Yes, it is what I required, but due to the matrix size, (7,7), and its randomly generated elements, it is difficult to compare how 3d Julia and Python matrices are serialized.
Instead I took the simplest case
z1=z2=z3 = [1 2 3; 4 5 6]
respectively
z1=z2=z3=np.array([[1, 2, 3], [4, 5, 6]])
and ran both PlotlyJS https://gist.github.com/empet/c41a8b131687cc2584485ab47ca6ddf4, and plotly.py https://gist.github.com/empet/ba5a8d0678cd75cd2119f3b74077425c versions of the code generating a single heatmap, with the same customdata. Then I displayed the json serialization of each figure. Based on the notebooks with my experiments, you should compare further your json version with the two listed in notebooks, and decide how customdata must be defined when you are using Pycall.
Moreover, take into account that the PlotlyJS’s author added automatically in the json file the transpose, z’, of the matrix that defines a heatmap. That’s why I took now an example of matrix of size (2,3) to notice when there is a discordance between the size of z and the sizes o matrices involved in customdata.
It’s mind-blowing!!!