Data Science: From School to Work, Part V

Make it work, then make it stunning, then if you happen to actually, actually should, make it quick. 90 % of the time, if you happen to make it stunning, it’ll already be quick. So actually, simply make it stunning! (Source)

— Joe Armstrong (co-designers of the Erlang programming language.)

article about Python for the sequence “Knowledge Science: From Faculty to Work.” Because the starting, you might have discovered how one can manage your Python project with UV, how to write a clean code using PEP and SOLID principles, how to handle errors and use loguru to log your code and how to write tests.

Now you’re ready to create working, production-ready code. However code isn’t excellent and might all the time be improved. A remaining (non-obligatory, however extremely beneficial) step in creating code is optimization.

To optimize your code, you want to have the ability to monitor what’s happening in it. To take action, we use instruments referred to as Profilers. They generate profiles of your code. It means a set of statistics that describes how usually and for the way lengthy varied components of this system executed. They make it doable to determine bottlenecks and components of the code that eat too many sources. In different phrases, they present the place your code must be optimized.

At present, there’s such a proliferation of profilers in Python that the default profiler in Pycharm is known as yappi for “But One other Python Profiler”.

This text is subsequently not an exhaustive listing of all current profilers. On this article, I current a device for every facet of the code we need to profile: reminiscence, time and CPU/GPU consumption. Different packages will probably be talked about with some references however won’t be detailed.

---

I – Reminiscence profilers

Reminiscence profiling is the strategy of monitoring and evaluating a program’s reminiscence utilization whereas working. This technique helps builders to find reminiscence leaks, optimizing reminiscence utilization, and comprehending their packages’ reminiscence consumption patterns. Reminiscence profiling is essential to forestall purposes from utilizing extra reminiscence than obligatory and inflicting sluggish efficiency or crashes.

1/ memory-profiler

memory_profiler is an easy-to-use Python module designed to profile reminiscence utilization of a script. It depends upon psutil module. To put in the package deal, merely kind:

pip set up memory_profiler # (in your digital atmosphere)
# or if you happen to use uv (what I encourage)
uv add memory_profiler

Profiling executable

One of many benefits of this package deal is that it isn’t restricted to pythonic use. It installs the mprof command that enables monitoring the exercise of any executable.

For example, you possibly can monitor the reminiscence consummation of purposes like ollama by working this command:

mprof run ollama run gemma3:4b
# or with uv
mprof run ollama run gemma3:4b

To see the consequence, you must set up matplotlib first. Then, you possibly can plot the recorded reminiscence profile of your executable by working:

mprof plot
# or with uv
mprof run ollama run gemma3:4b

The graph then seems to be like this:

Profiling Python code

Let’s get again to what brings us right here, the monitoring of a Python code.

memory_profiler works with a line-by-line mode utilizing a easy decorator @profile. First, you adorn the curiosity perform and then you definitely run the script. The output will probably be written on to the terminal. Think about the next monitoring.py script:

@profile
def my_func():
    a = [1] * (10 ** 6)
    b = [2] * (2 * 10 ** 7)
    del b
    return a


if __name__ == '__main__':
    my_func()

You will need to discover that it isn’t essential to import the package deal from memory_profiler import profile on the start of the script. On this case you must specify some particular arguments to the Python interpreter.

python-m memory_profiler monitoring.py # with an area between python and -m
# or
uv run -m memory_profiler monitoring.py

And you’ve got the next output with a line-by-line particulars:

The output is a desk with 5 columns.

• Line #: The road variety of the profiled code
• Mem utilization: The reminiscence utilization of the Python interpreter after executing that line.
• Increment: The change in reminiscence utilization in comparison with the earlier line.
• Occurrences: The variety of instances that line was executed.
• Line Contents: The precise supply code.

This output could be very detailed and permits very fantastic monitoring of a particular perform.

Essential: Sadly, this package deal is now not actively maintained. The creator is on the lookout for a substitute.

2/ tracemalloc

tracemalloc is a built-in module in Python that tracks reminiscence allocations and deallocations. Tracemalloc offers an easy-to-use interface for capturing and analyzing reminiscence utilization snapshots, making it a useful device for any Python developer.

It gives the next particulars:

• Exhibits the place every object was allotted by offering a traceback.
• Offers reminiscence allocation statistics by file and line quantity, together with the general measurement, rely, and common measurement of reminiscence blocks.
• Lets you examine two snapshots to determine potential reminiscence leaks.

The package deal tracemalloc could also be usefull to determine reminiscence leak in your code.

Personally, I discover it much less intuitive to arrange than the opposite packages introduced on this article. Listed below are some hyperlinks to go additional:

---

II – Time profilers

Time profiling is the method of measuring the time spent in several components of a program. By figuring out efficiency bottlenecks, you possibly can focus their optimization efforts on the components of the code that can have probably the most important affect.

1/ line-profiler

The line-profiler package deal is kind of much like memory-profiler, however it serves a distinct objective. It’s designed to profile particular features by measuring the execution time of every line inside these features. To make use of LineProfiler successfully, it’s essential to explicitly specify which features you need it to profile by merely including the @profile decorator above them.

To put in it simply kind:

pip set up line_profiler # (in your digital atmosphere)
# or
uv add line_profiler

Contemplating the next script named monitoring.py

@profile
def create_list(lst_len: int):
    arr = []
    for i in vary(0, lst_len):
        arr.append(i)


def print_statement(idx: int):
    if idx == 0:
        print("Beginning array creation!")
    elif idx == 1:
        print("Array created efficiently!")
    else:
        elevate ValueError("Invalid index offered!")


@profile
def foremost():
    print_statement(0)
    create_list(400000)
    print_statement(1)


if __name__ == "__main__":
    foremost()

To measure the execution time of the perform foremost() and create_list(), we add the decorator @profile.

The best technique to get a time profiling of this script to make use of the kernprof script.

kernprof -lv monitoring.py # (in your digital atmosphere)
# or
uv run kernprof -lv monitoring.py

It would create a binary file named your_script.py.lprof. The argument -v permits to indicate directyl the output within the terminal.
In any other case, you possibly can view the outcomes later like so:

python-m line_profiler monitoring.py.lprof # (in your digital atmosphere)
# or
uv run python -m line_profiler monitoring.py.lprof

It offers the next informations:

There are two tables, one by profiled perform. Every desk containes the next informations

• Line #: The road quantity within the file.
• Hits: The variety of instances that line was executed.
• Time: The whole period of time spent executing the road within the timer’s models. Within the header data earlier than the tables, you will note a line “Timer unit:” giving the conversion issue to seconds. It might be completely different on completely different techniques.
• Per Hit: The typical period of time spent executing the road as soon as within the timer’s models
• % Time: The proportion of time spent on that line relative to the whole quantity of recorded time spent within the perform.
• Line Contents: The precise supply code.

1/ cProfile

Python comes with two built-in profilers:

• cProfile: A C extension with affordable overhead that makes it appropriate for profiling long-running packages. It is suggested for many customers.
• profile: A pure Python module whose interface is imitated by cProfile, however which provides important overhead to profiled packages. It may be a invaluable device when it’s essential to lengthen or customise the profiling performance.

The bottom syntax is cProfile.run(assertion, filename=None, kind=-1). The filename argument could be handed to avoid wasting the output. And the kind argument can be utilized to specify how the output must be printed. By default, it’s set to -1( no worth).

For example, if you happen to modify the monitoring script like this:

import cProfile


def create_list(lst_len: int):
    arr = []
    for i in vary(0, lst_len):
        arr.append(i)


def print_statement(idx: int):
    if idx == 0:
        print("Beginning array creation!")
    elif idx == 1:
        print("Array created efficiently!")
    else:
        elevate ValueError("Invalid index offered!")


def foremost():
    print_statement(0)
    create_list(400000)
    print_statement(1)


if __name__ == "__main__":
    cProfile.run("foremost()")

we’ve got the next output:

First, we’ve got the script outputs: print_statement(0) and print_statement(1).

Then, we’ve got the profiler output: The primary line exhibits the variety of perform calls and the time it took to run. The second line is a reminder of the sorted parameter. And, the profiler offers a desk with six columns:

1. ncalls: Exhibits the variety of calls made
2. tottime: Whole time taken by the given perform. Notice that the time made in calls to sub-functions are excluded.
3. percall: Whole time / No of calls. (the rest is disregarded)
4. cumtime: In contrast to tottime, this consists of time spent on this and all subfunctions that the higher-level perform calls. It’s most helpful and is correct for recursive features.
5. percall: The percall following cumtime is calculated because the quotient of cumtime divided by primitive calls. The primitive calls embody all of the calls that weren’t included via recursion.
6. filename: The identify of the strategy.

The primary and the final rows of the desk come from cProfile. The opposite rows are in regards to the script.

You may customise the output through the use of the Profile() class. First, you must initialize an occasion of Profile class and utilizing the strategy allow() and disable() to, respectively, begin and to finish the accumulating of profiling knowledge. Then, the pstats module can be utilized to control the outcomes collected by the profiler object.

To kind output by cumulative time, as a substitute of the usual identify the earlier code could be rewritten like this:

import cProfile, pstats


# ... 
# Identical as earlier than


if __name__ == "__main__":
    profiler = cProfile.Profile()
    profiler.allow()
    foremost()
    profiler.disable()
    stats = pstats.Stats(profiler).sort_stats('cumtime')
    stats.print_stats()

And the output turns into:

As you possibly can see, now the desk is sorted by cumtime. And the 2 rows of cProfile of the earlier desk usually are not on this desk.

Visualize profiling with Snakeviz.

The output could be very simple to analyse. However, it may possibly grow to be unreadable if the profiled code turns into too massive.

One other technique to analyse the ouput is to visualise knowledge as a substitute of learn it. To take action, we use the Snakeviz package deal. To put in it, merely kind:

pip set up snakeviz # (in your digital atmosphere)
# or
uv add snakeviz

Then, exchange stats.print_stats() by stats.dump_stats("profile.prof") to avoid wasting profiling knowledge. Now, you possibly can have a visualization of your profiling by typing:

snakeviz profile.prof

It launches a file browser interface from which you’ll select amongst two knowledge visualizations: Icicle and Sunburst.

It’s simpler to learn than the print_stats() output as a result of you possibly can work together with every component by transferring your mouse over it. For example, you possibly can have extra particulars in regards to the perform create_list()

Create a name graph with gprof2dot

A name graph is a visible illustration of the relationships between features or strategies in a program, displaying which features name others and the way lengthy every perform or technique takes. It may be seen as a map of your code.

pip set up gprof2dot # (in your digital atmosphere)
# or
uv add gprof2dot

Then exectute your by typing

python-m cProfile -o monitoring.pstats .monitoring.py # (in your digital atmosphere)
# or
uv run python-m cProfile -o monitoring.pstats .monitoring.py

It would create a monitoring.pstats that may be flip right into a name graph utilizing the next command:

gprof2dot -f pstats monitoring.pstats | dot -Tpng -o monitoring.png # (in your digital atmosphere)
# or
uv run gprof2dot -f pstats monitoring.pstats | dot -Tpng -o monitoring.png

Then the decision graph is saved right into a png file named monitoring.png

2/ Different attention-grabbing packages

a/ PyCallGraph

PyCallGraph is a Python module that creates name graph visualizations. To make use of it, you must :

To create a name graph of your code, provide run it a PyCallGraph context like this:

from pycallgraph import PyCallGraph
from pycallgraph.output import GraphvizOutput

with PyCallGraph(output=GraphvizOutput()):
    # code you need to profile

Then, you get a png of the decision graph of your code is called by default pycallgraph.png.

I’ve made the decision graph of the earlier instance:

In every field, you might have the identify of the perform, the time spent in and the variety of calls. Like with snakeviz, the graph could also be very complicated in case your code has many dependencies. However the coloration signifies the bottlenecks. In complicated code, it’s very attention-grabbing to check it to see the dependencies and relationships.

b/ PyInstrument

PyInstrument can also be a Python profiler very simple to make use of. You may add the profiler in your script by surredning the code like this:

from pyinstrument import Profiler

profiler = Profiler()
profiler.begin()

# code you need to profile

profiler.cease()
print(profiler.output_text(unicode=True, coloration=True))

The output offers

It’s much less detailled than cProfile however additionally it is extra readable. Your features are highlighted and sorted by time.

Butthe true curiosity of PyInstrument comes with its html output. To get this html output merely kind within the terminal:

pyinstrument --html .monitoring.py
# or
uv run pyinstrument --html .monitoring.py

It launches a file browser interface from which you’ll select amongst two knowledge visualizations: Name stack and Timeline.

Right here, the profile is extra detailed and you’ve got many choices to filter.

---

CPU/GPU profiler

CPU and GPU profiling is the method of analyzing the utilization and efficiency of a program on the central processing unit (CPU) and graphics processing unit (GPU). By measuring how a lot sources are spent on completely different components of the code on these processing models, builders can determine efficiency bottlenecks, perceive the place their code is being executed, and optimize their utility to realize higher efficiency and effectivity.

So far as I do know, there is just one package deal that may profile GPU energy consumption.

1/ Scalene

Scalene is a high-performance CPU, GPU and reminiscence profiler designed particularly for Python. It’s an open-source package deal that gives detailed insights. It’s designed to be quick, correct, and straightforward to make use of, making it a wonderful device for builders trying to optimize their code.

• CPU/GPU Profiling: Scalene offers detailed data on CPU/GPU utilization, together with the time spent in several components of your code. It will possibly enable you determine efficiency bottlenecks and optimize your code for higher execution instances.
• Reminiscence Profiling: Scalene tracks reminiscence allocation and deallocation, serving to you perceive how your code makes use of reminiscence. That is notably helpful for figuring out reminiscence leaks or optimizing memory-intensive purposes.
• Line-by-Line Profiling: Scalene offers line-by-line profiling, which provides you an in depth breakdown of the time spent in every line of your code. This function is invaluable for pinpointing efficiency points.
• Visualization: Scalene features a graphical interface for visualizing profiling outcomes, making it simpler to grasp and navigate the information.

To spotlight all the benefits of Scalene, I’ve developed features with the only real goal of consuming reminiscence memory_waster(), CPU cpu_waster() and GPU gpu_convolution(). All of them are in a script scalene_tuto.py.

import random
import copy
import math
import cupy as cp
import numpy as np


def memory_waster():
    """Wastes reminiscence however in a managed method"""
    memory_hogs = []

    # Create reasonably sized redundant knowledge buildings
    for i in vary(100):
        garbage_data = []
        for j in vary(1000):
            waste = f"Ineffective string #{j} repeated " * 10
            garbage_data.append(waste)
            garbage_data.append(
                {
                    "id": j,
                    "knowledge": waste,
                    "numbers": [random.random() for _ in range(50)],
                    "range_data": listing(vary(100)),
                }
            )
        memory_hogs.append(garbage_data)

    for iteration in vary(4):
        print(f"Creating copy #{iteration}...")
        memory_copy = copy.deepcopy(memory_hogs)
        memory_hogs.lengthen(memory_copy)

    return memory_hogs


def cpu_waster():
    meaningless_result = 0

    for i in vary(10000):
        for j in vary(10000):
            temp = (i**2 + j**2) * random.random()
            temp = temp / (random.random() + 0.01)
            temp = abs(temp**0.5)
            meaningless_result += temp

            # Some trigonometric operations
            angle = random.random() * math.pi
            temp += math.sin(angle) * math.cos(angle)

        if i % 100 == 0:
            random_mess = [random.randint(1, 1000) for _ in range(1000)]  # Smaller listing
            random_mess.kind()
            random_mess.reverse()
            random_mess.kind()

    return meaningless_result


def gpu_convolution():
    image_size = 128
    kernel_size = 64

    picture = np.random.random((image_size, image_size)).astype(np.float32)
    kernel = np.random.random((kernel_size, kernel_size)).astype(np.float32)

    image_gpu = cp.asarray(picture)
    kernel_gpu = cp.asarray(kernel)

    consequence = cp.zeros_like(image_gpu)

    for y in vary(kernel_size // 2, image_size - kernel_size // 2):
        for x in vary(kernel_size // 2, image_size - kernel_size // 2):
            pixel_value = 0
            for ky in vary(kernel_size):
                for kx in vary(kernel_size):
                    iy = y + ky - kernel_size // 2
                    ix = x + kx - kernel_size // 2
                    pixel_value += image_gpu[iy, ix] * kernel_gpu[ky, kx]
            consequence[y, x] = pixel_value

    result_cpu = cp.asnumpy(consequence)
    cp.cuda.Stream.null.synchronize()

    return result_cpu


def foremost():
    print("n1/ Losing some reminiscence (managed)...")
    _ = memory_waster()

    print("n2/ Losing CPU cycles (managed)...")
    _ = cpu_waster()

    print("n3/ Losing GPU cycles (managed)...")
    _ = gpu_convolution()


if __name__ == "__main__":
    foremost()

For the GPU perform, you must set up cupy based on your cuda model (nvcc --version to get it)

pip set up cupy-cuda12x # (in your digital atmosphere)
# or
uv add set up cupy-cuda12x

Additional particulars on putting in cupy could be discovered within the documentation.

To run Scalene, use the command

scalene scalene_tuto.py
# or
uv run scalene scalene_tuto.py

It profiles each CPU, GPU, and reminiscence by default. Should you solely need one or a few of the choices, use the flags --cpu, --gpu, and --memory.

Scalene offers a line-level and a perform stage profiling. And it has two interfaces: the Command Line Interface (CLI) and the net interface.

Essential: It’s higher to make use of Scalene with Ubuntu utilizing WSL. In any other case, the profiler doesn’t retrieve reminiscence consumption data.

a) Command Line Interface

By default, Scalene’s output is the net interface. To acquire the CLI as a substitute, add the flag --cli.

scalene scalene_tuto.py --cli
# or
uv run scalene scalene_tuto.py --cli

You’ve the next outcomes:

By default, the code is displayed in darkish mode. So if, like me, you’re employed in gentle mode, the consequence isn’t very fairly.

The visualization is categorized into three distinct colours, every representing a distinct profiling metric.

• The blue part represents CPU profiling, which offers a breakdown of the time spent executing Python code, native code (equivalent to C or C++), and system-related duties (like I/O operations).
• The inexperienced part is devoted to reminiscence profiling, displaying the proportion of reminiscence allotted by Python code, in addition to the general reminiscence utilization over time and its peak values.
• The yellow part focuses on GPU profiling, displaying the GPU’s working time and the quantity of knowledge copied between the GPU and CPU, measured in mb/s. It’s price noting that GPU profiling is at present restricted to NVIDIA GPUs.

b) The online interface.

The online interface is split in three components.

The colour code is similar as within the command lien interface. However some icons are added:

• 💥: Optimizable code area (efficiency indication within the Perform Profile part).
• ⚡: Optimizable traces of code.

c) AI Solutions

One of many nice benefits of Scalene is the flexibility to make use of AI to enhance the slowness and/or overconsumption you might have recognized. It at present helps OpenAI API, Amazon BedRock, Azure OpenAI and ollama in native

After deciding on your instruments, you simply should click on on 💥 or ⚡if you wish to optimize part of the code or only a line.

I take a look at it with codellama:7b-python from ollama to optimize the gpu_convolution() perform. Sadly, as talked about within the interface:

Notice that optimizations are AI-generated and is probably not appropriate.

Not one of the urged optimizations labored. However the codebase was not conducive to optimization because it was artificially difficult. Simply take away pointless traces to avoid wasting time and reminiscence. Additionally, I used a small mannequin, which could possibly be the rationale.

Although my assessments had been inconclusive, I feel this feature could be attention-grabbing and can certainly proceed to enhance.

---

Conclusion

These days, we’re much less involved in regards to the useful resource consumption of our developments, and really rapidly these optimization deficits can accumulate, making the code gradual, too gradual for manufacturing, and generally even requiring the acquisition of extra highly effective {hardware}.
Code profiling instruments are indispensable with regards to figuring out areas in want of optimization.

The mix of the reminiscence profiler and line profiler offers an excellent preliminary evaluation: simple to arrange, with easy-to-understand studies.

Instruments equivalent to cProfile and Scalene are full and have graphical representations, however require extra time to investigate. Lastly, the AI optimization possibility supplied by Scalene is an actual asset, even when in my case the mannequin used was not adequate to offer something related.

---

Interested in Python & Knowledge Science?
Observe me for extra tutorials and insights!