Code organization and refactoring

Introduction to Data Versioning with DVC

Ravi Bhadauria

Machine Learning Engineer

Prototyping vs production code

Prototyping code allows rapid iteration
But not suitable for production
- Untested and prone to errors
- Not modular with many repeated code blocks
- Likely not reproducible

Features of good production code

Reproducible: recreate same outputs in different environments and time
Modular: written as distinct, independent, and testable modules
Consistent: Single source of truth for all parameters
- A configuration/parameter file

A schematic diagram showing three properties of production code

Configuration files and YAML

Files should be in supported format
- YAML, JSON, TOML, Python
- Default is params.yaml
We'll work with YAML
- YAML Ain't Markup Language
- Allows a standard format to transfer data between languages or applications
- Simple and clean format
- Valid file extensions: .yaml or .yml

¹ https://dvc.org/doc/command-reference/params#description

YAML Syntax

Specify parameters as dictionaries
- Keys and values separated by :
Comments start with #
Data types:
- Integer, Floats, Strings
Data structures:
- Arrays
- Nested Dictionaries
Indentation is important

# Key-value pairs
a: 1
b: 1.2
c: "String value"

# Arrays
a: [1, 2.2, 3, 4.8]
b:
  - 5
  - "String value"

# Nested dictionaries
a:
  b: "Some value"
  c: "Some other value"

Example configuration file

# Data preprocessing paramters
preprocess:
  ...
  target_column: RainTomorrow
  categorical_features:
    - Location
    - WindGustDir
    - ...

# Model training/evaluation paramters
train_and_evaluate:
  rfc_params:
    n_estimators: 2
    ...

Example modular function

# In model.py
def evaluate_model(model, X_test, y_test):
    """Evaluate a model on a test set and return metrics."""
    y_pred = model.predict(X_test)
    precision = precision_score(y_test, y_pred)
    ...
    return { "accuracy": accuracy, "precision": precision,
        "recall": recall, "f1_score": f1 }

# In entry-point code (train_and_evaluate.py)
from model import evaluate_model
metrics = evaluate_model(model, X_test, y_test)

Sample project code layout

Image of code layout in a Machine Learning repository

Let's practice!

Introduction to Data Versioning with DVC