Advanced uses: notebooks

Remembering state between runs

While it’s possible to change the visualization settings in python using hiplot.Experiment.display_data() and hiplot.ValueDef, it is more convenient to define XY plots, change column types or coloring in the user interface directly. However, all of this is reset when executing the cell a second time.

If you set store_state_key to a unique identifier in hiplot.Experiment.display(), HiPlot’s state will be stored in the URL, and restored when hiplot.Experiment.display() is called a second time later with the same parameter.

# This visualization's state will be recovered when the cell is re-run
import hiplot as hip
data = [{'dropout':0.1, 'lr': 0.001, 'loss': 10.0, 'optimizer': 'SGD'},
        {'dropout':0.15, 'lr': 0.01, 'loss': 3.5, 'optimizer': 'Adam'},
        {'dropout':0.3, 'lr': 0.1, 'loss': 4.5, 'optimizer': 'Adam'}]
hip.Experiment.from_iterable(data).display(store_state_key="cell1")

Interactions with displayed visualization

Note

This feature is currently supported for Jupyter Notebooks only. In particular, it does not work with Jupyter Lab.

HiPlot visualizations are highly interactive, allowing the user to filter and slice through the data. When running in Jupyter Notebooks, it is possible to retrieve some information about the state of the visualization.

When calling hiplot.Experiment.display(), a hiplot.ExperimentDisplayed object is returned. Behind the hood, a Javascript-Python communication channel is established with the python notebook kernel, that HiPlot’s javascript frontend uses to send information about selected datapoints, filtering, …

# A Jupyter Notebook cell
import hiplot as hip
data = [{'dropout':0.1, 'lr': 0.001, 'loss': 10.0, 'optimizer': 'SGD'},
        {'dropout':0.15, 'lr': 0.01, 'loss': 3.5, 'optimizer': 'Adam'},
        {'dropout':0.3, 'lr': 0.1, 'loss': 4.5, 'optimizer': 'Adam'}]
displayed_exp = hip.Experiment.from_iterable(data).display()

After the visualization has loaded, and in another cell, python code can query information from HiPlot

>>> displayed_exp.get_selected()  # Return all the datapoints currently selected
 [<hiplot.experiment.Datapoint at 0x10b32dc10>]

>>> displayed_exp.get_brush_extents()  # Retrieve current brush extents in the parallel plot
 {'dropout': {'type': 'numeric',
  'brush_extents_normalized': [1, 0.6015169902912622],
  'range': [0.3, 0.22030339805825244]},
 'lr': {'type': 'numeric',
  'brush_extents_normalized': [1, 0.5918082524271845],
  'range': [0.1, 0.05958901699029127]},
 'optimizer': {'type': 'categorical',
  'brush_extents_normalized': [1, 0],
  'values': ['Adam', 'SGD']}}

Note

While all columns will have a type and a brush extent (brush_extents_normalized), only numeric columns have a range field. Categorical fields have a list of selected values instead

Warning

The user can modify the data type of variables - so the data type returned by hiplot.ExperimentDisplayed.get_brush_extents() does not always match the one provided in Python.