Plotnine performance issue

Firstly, thanks for creating this project - I’ve been looking for a python grammar of graphics implementation for a while.

I’m having a performance issue plotting a large Pandas data frame.

The data is found in prob_w_d_df, which is a (600000, 3) data frame; prob_w_d_df.describe() gives:

	doc_id		prob
	------		----
count	600000		6.000000e+05
mean	29		1.000000e-04
std	17.318117	4.862052e-04
min	0.000000	2.285988e-07
25%	14.750000	1.205222e-06
50%	29.500000	5.755733e-06
75%	44.250000	3.625872e-05
max	59.000000	9.012307e-02

The plot code is:

ggplot(prob_w_d_df, aes(x='feature', y='prob', group=1)) + 
 geom_line() + 
 xlab("Word") + 
 ylab("P(word)") +
 facet_wrap('~doc_id', scales="free") +
 theme_minimal(base_family="Arial") +
 theme(figure_size=(15,15), axis_text_x=element_blank())
) 

This takes 1,543 seconds (~27 minutes) to generate the plot.

Now I understand it’s a large data frame so will be slow; however, when I replicate the exact above code with ggplot2 in R it takes 65 seconds. I understand R may be faster but the difference seems to great.

Is there anything I’m doing wrong or should do you improve performance?

(I’ve shared a CSV dump of prob_w_d_df if it helps: prob_w_d_df.csv.)

Thanks.

[bevankoopman]

If I remove the facet_wrap(...) (i.e., plot a single panel) it’s still takes 20 minutes so it seems multiple axes are not the real issue.

[grgmiller]

I was having a similar issue with performance when plotting data from a Pandas dataframe. What I found was that plotnine was having trouble if one of my variables was an object datatype. My x-axis was a timeseries of minutes of the day, which I had converted to strings, like “12:05”, “12:10”, etc.

I found that when I converted these timestamps to float datatypes, like 12.0833, 12.1667, etc, it improved the performance considerably.