Profiling#

%scikitplot_profile runs a statement, %%scikitplot_profile a cell, under cProfile and shows the functions that took the time:

%scikitplot_profile -s ncalls -l 5 sorted(range(10_000), reverse=True)

%%scikitplot_profile -l 10
def fib(n):
    return n if n < 2 else fib(n - 1) + fib(n - 2)
fib(18)

Option

Meaning

-s KEY

sort by cumulative (default), tottime or ncalls

-l N

rows of the table (default 25)

-D FILE

also write the profile to FILE (pstats, snakeviz)

--snakeviz

open the profile in snakeviz (see below)

--stop-snakeviz

stop the snakeviz servers of this session

The result is a ProfileResult: .rows (ncalls, tottime, cumtime, location), .total_calls, .to_dict(). Code that raises is profiled up to the error; the table is shown, then the error.

Only one profiler can record a process: when another one is active (an outer profile, python -m cProfile, a profiling IDE) the magic refuses instead of silently stopping it.

snakeviz#

--snakeviz starts snakeviz on 127.0.0.1 (never on every network interface) and shows it in the notebook. That works when the browser runs on the kernel’s machine, as with a local Jupyter. On a remote kernel, keep the table, or write the profile with -D and open the file with snakeviz where you are. Servers stop with --stop-snakeviz, when the extension is unloaded (%unload_ext scikitplot) and when the kernel exits.