Skip to content

Latest commit

 

History

History
315 lines (228 loc) · 13.7 KB

performance.md

File metadata and controls

315 lines (228 loc) · 13.7 KB
jupyter
jupytext kernelspec language_info plotly
notebook_metadata_filter text_representation
all
extension format_name format_version jupytext_version
.md
markdown
1.3
1.16.4
display_name language name
Python 3 (ipykernel)
python
python3
codemirror_mode file_extension mimetype name nbconvert_exporter pygments_lexer version
name version
ipython
3
.py
text/x-python
python
python
ipython3
3.11.10
description display_as language layout name order permalink redirect_from thumbnail
Recommendations for increased speed, improved interactivity, and the ability to plot even more data!
basic
python
base
High Performance Visualization
14
python/performance/
python/webgl-vs-svg/
python/datashader/
thumbnail/webgl.jpg

DataFrame Types

New in Plotly.py version 6

Plotly Express natively supports various dataframe libraries, including pandas, Polars, and PyArrow. When building figures with Plotly Express, changing your dataframe library may help improve performance.

In versions of Plotly.py prior to version 6, Plotly Express functions accepted non-pandas dataframes as input but used the dataframe interchange protocol or converted those dataframes to pandas internally.

See the Plotly Express Arguments page for full details on supported dataframe libraries.

NumPy and NumPy Convertible Arrays for Improved Performance

New in Plotly.py version 6

You can improve the performance of generating Plotly figures that use a large number of data points by passing data as NumPy arrays, or in a format that Plotly can convert easily to NumPy arrays, such as pandas and Polars Series or DataFrames. These formats will usually show better performance than passing data as a Python list.

Plotly.py uses Plotly.js for rendering, which supports typed arrays. In Plotly.py, NumPy arrays and NumPy-convertible arrays are base64 encoded before being passed to Plotly.js for rendering.

Arrays and Data Types Supported

The following types of objects in Python are supported for base64 encoding for rendering with Plotly.js.

  • NumPy numpy.ndarray objects.
  • pandas Index, pandas Series, Polars Series, and PyArrow Chunked Array objects.
  • When working with Plotly Express, pandas DataFrame, Polars DataFrame and PyArrow DataFrame objects passed to the data_frame argument of px functions.
  • Array objects that can be converted to numpy.ndarray objects, i.e., they implement "__array__" or "__array_interface__" and return a numpy.ndarray.

The following array data types are supported:

  • float32
  • float64
  • int8
  • uint8
  • int16
  • uint16
  • int32
  • uint32

*If the array dtype is int64 or uint64, often the default dtype for arrays in NumPy when no dtype is specified, those dtypes will be changed to supported types internally by Plotly.py where possible. When working with NumPy directly, you can also specify the dtype when creating ndarray objects, and Plotly.py won't need to make the conversion internally.

Arrays or data types that are not supported for base64 encoding to Plotly.js's typed arrays specification will still work and render correctly with Plotly. Those arrays and or data types just won't have the performance benefits that Plotly.js's base64 typed arrays feature provides.

Dash Design Kit Compatibility

The Dash Design Kit Graph component from versions of Dash Design Kit earlier than 1.13 doesn't work with Plotly.py version 6 when using Python objects listed in the previous "Arrays and Data Types Supported". Upgrade Dash Design Kit to 1.13 or later to resolve this issue.

Unsupported Attributes

Arrays passed to attributes with the following names are not supported for base64 encoding for rendering with Plotly.js.

geojson, layers, and range.

Attributes that are not supported for base64 encoding to Plotly.js's typed arrays specification will still work and render correctly. Those attributes just won't have the performance benefits that Plotly.js's base64 typed arrays feature provides.

Example with NumPy Arrays

Here, we use NumPy arrays with a go.Scatter3d figure.

import plotly.graph_objects as go
import numpy as np

np.random.seed(1)

# Number of data points
N = 10000

# Generate random data
x = np.random.randn(N)
y = np.random.randn(N).astype('float32')
z = np.random.randint(size=N, low=0, high=256, dtype='uint8')
c = np.random.randint(size=N, low=-10, high=10, dtype='int8')

fig = go.Figure(data=[go.Scatter3d(
    x=x,
    y=y,
    z=z,
    marker=dict(color=c),
    mode='markers',
    opacity=0.2
)])

fig.show()

WebGL

plotly figures are rendered by web browsers, which broadly speaking have two families of capabilities for rendering graphics:

  • The SVG API, which supports vector rendering.
  • The Canvas API, which supports raster rendering, and can exploit GPU hardware acceleration via a browser technology known as WebGL.

Each plotly trace type is rendered with either SVG or WebGL. The following trace types use WebGL for rendering:

  • Accelerated versions of SVG trace types: scattergl, scatterpolargl,
  • High-performance multidimensional trace types: splom, or parcoords
  • 3D trace types scatter3d, surface, mesh3d, cone, streamtube, isosurface, volume
  • Mapbox Gl JS-powered trace types: scattermap, choroplethmap, densitymap

WebGL Limitations and Tradeoffs

WebGL is a powerful technology for accelerating rendering but comes with some strict limitations:

  1. GPU requirement: WebGL is a GPU (graphics card) technology and therefore requires specific hardware which is available in most but not all cases and is supported by most but not all browsers.
  2. Rasterization: WebGL-rendered data is drawn as a grid of pixels rather than as individual shapes, so can appear pixelated or fuzz in certain cases, and when exported to static file formats will appear pixelated on zoom. In addition, text rendering will differ between SVG and WebGL-powered traces.
  3. Context limits: browsers impose a strict limit on the number of WebGL "contexts" that any given web document can access. WebGL-powered traces in plotly can use multiple contexts in some cases but as a general rule, it may not be possible to render more than 8 WebGL-involving figures on the same page at the same time. See the following section, Multiple WebGL Contexts, for more details.
  4. Size limits: browsers impose hardware-dependent limits on the height and width of figures using WebGL which users may encounter with extremely large plots (e.g. tens of thousands of pixels of height).

In addition to the above limitations, the WebGL-powered version of certain SVG-powered trace types (scattergl, scatterpolargl) are not complete drop-in replacements for their SVG counterparts yet

  • Available symbols will differ.
  • Area fills are not yet supported in WebGL.
  • Range breaks on time-series axes are not yet supported.
  • Axis range heuristics may differ.

Multiple WebGL Contexts

New in 5.19

Most browsers have a limit of between 8 and 16 WebGL contexts per page. A Plotly WebGL-based figure may use multiple WebGL contexts, but generally you'll be able to render between 4 and 8 figures on one page.

If you exceed the browser limit on WebGL contexts, some figures won't render and you'll see an error. In the console in Chrome, for example, you'll see the error: "Too many active WebGL contexts. Oldest context will be lost".

If you encounter WebGL context limits when using WebGL-based figures, you can use Virtual WebGL, which virtualizes a single WebGL context into multiple contexts.

To use it, in the environment where your Plotly figures are being rendered, load the Virtual WebGL script, "https://unpkg.com/[email protected]/src/virtual-webgl.js", for example, using a <script> tag. Performance when using Virtual WebGL will be slower than when not using Virtual WebGL.

In a Jupyter notebook environment that supports magic commands, you can load it with the HTML magic command:

%%html
<script src=“https://unpkg.com/[email protected]/src/virtual-webgl.js”></script>

WebGL for Scatter Performance

In the examples below we show that it is possible to represent up to around a million points with WebGL-enabled traces. For larger datasets, or for a clearer visualization of the density of points, it is also possible to use datashader.

WebGL with Plotly Express

The render_mode argument to supported Plotly Express functions (e.g. scatter and scatter_polar) can be used to enable WebGL rendering.

Note The default render_mode is "auto", in which case Plotly Express will automatically set render_mode="webgl" if the input data is more than 1,000 rows long. In this case, WebGl can be disabled by setting render_mode=svg.

Here is an example that creates a 100,000 point scatter plot using Plotly Express with WebGL rendering explicitly enabled.

import plotly.express as px

import pandas as pd
import numpy as np

np.random.seed(1)

N = 100000

df = pd.DataFrame(dict(x=np.random.randn(N),
                       y=np.random.randn(N)))

fig = px.scatter(df, x="x", y="y", render_mode='webgl')

fig.update_traces(marker_line=dict(width=1, color='DarkSlateGray'))

fig.show()

WebGL with 1,000,000 points with Graph Objects

If Plotly Express does not provide a good starting point for creating a chart, you can use the more generic go.Scattergl class from plotly.graph_objects.

import plotly.graph_objects as go

import numpy as np

N = 1_000_000

fig = go.Figure()

fig.add_trace(
    go.Scattergl(
        x = np.random.randn(N),
        y = np.random.randn(N),
        mode = 'markers',
        marker = dict(
            line = dict(
                width = 1,
                color = 'DarkSlateGrey')
        )
    )
)

fig.show()

See https://plotly.com/python/reference/scattergl/ for more information and chart attribute options!

Datashader

Use Datashader to reduce the size of a dataset passed to the browser for rendering by creating a rasterized representation of the dataset. This makes it ideal for working with datasets of tens to hundreds of millions of points.

Passing Datashader Rasters as a Tile Map Image Layer

The following example shows the spatial distribution of taxi rides in New York City, which are concentrated on major avenues. For more details about tile-based maps, see the tile map layers tutorial.

import pandas as pd
df = pd.read_csv('https://raw.githubusercontent.com/plotly/datasets/master/uber-rides-data1.csv')
dff = df.query('Lat < 40.82').query('Lat > 40.70').query('Lon > -74.02').query('Lon < -73.91')

import datashader as ds
cvs = ds.Canvas(plot_width=1000, plot_height=1000)
agg = cvs.points(dff, x='Lon', y='Lat')
# agg is an xarray object, see http://xarray.pydata.org/en/stable/ for more details
coords_lat, coords_lon = agg.coords['Lat'].values, agg.coords['Lon'].values
# Corners of the image
coordinates = [[coords_lon[0], coords_lat[0]],
               [coords_lon[-1], coords_lat[0]],
               [coords_lon[-1], coords_lat[-1]],
               [coords_lon[0], coords_lat[-1]]]

from colorcet import fire
import datashader.transfer_functions as tf
img = tf.shade(agg, cmap=fire)[::-1].to_pil()

import plotly.express as px
# Trick to create rapidly a figure with map axes
fig = px.scatter_map(dff[:1], lat='Lat', lon='Lon', zoom=12)
# Add the datashader image as a tile map layer image
fig.update_layout(
    map_style="carto-darkmatter",
    map_layers=[{"sourcetype": "image", "source": img, "coordinates": coordinates}],
)
fig.show()

Exploring Correlations of a Large Dataset

Here we explore the flight delay dataset from https://www.kaggle.com/usdot/flight-delays. In order to get a visual impression of the correlation between features, we generate a datashader rasterized array which we plot using a Heatmap trace. It creates a much clearer visualization than a scatter plot of (even a fraction of) the data points, as shown below.

import plotly.graph_objects as go
import pandas as pd
import numpy as np
import datashader as ds
df = pd.read_parquet('https://raw.githubusercontent.com/plotly/datasets/master/2015_flights.parquet')
fig = go.Figure(go.Scattergl(x=df['SCHEDULED_DEPARTURE'][::200],
                             y=df['DEPARTURE_DELAY'][::200],
                             mode='markers')
)
fig.update_layout(title_text='A busy plot')
fig.show()
import plotly.express as px
import pandas as pd
import numpy as np
import datashader as ds
df = pd.read_parquet('https://raw.githubusercontent.com/plotly/datasets/master/2015_flights.parquet')

cvs = ds.Canvas(plot_width=100, plot_height=100)
agg = cvs.points(df, 'SCHEDULED_DEPARTURE', 'DEPARTURE_DELAY')
zero_mask = agg.values == 0
agg.values = np.log10(agg.values, where=np.logical_not(zero_mask))
agg.values[zero_mask] = np.nan
fig = px.imshow(agg, origin='lower', labels={'color':'Log10(count)'})
fig.update_traces(hoverongaps=False)
fig.update_layout(coloraxis_colorbar=dict(title='Count', tickprefix='1.e'))
fig.show()

Instead of using Datashader, it would theoretically be possible to create a 2d histogram with Plotly, but this is not recommended because you would need to load the whole dataset of around 5M rows in the browser for plotly.js to compute the heatmap.