numexpr VS line_profiler

You can just slap numexpr on top of it to compile this line on the fly.

https://github.com/pydata/numexpr

Are you doing any costly chained NumPy operations in your preprocessing? E.g. max(abs(large_ary)), this produces multiple copies of your data, https://github.com/pydata/numexpr can greatly reduce time spent with such operations

What is not entirely obvious here is that under the hood you can install a nice library called numexpr (docs, src) that exists to make calculations with large NumPy (and pandas) objects potentially much faster. When you use query or eval, this expression is passed into numexpr and optimized using its bag of tricks. Expected performance improvement can be between .95x and up to 20x, with average performance around 3-4x for typical use cases. You can read details in the docs, but essentially numexpr takes vectorized operations and makes them work in chunks that optimize for cache and CPU branch prediction. If your arrays are really large, your cache will not be hit as often. If you break your large arrays into very small pieces, your CPU won’t be as efficient.

LineProfiler is the best tool to learn how to write performant Python and code optimization.

https://github.com/pyutils/line_profiler

You can literally see the hot spot of your code, then you can grind different algorithms or change the whole architecture to make it faster.

For example replace short for loops to list comprehensions, vectorize all numpy operations (only vectorize partially do not help the issue), using 'not any()' instead or 'all()' for boolean, etc.

Doing this for like 2 weeks, basically you can automatically recognized most bad code patterns in a glance.

You may want to try using a package like line_profiler to narrow down where the time is spent.

When tuning for performance, always measure. Never assume you know where the slow parts are. Run a line profiler and see where all the time is actually going.

Either way I would highly recommend running a profiler on your code to see where the program is spending most of its time. line_profiler is a very nice one, as it shows you execution time for each line.

There are dedicated tools like line_profiler for python - if this doesn't do exactly what you need it can be easily modified.

It's surprising to me that an sklearn pipeline and a spacy pipeline both doing simple regexing are vastly different in performance. I would go one layer deeper with measurement with something like line_profiler, which I've used to great effect to get line-by-line perf stats. This should illuminate why.

This looks really nice! Does it use line_profiler or is it a different implementation for the profiling? Either way the interface is fantastic!

# https://github.com/pyutils/line_profiler pip install line_profiler kernprof -l -v some-code.py # This might take a while... Wrote profile results to some-code.py.lprof Timer unit: 1e-06 s Total time: 13.0418 s File: some-code.py Function: exp at line 3 Line # Hits Time Per Hit % Time Line Contents ============================================================== 3 @profile 4 def exp(x): 5 1 4.0 4.0 0.0 getcontext().prec += 2 6 1 0.0 0.0 0.0 i, lasts, s, fact, num = 0, 0, 1, 1, 1 7 5818 4017.0 0.7 0.0 while s != lasts: 8 5817 1569.0 0.3 0.0 lasts = s 9 5817 1837.0 0.3 0.0 i += 1 10 5817 6902.0 1.2 0.1 fact *= i 11 5817 2604.0 0.4 0.0 num *= x 12 5817 13024902.0 2239.1 99.9 s += num / fact 13 1 5.0 5.0 0.0 getcontext().prec -= 2 14 1 2.0 2.0 0.0 return +s