'
  • index: the index to highlight
  • even_color: color of even rows
  • even_bg_color: background color of even rows

Theme names: 'yellow light', 'grey light', 'blue light', 'orange light', 'green light', 'red light', 'yellow dark', 'grey dark', 'blue dark', 'orange dark', 'green dark', 'red dark', or 'comet'.

For more details on ui.display(), see the Python Panel Examples.

ui.display_figure()

This method is used to display a matplotlib figure.

python ui.display_figure(plt) ui.display_figure(figure)

Displays a matplotlib figure.

Examples:

```python from comet_ml import ui import matplotlib.pyplot as plt

plt commands here

ui.display(plt) ```

Or calling with the figure:

```python from comet_ml import ui import matplotlib.pyplot as plt

fig, ax = plt.subplots()

Figure commands here

ui.display(fig) ```

ui.display_image()

This method is used to display an image, either from a logged asset, or from a PIL Image.

Example from logged asset:

```python from comet_ml import API, ui

api = API()

for experiment in api.get_panel_experiments(): for json in experiment.get_asset_list("image"): if json["fileName"].endswith("jpg"): data = experiment.get_asset(json["assetId"]) ui.display_image(data, format="jpg") ```

Displays image strings in the given format ("jpg", "gif", or "png").

Example from PIL Image:

```python from comet_ml import ui from PIL import Image import random

Create a PIL Image

image = Image.new("RGB", (500, 500))

process image here

ui.display(image) ```

For more details, see the Python Panel Examples.

ui.display_text()

This method is useful for displaying plain text.

python ui.display_text(text)

Displays text that otherwise would have characters interpreted as HTML.

ui.display_markdown()

This method is useful for displaying text formatted as markdown.

python ui.display_markdown(*text)

Displays text as markdown, ignoring any indentation.

ui Widget methods

This section describes so-called widgets. These are elements that have a GUI representation and trigger a change to re-run your code. This trigger can be caused by a widget's event firing (such as clicking a button) or (in the case of the input widget) the value and focus changes. This is best shown through an example.

Consider the following Python code:

```python from comet_ml import ui

if ui.button("Click me!"): ui.display("You clicked the button") ui.display("Now waiting for you to click the button again") else: ui.display("Waiting for you to click the button") ```

Running this code in the Python Code Panel editor looks like the following:

Project View

First, note that the ui component is used directly by calling ui.button().

Also note that there is no explicit loop. However, whenever a widget is changed (or clicked) your code will run again, this time reflecting the change. For this example, the flow is:

  1. The code is run.
  2. The message "Waiting for you to click the button" is displayed.
  3. You click the button.
  4. The code runs again.
  5. This time, the messages "You clicked the button" and "Now waiting for you to click the button again" are displayed.
  6. If you click the button again, you go back to step 4, and the cycle repeats.

All of the following widgets have this same effect. Therefore, you will need to be mindful when writing Python Code Panels that the values of ui components will change, ordering of the widgets matter, and that your code will run repeatedly. This has a number of implications, such as all of the code will run repeatedly (even long-running initialization code). However, there are techniques to handle such issues (see below).

ui.dropdown()

The ui.dropdown() element does two things at once:

  1. creates a dropdown (selection) list of options on the Panel canvas
  2. returns the selected item

Signature and Use:

python choice = ui.dropdown(label, options, index=0, format_func=None, key=None, on_change=None, args=None, kwargs=None, multiple=False, classes=None)

Let's take a look at a specific example.

```python from comet_ml import ui

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

When you first run this code, you will see:

Note that this is very different from the usual manner of creating GUI elements. In this manner, there are no "callbacks" but merely the above code. By default, the dropdown has been shown on the screen and the default options (index=0) has been selected. The code continues, and so you see choice "A" already set as the choice.

If you then select a different item, your code runs again, updating the GUI and the selected item:

If you would like to separate code that should only run once (say, because it is expensive to compute) you can separate the code to run when the GUI is updated by placing it in a main function, like this:

```python from comet_ml import ui

Code that is expensive to run:

choices = ...

def main(): # The fast GUI-based code:

choice = ui.dropdown("Choose one:", choices)
ui.display("You picked", choice)

```

You may provide an parameter index as a numeric value representing the row to show as the initial choice.

The format_func is a function that takes a row from options, and returns a string to be used in the dropdown selection list. This is useful if you would like to provide to options something other than what you would like displayed. For example, consider:

```python from comet_ml import API, ui

api = API()

api_experiments = api.get_panel_experiments()

api_experiment = ui.dropdown( "Choose an experiment by key:", api_experiments, format_func= lambda experiment: experiment.key ) ```

In this example, the experiment's key is used in the dropdown list, but options is a list of APIExperiments. In this manner, you can pass in options in any format, but you should then provide a format_func for easy selection.

Arguments:

  • label: (str) label for the dropdown list
  • options: (list) list of choices to choose from
  • index: (int or list, optional) use initial index for single choice, or use a list of strings if multiple choice
  • format_func: (function, optional) function that takes an option and returns a string to use in the option list
  • key: (str, optional) when generating dropdowns in a loop - this is useful to assign unique keys for the dropdown
  • on_change: (function, optional) function to run when an option is selected
  • args: (list or tuple, optional) positional args to send to on_change function
  • kwargs: (dict, optional) keyword args to send to on_change function
  • multiple: (bool, optional) if True, allow user to select multiple options
  • classes: (list, optional) list of CSS class names to attach to the dropdown

ui.input()

The ui.input() widget is used in order to get textual input from the user. Pressing TAB or ENTER will trigger the script to run again with the new input value.

Signature and Use:

python value = ui.input(label, value="", key=None, on_click=None, args=None, kwargs=None, classes=None)

Arguments:

  • label: (str) textual description that preceeds the input area
  • value: (str, optional) default text in input area
  • key: (str, optional) when generating in a loop - this is useful to assign unique keys for the checkbox widget
  • on_click: (function, optional) function to run when text is ready
  • args: (list or tuple, optional) positional args to send to on_change function
  • kwargs: (dict, optional) keyword args to send to on_click function
  • classes: (list, optional) list of CSS class names to attach to the input widget

ui.checkbox()

The ui.checkbox() widget is used in order to get a binary choice from the user.

Signature and Use:

python value = ui.checkbox(label, value=False, key=None, on_click=None, args=None, kwargs=None, classes=None)

Arguments:

  • label: (str) textual description that follows the checkbox widget
  • value: (bool, optional) default value of checkbox
  • key: (str, optional) when generating in a loop - this is useful to assign unique keys for the checkbox widget
  • on_click: (function, optional) function to run when text is ready
  • args: (list or tuple, optional) positional args to send to on_change function
  • kwargs: (dict, optional) keyword args to send to on_click function
  • classes: (list, optional) list of CSS class names to attach to the checkbox widget

ui.button()

The ui.button() widget is used to trigger an action at a specific time.

Signature and Use:

python if ui.button(label, key=None, on_click=None, args=None, kwargs=None, classes=None): # do action else: # wait for button to be pressed Arguments:

  • label: (str) textual description that appears on the button
  • key: (str, optional) when generating in a loop, this is useful to assign unique keys for the button
  • on_click: (function, optional) function to run when text is ready
  • args: (list or tuple, optional) positional args to send to on_change function
  • kwargs: (dict, optional) keyword args to send to on_click function
  • classes: (list, optional) list of CSS class names to attach to the button

ui.progress()

The ui.progress() widget is used in order to show progress for long-running processes.

Use:

```python from comet_ml import ui import time

def long_running_process(): start_time = time.time() while time.time() - start_time < 3: pass

if ui.progress("Getting started", 0): # initialize panel, don't do anything yet! pass elif ui.progress("Processing 25% done...", 25): long_running_process() # first 25% elif ui.progress("Half-way done", 50): long_running_process() # second 25% elif ui.progress("Almost done!", 75): long_running_process() # running 3rd 25% elif ui.progress("Done!", 100): # clears screen long_running_process() # last 25% else: ui.display("Ok! What's next?") ```

Arguments:

  • label: (str) label for the dropdown list
  • percent: (int) number between 1 and 100
  • on_load: (function, optional) function to run
  • args: (list or tuple, optional) positional args to send to on_load function
  • kwargs: (dict, optional) keyword args to send to on_load function
  • classes: (list, optional) list of CSS class names to attach to the progress widget

ui.columns()

The ui.columns() widget is used to break the current display area into a series of columns. Columns can be nested.

Signature and Use:

python columns = ui.columns(items, classes=None, **styles)

By default, any item on which you perform ui.display() goes to the top-level Panel area. However, if you would like, you can place displayed items into different columns using this widget.

Examples:

```python columns = ui.columns(3)

for i, column in enumerate(columns): column.display("Column %s" % i)

columns = ui.columns([1, 3, 1])

middle column is thre times the width of first and third

columns[0].display("Column one") columns[1].display("Column two is wide") columns[2].display("Column three") ```

Arguments:

  • items: (int, or list of numbers) if an integer, then break the current area into even columns; if a list of numbers, then divide proportionally by the number's value divided by sum of numbers.
  • classes: (list, optional) list of CSS class names to attach to the columns
  • styles: (optional) dictionary of CSS items to apply to the columns

ui Utility methods

These are the ui utility functions:

  • ui.get_theme_names()
  • ui.set_css(css_text)
  • ui.add_css(css_text)

These are described below.

ui.get_theme_names()

Get the names of color themes for displaying pandas' Dataframes.

Example:

```python from comet_ml import ui

color_theme_names = ui.get_theme_names() color_theme_name = ui.dropdown.dropdown( "Select a theme:", color_theme_name )

create a draframe here

ui.display(df, theme=color_theme_name) ```

As shown, the color_theme_names could be useful in creating a dropdown selection showing a Dataframe.

ui.set_css()

This method allows you to set additional CSS for items for display.

Warning

This is experimental and may be removed in a future version.

ui.add_css()

This method allows you to add CSS for items for display.

Warning

This is experimental and may be removed in a future version.

Troubleshooting

If a panel is reporting an error, then it could be caused by the following:

  1. Using an unsupported browser. Currently Python Panels is known to work under the following browsers:

    • Chrome, version 71.0 and greater
    • Firefox, version 70.0 and greater
    • Edge, version 80 and greater
  2. The panel has an error. If this is your panel, you can edit it to fix what is wrong. If it is another user's panel, you can contact them to let them know there is an issue.

Debugging

Note that using print() will display items in the Console area (bottom, right-hand corner). This is very useful for debugging.

Technical Details

Python Panels also allows you to use many other Python support modules, including: asciitree, atomicwrites, attrs, beautifulsoup4, bleach, cloudpickle, cssselect, cycler, cytoolz, decorator, distlib, docutils, freesasa, future, glpk, html5lib, Jinja2, imageio, iniconfig, jedi, joblib, kiwisolver, libiconv, libxml, libxslt, libyaml, lxml, markdown, MarkupSafe, micropip, mne, more-itertools, mpmath, msgpack, networkx, nlopt, nose, numcodecs, optlang, packaging, parso, patsy, pluggy, py, Pygments, pyparsing, pyrtl, pytest, python-dateutil, python-sat, pytz, pywavelets, pyyaml, regex, retrying, setuptools, six, soupsieve, statsmodels, swiglpk, sympy, toolz, traits, typing-extensions, uncertainties, webencodings, xlrd, yt, zarr, and zlib.

Python Panels run in the browser rather than from a server. That has a number of implications. First and foremost, there is a bit of an overhead to load the libraries needed in order to execute Python code in the browser. Typically, this delay only happens when required, typcially once per project or experiment view. Note also, that Python Panels can use more memory than usual.

To create responsive Python Panels, it is suggested to:

  • limit the number of modules loaded
  • limit the number of requests (e.g., api.get_...() calls)
  • limit the amount of information displayed

The following Python modules are known to require noticeable load times:

  • matplotlib (medium load time)
  • pandas (large load time)
  • Plotly Express (large load times, as it requires pandas)

Known Limitiations

  • Python Panels do not automatically update with new experiments nor 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.
  • Limited types of widgets: currently, there is only ui.dropdown(); what else would you like?
  • Line numbers in Python's tracebacks are off; see your webbrowser's Console for current line numbers
  • Pandas' DataFrame.style is not available

References for Panel Resources

For more information on Panels, see also: