Further improve `HParam` performance by 18x

This uses traceback.extract_stack(limit=2)[-2] instead of inspect.stack(0)[1] because it's faster. This is the benchmark:

import inspect
import sys
import time
import timeit
import traceback

def deeper(func, depth):
    if depth > 0:
        return deeper(func, depth - 1)
    else:
        return func()

def no_stack():
    pass

def bench_stack():
    return inspect.stack()[1].filename, inspect.stack()[1].lineno

def bench_stack_limit():
    return inspect.stack(0)[1].filename, inspect.stack(0)[1].lineno

def bench_trace():
    return traceback.extract_stack()[-2].filename, traceback.extract_stack()[-2].lineno

def bench_trace_limit():
    return (traceback.extract_stack(limit=2)[-2].filename,
            traceback.extract_stack(limit=2)[-2].lineno)

assert deeper(bench_trace_limit, depth=5) == deeper(bench_stack, depth=5)
assert deeper(bench_trace, depth=5) == deeper(bench_stack, depth=5)
assert deeper(bench_stack_limit, depth=5) == deeper(bench_stack, depth=5)

samples = 10000
print('baseline:', timeit.timeit(lambda: deeper(no_stack, depth=5), number=samples))
print('bench_stack:', timeit.timeit(lambda: deeper(bench_stack, depth=5), number=samples))
print('bench_stack_limit:',
      timeit.timeit(lambda: deeper(bench_stack_limit, depth=5), number=samples))
print('bench_trace:', timeit.timeit(lambda: deeper(bench_trace, depth=5), number=samples))
print('bench_trace_limit:',
      timeit.timeit(lambda: deeper(bench_trace_limit, depth=5), number=samples))

The benchmark was inspired by: https://gist.github.com/JettJones/c236494013f22723c1822126df944b12

These were the results:

baseline: 0.008920529999999996
bench_stack: 17.366878454000002
bench_stack_limit: 4.497257854999997
bench_trace: 0.9375153429999976
bench_trace_limit: 0.2445523019999989

PetrochukM / HParams

Further improve `HParam` performance by 18x #4

Codecov Report