Exploring best solution when generating custom plots: export `get_data()` or similar from campaign to manually plot, modify, or rearrange data after camapaign execution

[acasadevall]

Overall I see class campaign does not have an easy way to export generated data once campaign has finalized the running part.

Use case:
I want to plot a custom graph which requires rearranging my data, adding other columns, etc., and then adding specific plotting features.

Issue:
Current campaign has campaign.generate_graph(...) which offers a straightforward solution to generate graphs based on x,y,hue params (seaborn/pandas style). This might be enough but for other more customized graphs requires adding pre/post callbacks. Example: composition of graphs, having FacetGrid vs non-FacetGrid.

Possible solutions

• (manually) Adding something like campaign.get_data() to get raw generated data in DataFrame (pandas) form. Example:

    output_data, gen_path = campaign.get_data() # <-- here we can also add data_frame callback similarly to current generate_graph approach
    # adding custom plot
    g = sns.catplot(data=processed_output, kind='bar', x='..', y='..', hue='..', palette='..', ...)
    g.fig.get_axes()[0].set_title("Title")
    g.set(ylabel="...", xlabel="...")
    g.fig.get_axes()[0].set_yscale('log')
    # saving using output path generated by benchkit/campaign
    g.fig.savefig(f"{fig_path}.png", transparent=False)
    print(f'[INFO] Saving campaign figure in "{gen_path }.png"')
    g.fig.savefig(f"{fig_path}.pdf", transparent=False)
    print(f'[INFO] Saving campaign figure in "{gen_path }.pdf"')

-- PROS: add post-process in the campaign
-- CONS: mix of responsibilities. current campaign class already has dependencies with Seaborn/Pandas when generating graph. Maybe campaign.get_data() should only return csv data rather than Pandas.

• (add more complexity into campaign.generate_graph) Adding more callbacks (pre/post) to add specific calls to the pipeline:

    campaign.generate_graph(
        plot_name="catplot",
        kind="bar",
        orient='v',
        x="...",
        y="...",
        hue="...",
        palette="...",
        ...,
        process_dataframe=df_callback,
        **graph_callback=post_graph_callback**
    )

-- PROS: already used in benchkit, no more methods are needed
-- CONS: adding more callbacks means adding more complexity. We cannot generate wrappers of wrappers to support custom plots. Generating graphs using campaign.generate_graph should not have more complexity than using standard Seaborn/Matplotlib way.

• (out of benchkit) Do a post-process afterwards on the csv/json files that are generated. This seems to be fair solution, but someone could it would be good to have only one pipeline from benchkit already.

[apaolillo]

I think returning the pandas DataFrame is a reasonable request. I can add that and the rest remains valid.