DEV Community

Cover image for Demystifying MLOps: Week 1
totalSophie
totalSophie

Posted on • Edited on

Demystifying MLOps: Week 1

#machinelearning #mlops #aws

Notes from MLOps ZoomCamp

1.1 What is MLOps

MLOps (Machine Learning Operations) refers to the practices, processes, and tools used to manage the entire lifecycle of machine learning models. It bridges the gap between data scientists, software engineers, and operations teams to ensure successful deployment and maintenance of ML models.

Key Components

  • Data Management and Versioning
  • Model Training and Evaluation
  • Deployment and Infrastructure
  • Continuous Integration and Delivery
  • Monitoring and Governance

1.2 Environment Preparation

You can use an EC2 instance or your local environment

Step 1

Download and install the Anaconda distribution of Python:

wget https://repo.anaconda.com/archive/Anaconda3-2022.05-Linux-x86_64.sh
bash Anaconda3-2022.05-Linux-x86_64.sh
Enter fullscreen mode Exit fullscreen mode

Step 2

Update existing packages:

sudo apt update
Enter fullscreen mode Exit fullscreen mode

Step 3

Install Docker:

sudo apt install docker.io
Enter fullscreen mode Exit fullscreen mode

Step 4

Create a separate directory for the installation and get the latest release of Docker Compose:

mkdir soft
cd soft
wget https://github.com/docker/compose/releases/download/v2.18.0/docker-compose-linux-x86_64 -O docker-compose
chmod +x docker-compose
nano ~/.bashrc
Enter fullscreen mode Exit fullscreen mode

Add the following line to the .bashrc file:

export PATH="${HOME}/soft:${PATH}"
Enter fullscreen mode Exit fullscreen mode

Save and exit the .bashrc file, then apply the changes:

source ~/.bashrc
Enter fullscreen mode Exit fullscreen mode

Step 5

Run Docker to check if it's working:

docker run hello-world
Enter fullscreen mode Exit fullscreen mode

1.3 Training a ride duration prediction model

Dataset

Dataset used is 2022 NYC green taxi trip records
Photo of a green cab
More information on the data is found at https://www.nyc.gov/assets/tlc/downloads/pdf/data_dictionary_trip_records_yellow.pdf

Download the dataset

!wget https://d37ci6vzurychx.cloudfront.net/trip-data/green_tripdata_2022-01.parquet
Enter fullscreen mode Exit fullscreen mode

Imports

Import required packages

import pandas as pd 
import pickle 
import seaborn as sns 
import matplotlib.pyplot as plt 

from sklearn.feature_extraction import DictVectorizer 
from sklearn.linear_model import LinearRegression 
from sklearn.linear_model import Lasso
from sklearn.linear_model import Ridge
from sklearn.metrics import mean_squared_error
Enter fullscreen mode Exit fullscreen mode

Reading the file: 

jan_data = pd.read_parquet("./data/green_tripdata_2022-01.parquet")
jan_data.head()
Enter fullscreen mode Exit fullscreen mode
VendorID lpep_pickup_datetime lpep_dropoff_datetime store_and_fwd_flag RatecodeID PULocationID DOLocationID passenger_count trip_distance fare_amount extra mta_tax tip_amount tolls_amount ehail_fee improvement_surcharge total_amount payment_type trip_type congestion_surcharge
0 2 2022-01-01 00:14:21 2022-01-01 00:15:33 N 1 42 42 1 0.44 3.5 0.5 0.5 0 0 0.3 4.8 2 1 0
1 1 2022-01-01 00:20:55 2022-01-01 00:29:38 N 1 116 41 1 2.1 9.5 0.5 0.5 0 0 0.3 10.8 2 1 0
2 1 2022-01-01 00:57:02 2022-01-01 01:13:14 N 1 41 140 1 3.7 14.5 3.25 0.5 4.6 0 0.3 23.15 1 1 2.75
3 2 2022-01-01 00:07:42 2022-01-01 00:15:57 N 1 181 181 1 1.69 8 0.5 0.5 0 0 0.3 9.3 2 1 0
4 2 2022-01-01 00:07:50 2022-01-01 00:28:52 N 1 33 170 1 6.26 22 0.5 0.5 5.21 0 0.3 31.26 1 1 2.75

jan_data.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 62495 entries, 0 to 62494
Data columns (total 20 columns):
 #   Column                 Non-Null Count  Dtype         
---  ------                 --------------  -----         
 0   VendorID               62495 non-null  int64         
 1   lpep_pickup_datetime   62495 non-null  datetime64[ns]
 2   lpep_dropoff_datetime  62495 non-null  datetime64[ns]
 3   store_and_fwd_flag     56200 non-null  object        
 4   RatecodeID             56200 non-null  float64       
 5   PULocationID           62495 non-null  int64         
 6   DOLocationID           62495 non-null  int64         
 7   passenger_count        56200 non-null  float64       
 8   trip_distance          62495 non-null  float64       
 9   fare_amount            62495 non-null  float64       
 10  extra                  62495 non-null  float64       
 11  mta_tax                62495 non-null  float64       
 12  tip_amount             62495 non-null  float64       
 13  tolls_amount           62495 non-null  float64       
 14  ehail_fee              0 non-null      object        
 15  improvement_surcharge  62495 non-null  float64       
 16  total_amount           62495 non-null  float64       
 17  payment_type           56200 non-null  float64       
 18  trip_type              56200 non-null  float64       
 19  congestion_surcharge   56200 non-null  float64       
dtypes: datetime64[ns](2), float64(13), int64(3), object(2)
memory usage: 9.5+ MB
Enter fullscreen mode Exit fullscreen mode

Calculate duration of trip from dropoff and pickup times 

jan_dropoff = pd.to_datetime(jan_data["lpep_dropoff_datetime"])
jan_pickup = pd.to_datetime(jan_data["lpep_pickup_datetime"])

jan_data["duration"] = jan_dropoff - jan_pickup

# Convert the values to minutes
jan_data["duration"] = jan_data.duration.apply(lambda td: td.total_seconds()/60)
Enter fullscreen mode Exit fullscreen mode

Check the distribution of the duration

jan_data.duration.describe(percentiles=[0.95, 0.98, 0.99])

count    62495.000000
mean        19.019387
std         78.215732
min          0.000000
50%         11.583333
95%         35.438333
98%         49.722667
99%         68.453000
max       1439.466667
Name: duration, dtype: float64
Enter fullscreen mode Exit fullscreen mode

sns.distplot(jan_data.duration)

Picture of the distplot

We can see the data is skewed due to the presence of outliers
Keeping only the records with the duration between 1 and 70 minutes

jan_data = jan_data[(jan_data.duration >= 1) & (jan_data.duration <= 60)]

One Hot Encoding

Using Dictionary Vectorizer for One Hot Encoding
Our categorical values that I will consider are the pickup and dropoff locations

categorical = ["PULocationID", "DOLocationID"]
numerical = ["trip_distance"]
Enter fullscreen mode Exit fullscreen mode

Convert the column type to string from integers


jan_data.loc[:, categorical] = jan_data[categorical].astype(str)

# Change our values to dictionaries

train_jan_data = jan_data[categorical + numerical].to_dict(orient='records')
Enter fullscreen mode Exit fullscreen mode 
dv = DictVectorizer()
X_train_jan = dv.fit_transform(train_jan_data)

# Convert the feature matrix to an array
fm_array = X_train_jan.toarray()
Enter fullscreen mode Exit fullscreen mode 
# Get the dimensionality of the feature matrix
fm_array.shape
Enter fullscreen mode Exit fullscreen mode

(59837, 471)

Python function that would do the above steps

Custom function to read and preprocess the data

def read_dataframe(filename):
    # Read the parquet file
    df = pd.read_parquet(filename)

    # Calculate the duration
    df_dropoff = pd.to_datetime(df["lpep_dropoff_datetime"])
    df_pickup = pd.to_datetime(df["lpep_pickup_datetime"])
    df["duration"] = df_dropoff - df_pickup

    # Remove outliers
    df["duration"] = df.duration.apply(lambda td: td.total_seconds()/60)
    df = df[(jan_data.duration >= 1) & (df.duration <= 60)]

    # Preparation for OneHotEncoding using DictVectorizer
    categorical = ["PULocationID", "DOLocationID"]
    df[categorical] = df[categorical].astype(str)

    return df
Enter fullscreen mode Exit fullscreen mode

Fitting Linear Regression Model 

# Using January data as train and Feb as Validation

df_train = read_dataframe("./data/green_tripdata_2022-01.parquet")
df_val = read_dataframe("./data/green_tripdata_2022-02.parquet")
Enter fullscreen mode Exit fullscreen mode 
dv = DictVectorizer()

categorical = ["PULocationID", "DOLocationID"]
numerical = ["trip_distance"]


train_dicts= df_train[categorical + numerical].to_dict(orient='records')
X_train = dv.fit_transform(train_dicts)

val_dicts= df_val[categorical + numerical].to_dict(orient='records')
X_val = dv.transform(val_dicts)
Enter fullscreen mode Exit fullscreen mode 
target = 'duration'
y_train = df_train[target].values
y_val = df_val[target].values
Enter fullscreen mode Exit fullscreen mode 
lr = LinearRegression()
lr.fit(X_train, y_train)

y_pred = lr.predict(X_val)

mean_squared_error(y_val, y_pred, squared=False)
Enter fullscreen mode Exit fullscreen mode

8.364575685718151

Try other models like lasso and Ridge

Save the model 

with open('models/lin_reg.bin', 'wb') as f_out:
    pickle.dump((dv, lr), f_out)
Enter fullscreen mode Exit fullscreen mode

Cover Photo by Alina Grubnyak on Unsplash

Top comments (1)

Collapse
 
endymion1818 profile image
Ben Read •

👏 Nice article, welcome to the community!

Read next

mikeyoung44 profile image

Undetectable Backdoors in Outsourced Machine Learning Models: A Theoretical Vulnerability

Mike Young -

michael_flanagan_red profile image

Building the Backbone: Entities Part 2, Agent

Michael Flanagan -

mikeyoung44 profile image

Demystifying CXL Heterogeneous Systems with Heimdall Benchmark

Mike Young -

mikeyoung44 profile image

Small But Mighty: Survey of Small Language Models in the LLM Era

Mike Young -