Python-bloggers

Parallelisation of Sci Kit Learning Python Models

This article was first published on Python – Hutsons-hacks , and kindly contributed to python-bloggers. (You can report issue about the content on this page here)
Want to share your content on python-bloggers? click here.

I wanted a way to effectively make parallel some of my prebuilt machine learning models. Luckily, the package has this capability inbuilt and it is easy to make massive performance gains in terms of your model runtime. I will show you how in the following article.

Loading imports and finding out my CPU cores

The first step is to load in the required libraries I am going to need:

from time import time
from sklearn.datasets import make_classification
from sklearn.ensemble import RandomForestClassifier
import os
import matplotlib.pyplot as plt

This loads in the time package, as we are going to create a function to fit the model and time how long it takes to run. I then use sklearn datasets and ensembles to extract the make_classification and RandomForestClassifier – these will make the data described below and fit the Random Forest classifier to the dataset. I will use the os package to get the registered CPUs on my machine, the most important part of this article and then import matplotlib.pyplot for generating some quick plots.

To find out my CPUs I have the following code:

# Get CPU cores
cpu_count = os.cpu_count()
print(f"This machine has {cpu_count} cores")
reserved_cpu = 1
final_cpu = int(cpu_count - reserved_cpu)
print("Saving one CPU so my PC does not lock up")

The cpu_count variable stores the number of CPUs your specific machine has, this will vary from machine to machine. This then prints out the number of CPUs. I then create a reserved_cpu variable to hold on to one of the cores for other processing, like graphics. The final_cpu variable just adjusts the CPU count and the following is returned to the console:

Creating a synthetic classification dataset

I am going to create a synthetic classification dataset in scikit learn by using the make_classification function. This is implemented in the code wrapper below:

# Create the dataset
X, Y = make_classification(n_samples=10000,
                           n_features=70, 
                           n_informative=65,
                           n_redundant=5)

To explain this:

This is then unpacked to the X and Y variables i.e. the independent variables (X) and the dependent variable (Y).

Create model fit timer function

In the next step we will create the model fit timer function to fit the model and time how long it takes to train:

# Function to run model and create a timer and list output of the model
def run_model(model, X, Y):
    start = time()
    model = model.fit(X,Y)
    end = time()
    result = end-start
    print("[MODEL INFO] The model ran in {:.2f}".format(result))
    return [model, result]

This run_model function takes in the model you want to fit, the X dependent variables and the Y dependent, or as I call it “the thing you want to predict”. Not sure if this will catch on?

Within this declaration:

Fitting a model on one CPU core

This will use the run_model function we created in the previous step, as well as initialise the model to be fitted:

# =============================================================================
# Single core example
# =============================================================================
#Initialise the model
model_singlecore = RandomForestClassifier(n_estimators=500, n_jobs=1)
# Create a timer and fit the model
model_single_list = run_model(model_singlecore, X, Y)

We initialise the RandomForestClassifier with 500 trees and explicitly declare 1 CPU core to use in the n_jobs parameter. I then store the model in the model_single_list variable, I have named it such, as it returns a list of two indexes – the first is the trained ML model and the second is the run time of the model training. The model trains in:

Fitting a model on all (but one) of my CPU cores

The next model will be initialised in the same way, but this time it will be trained on all (but one) of my CPU cores:

# =============================================================================
# Multicore example
# =============================================================================
model_multi = RandomForestClassifier(n_estimators=500, n_jobs=final_cpu)
model_ran = run_model(model_multi, X, Y)

This uses the final_cpu variable we calculated at the head of the script to set the number of parallel jobs to spin up and then the call the run_model function to trigger the model fit timer function. This runs in:

Okay, so we can already see an improvement in the data, it is nearly 5 times faster than the 1 core example.

This is a small dataset, but imagine this performance saving on millions of records and the model having to iterate over this, it soon mounts up. Even more, imagine distributing this over an Azure cloud server and allowing you to distribute over 16 – 32 cores, this would really aid the model training.

Benchmarking by one core increase

The code in the next section trains the model against each of the cores and outputs the results in a list for them to be plotted. I will explain each section of the code in greater depth under this code block:

results_list = list()
#Generate a list with each core iterate by one - this could be a range object equally
n_cores = [1,2,3,4,5,6,7,8]

for model_it in n_cores:
    model = RandomForestClassifier(n_estimators=500, n_jobs=model_it)
    model_fit = run_model(model, X, Y)
    results = model_fit[1] #Slice the index of the model to get the fit
    results_list.append(results)

The variables and components of the model are:

I could just print the run times to console, and that is exactly what the function will do, but next, I want to visualise the results of this in matplotlib:

# Generate plot of results
def plot_results(x_val, y_val):
   plt.plot(x_val, y_val, color="blue", linestyle="--", marker="o")
   plt.xlabel("Number of cores")
   plt.ylabel("Run time (secs)")
   plt.show() 
   
   
plot_results(n_cores, results_list)

This is a nice little wrapper function that will take the x and y inputs and generate a CPU core benchmark plot. This then produces our benchmark plot:

The run time reduces and performance gains start to level out at around 7 cores. We have taken an algorithm that trains in 30+ seconds to one that can be trained in a little over 5 seconds. Such a performance improvement by distributing the load across multiple cores.

What’s next?

The code for this tutorial can be found on GitHub.

In the next article/tutorial I will show you how to optimise sci-kit learn for hyperparameter tuning and cross-validation resampling of the data.

Until then:

To leave a comment for the author, please follow the link and comment on their blog: Python – Hutsons-hacks .

Want to share your content on python-bloggers? click here.
Exit mobile version