Poor man runtime callgraph. You need linux perf (perf-tools under debian/ubuntu) and valgrind for the corresponding functions

minilib.py

import contextlib
import os
import signal
import subprocess

import numpy as np

def enable_callgrind():
    subprocess.check_call(["callgrind_control", "--instr=on", str(os.getpid())])

def disable_callgrind():
    subprocess.check_call(["callgrind_control", "--instr=off", str(os.getpid())])

def reset_callgrind():
    subprocess.check_call(["callgrind_control", "-z", str(os.getpid())])

@contextlib.contextmanager
def under_callgrind(reset=True):
    if reset:
        reset_callgrind()

    enable_callgrind()
    yield
    disable_callgrind()

@contextlib.contextmanager
def under_perf():
    pid = str(os.getpid())
    p = subprocess.Popen(["perf", "record", "-g", "-a", "-F", "1000", "-p", pid])
    yield
    os.kill(p.pid, signal.SIGINT)