write_image (via kaleido) visually distorts large square heatmaps through aliasing artifacts

Offen
  • Stimmt etwas nicht?
#186 12 Kommentare 4 Reaktionen 1 zugewiesene Person Auf GitHub ansehen

@ayjayt arbeitet bereits daran.

Seit 03.6.2025.

Bewertung

Dieses Issue wurde noch nicht bewertet.

Beschreibung

bug P1 stale

I have a relatively large matrix that I want to plot usinggraph_objects.Heatmap. The matrix has some structure that should become immediately apparent to the viewer (i.e. it has a constant diagonal and is symmetric). However due to visual artifacts, it looks like these properties are violated in the image produced by write_image (with the kaleido backend, haven't tested orca).

It was difficult to write a simple example script to reproduce these distortions, as variations in both the matrix and output image size can affect the issue significantly (both in intensity and quality). In any case it looks to be an aliasing artifact related to the selection of heatmap cell data that results in a non-bijective map between the heatmap cells and regions within the (vector) output image.

In the script below, one output format is PDF to rule out image compression as a culprit (as there the heatmap appears to be stored as a vector image as its cells remain sharp when zoomed in). However, the issue can be reproduced also when the output format is PNG. I assume that in the vector-output case, the heatmap data is (badly) rescaled like a bitmap image whose pixels are then re-interpreted as an array of square cells, even though those cells then do not match the heatmap/data cells any more.

#!/usr/bin/env python3

import numpy as np
import plotly.graph_objects as go
import plotly.io as pio

resultions = [
    # Gives very unequal cell sizes, making the result look like a mosaic:
    (200, 400, 400),
    # Looks almost correct, but some diagonal elements show bad values:
    (400, 400, 400),
]

for n, width, height in resultions:
    # Produce a random matrix with key properties (symmetric, constant
    # diagonal) matching my actual data.
    np.random.seed(1)
    A = np.random.random(size=(n, n))
    A += A.T
    np.fill_diagonal(A, -1)

    # Plot the matrix as a square heatmap.
    fig = go.Figure()
    fig.add_trace(go.Heatmap(z=A))
    fig.update_layout(
        autosize=False,
        width=width,
        height=height,
        margin=dict(b=0, l=0, r=0, t=0),
        yaxis1=dict(scaleanchor="x1", constrain="domain"),
        yaxis2=dict(scaleanchor="x2", constrain="domain"),
    )
    for ext in ("pdf", "png"):
        pio.write_image(fig, f"n={n}_{width}x{height}.{ext}")

The outcomes are:

PDF:
n=200_400x400.pdf
n=400_400x400.pdf

PNG (note in the first image that only the heatmap, not the text looks blurred):
n=200_400x400
n=400_400x400

  1. For n = 200 and a layout size of 400x400, many adjacent rows and columns appear to use the same heatmap data. As a result the diagonal looks to be composed of alternating small and large blue cells. On closer inspection you can see that there are two locations where a double row and a double column meet off-diagonal, so that the result looks asymmetric.
  2. For n = 400 and a layout size of 400x400, the result looks mostly correct but some main diagonal cells are not blue, i.e. they have not been sampled from the data's main diagonal.

Given that my understanding of the problem is correct, I would suggest:

  1. For bitmap output, use a better interpolation method to scale the heatmap image to its output size.
  2. For vector output, if possible, do not go the middle route of rescaling a bitmap image. Instead rescale the vector data.
Vorherrschende Sprache
Python
Sterne
504
Forks
65
Ø Merge
6 T. 9 Std.
Gemergte PRs (30 T.)
2

Entwicklungsumgebung

Dieses Projekt bietet weder Dev-Container noch Dockerfile noch Beitragsleitfaden – die Einrichtung liegt bei Ihnen. Beginnen Sie mit der README; die allgemeinen Schritte stehen in unserem Leitfaden für den ersten Beitrag.

Erste Schritte

  1. Lesen Sie das ganze Issue und danach den Beitragsleitfaden des Projekts.
  2. Schreiben Sie ins Issue, dass Sie es übernehmen — das erspart doppelte Arbeit.
  3. Forken Sie das Repository und arbeiten Sie in einem Branch.
  4. Öffnen Sie einen Pull Request, der die Issue-Nummer nennt.

Mehr aus plotly/kaleido

Alle Issues in plotly/kaleido

Ähnliche Issues

Weitere Issues zu Python

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.