[magnio]

The problem with matplotlib has never been its capability. I consider its rendering backend and resulting image quality to be top-notch, and there are a billion options for you to customize.

That second part is the biggest problem with matplotlib: the imperative API cloned from Matlab is just terrible. The vast majority of matplotlib examples are rife with global states and overriding behavior (ever have to set font size twice to force it?) At first I thought that it's a bit clunky due to my inexperience, but no, the API is just that bad. See https://ryxcommar.com/2020/04/11/why-you-hate-matplotlib/

Now, I mean no hate to matplotlib, which is still a solid library. I appreciate their commitment to backward compatibility by keeping all those baggage, and I don't want them to yank them out, given that seaborn already mitigates most of the API problems, and Altair already brings enlightenment to Python plotting.

Now if only seaborn works with polars natively ...

[nerdponx]

The object-oriented API and its documentation have improved markedly over the years, I encourage you to give it a try. Most of the official documentation examples use it now. Yes, there are lots of examples out there using the pyplot interface still, but there is enough information on the OO API now that you can get by.

Personally I think pyplot is still a perfectly fine interface for just banging out quick data visualization without too much effort, but I find myself very quickly switching to the OO interface as soon as I want a little bit of control over what's going on.

There are still some lingering issues, like the interface discrepancy between scatter and other plotting methods, and the difficulty of making a nicely-formatted color bar. But it's really coming a long way since I started using it in ~2015, And in that time I've actually come to prefer it over both base R and ggplot2, which is what I started with originally.

I have much stronger complaints about Seaborn and he will never find me recommending it to anyone.

[magnio]

> The object-oriented API and its documentation have improved markedly over the years, I encourage you to give it a try. Most of the official documentation examples use it now. Yes, there are lots of examples out there using the pyplot interface still, but there is enough information on the OO API now that you can get by.

I agree. I used seaborn and matplotlib recently, and I was pleasantly surprised at both how much more consistent the API is and how much more comprehensive the documentation is. I dreaded using it at first but quickly became more accustomed.

> I have much stronger complaints about Seaborn and he will never find me recommending it to anyone.

Interesting, what are your pain points? For me, most of the time I start with seaborn, then modify the underlying matplotlib figure and axes for more customization; only rarely do I start with plt.subplots(). This gives me nice default styles while still allowing me to bend it to my will.

[nerdponx]

> start with seaborn, then modify the underlying matplotlib figure

Maybe this is the heart of my complaint: last I remember trying Seaborn, I found it hard to get access to the underlying Axes objects, and I also found it hard/impossible to add elements incrementally to an existing Axes. This might be a PEBCAK issue so I'm willing to try it again.

[magnio]

Understandable. A reason why I keep going back and forth between matplotlib and seaborn in the past is due to not knowing how to customize seaborn. Now that I know they are just regular matplotlib objects, it's very easy to mix and match them when seaborn is just too opinionated. You can read more here https://seaborn.pydata.org/tutorial/introduction.html#opinio...

Some examples I did recently: to set the layout engine of a seaborn plot

    g = sns.relplot(df, x="value", y="y", col="variable", col_wrap=3)
    g.figure.set_layout_engine("constrained")
    g.figure.set_size_inches(16, 12)

Set the subplot titles

    g = sns.displot(df, x="value", col="model")
    g.figure.set_size_inches(16, 8)
    for (ax, title) in zip(g.axes[0], titles):
        ax.set_title(f"Title: {title}")

Multiple seaborn plots with two lines, error band, and scatter placed onto existing matplotlib subplots and custom colors

  colors = sns.color_palette()
  fig, axs = plt.subplots(1, len(ys), sharey=True, figsize=(16, 6), constrained_layout=True)
  for (X, y, ax, title) in zip(Xs, ys, axs, titles):
    sns.lineplot(x=X, y=y, color=colors[0], ax=ax)
    sns.lineplot(x=X, y=y_predict, color=colors[1], ax=ax)
    ax.fill_between(X, y_predict - std, y_predict + std, color=colors[1], alpha=0.3)
    
    sns.scatterplot(x=X_train, y=y_train, alpha=0.7, color=colors[7], ax=ax)
    
    ax.set_title(title)
    ax.tick_params(rotation=0)

(You can also prepare the data in pandas the way seaborn like it and create faceted plot declaratively; the point is you can mix and match them freely and intuitively.)